[Dynamo] Support for proxying frozen dataclasses

ghstack-source-id: fb6556cd2f9424fe223147471fe95126441954d9 Pull Request resolved: https://github.com/pytorch/pytorch/pull/134846
2025-10-28 10:34:54 +08:00 · 2024-09-01 13:30:12 -07:00
609 changed files with 12861 additions and 24211 deletions
--- a/.ci/docker/build.sh
+++ b/.ci/docker/build.sh
@ -236,7 +236,7 @@ case "$image" in
    TRITON=yes
    ;;
  pytorch-linux-focal-py3-clang10-onnx)
-    ANACONDA_PYTHON_VERSION=3.9
+    ANACONDA_PYTHON_VERSION=3.8
    CLANG_VERSION=10
    PROTOBUF=yes
    DB=yes
@ -245,7 +245,7 @@ case "$image" in
    ONNX=yes
    ;;
  pytorch-linux-focal-py3-clang9-android-ndk-r21e)
-    ANACONDA_PYTHON_VERSION=3.9
+    ANACONDA_PYTHON_VERSION=3.8
    CLANG_VERSION=9
    LLVMDEV=yes
    PROTOBUF=yes
@ -254,8 +254,8 @@ case "$image" in
    GRADLE_VERSION=6.8.3
    NINJA_VERSION=1.9.0
    ;;
-  pytorch-linux-focal-py3.9-clang10)
-    ANACONDA_PYTHON_VERSION=3.9
+  pytorch-linux-focal-py3.8-clang10)
+    ANACONDA_PYTHON_VERSION=3.8
    CLANG_VERSION=10
    PROTOBUF=yes
    DB=yes
@ -276,8 +276,8 @@ case "$image" in
    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-  pytorch-linux-focal-py3.9-gcc9)
-    ANACONDA_PYTHON_VERSION=3.9
+  pytorch-linux-focal-py3.8-gcc9)
+    ANACONDA_PYTHON_VERSION=3.8
    GCC_VERSION=9
    PROTOBUF=yes
    DB=yes
@ -318,8 +318,8 @@ case "$image" in
    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-    pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
-    ANACONDA_PYTHON_VERSION=3.9
+    pytorch-linux-jammy-py3.8-gcc11-inductor-benchmarks)
+    ANACONDA_PYTHON_VERSION=3.8
    GCC_VERSION=11
    PROTOBUF=yes
    DB=yes
@ -330,8 +330,8 @@ case "$image" in
    DOCS=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-clang12)
-    ANACONDA_PYTHON_VERSION=3.9
+  pytorch-linux-jammy-cuda11.8-cudnn9-py3.8-clang12)
+    ANACONDA_PYTHON_VERSION=3.8
    CUDA_VERSION=11.8
    CUDNN_VERSION=9
    CLANG_VERSION=12
@ -355,8 +355,8 @@ case "$image" in
    CONDA_CMAKE=yes
    VISION=yes
    ;;
-  pytorch-linux-jammy-py3.9-gcc11)
-    ANACONDA_PYTHON_VERSION=3.9
+  pytorch-linux-jammy-py3.8-gcc11)
+    ANACONDA_PYTHON_VERSION=3.8
    GCC_VERSION=11
    PROTOBUF=yes
    DB=yes
--- a/.ci/docker/centos-rocm/Dockerfile
+++ b/.ci/docker/centos-rocm/Dockerfile
@ -108,10 +108,10 @@ ENV CMAKE_C_COMPILER cc
 ENV CMAKE_CXX_COMPILER c++
 COPY ./common/install_triton.sh install_triton.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/triton.txt triton.txt
+COPY ci_commit_pins/triton-rocm.txt triton-rocm.txt
 COPY triton_version.txt triton_version.txt
 RUN if [ -n "${TRITON}" ]; then bash ./install_triton.sh; fi
-RUN rm install_triton.sh common_utils.sh triton.txt triton_version.txt
+RUN rm install_triton.sh common_utils.sh triton-rocm.txt triton_version.txt

 # Install AOTriton (Early fail)
 COPY ./aotriton_version.txt aotriton_version.txt
--- a/.ci/docker/ci_commit_pins/executorch.txt
+++ b/.ci/docker/ci_commit_pins/executorch.txt
@ -1 +1 @@
-cd1c833b079adb324871dcbbe75b43d42ffc0ade
+69472e5c43481324ad923ceb29392ab72830acee
--- a/.ci/docker/ci_commit_pins/halide.txt
+++ b/.ci/docker/ci_commit_pins/halide.txt
@ -1 +1 @@
-461c12871f336fe6f57b55d6a297f13ef209161b
+340136fec6d3ebc73e7a19eba1663e9b0ba8ab2d
--- a/.ci/docker/ci_commit_pins/triton-rocm.txt
+++ b/.ci/docker/ci_commit_pins/triton-rocm.txt
@ -0,0 +1 @@
+21eae954efa5bf584da70324b640288c3ee7aede
--- a/.ci/docker/ci_commit_pins/triton-xpu.txt
+++ b/.ci/docker/ci_commit_pins/triton-xpu.txt
@ -1 +1 @@
-cc981feba10a3f4c2e46f3fe368e8fcf5f5643df
+1b2f15840e0d70eec50d84c7a0575cb835524def
--- a/.ci/docker/ci_commit_pins/triton.txt
+++ b/.ci/docker/ci_commit_pins/triton.txt
@ -1 +1 @@
-757b6a61e7df814ba806f498f8bb3160f84b120c
+dedb7bdf339a3546896d4820366ca562c586bfa0
--- a/.ci/docker/common/install_onnx.sh
+++ b/.ci/docker/common/install_onnx.sh
@ -15,7 +15,7 @@ pip_install \
  flatbuffers==2.0 \
  mock==5.0.1 \
  ninja==1.10.2 \
-  networkx==2.5 \
+  networkx==2.0 \
  numpy==1.24.2

 # ONNXRuntime should be installed before installing
@ -30,9 +30,10 @@ pip_install \

 pip_install coloredlogs packaging

-pip_install onnxruntime==1.18.1
-pip_install onnx==1.16.2
-pip_install onnxscript==0.1.0.dev20240831 --no-deps
+pip_install onnxruntime==1.18
+pip_install onnx==1.16.0
+# pip_install "onnxscript@git+https://github.com/microsoft/onnxscript@3e869ef8ccf19b5ebd21c10d3e9c267c9a9fa729" --no-deps
+pip_install onnxscript==0.1.0.dev20240613 --no-deps
 # required by onnxscript
 pip_install ml_dtypes

--- a/.ci/docker/common/install_triton.sh
+++ b/.ci/docker/common/install_triton.sh
@ -12,7 +12,10 @@ conda_reinstall() {
  as_jenkins conda install -q -n py_$ANACONDA_PYTHON_VERSION -y --force-reinstall $*
 }

-if [ -n "${XPU_VERSION}" ]; then
+if [ -n "${ROCM_VERSION}" ]; then
+  TRITON_REPO="https://github.com/openai/triton"
+  TRITON_TEXT_FILE="triton-rocm"
+elif [ -n "${XPU_VERSION}" ]; then
  TRITON_REPO="https://github.com/intel/intel-xpu-backend-for-triton"
  TRITON_TEXT_FILE="triton-xpu"
 else
--- a/.ci/docker/requirements-ci.txt
+++ b/.ci/docker/requirements-ci.txt
@ -30,14 +30,9 @@ dill==0.3.7
 #Pinned versions: 0.3.7
 #test that import: dynamo/test_replay_record.py test_dataloader.py test_datapipe.py test_serialization.py

-expecttest==0.2.1
+expecttest==0.1.6
 #Description: method for writing tests where test framework auto populates
 # the expected output based on previous runs
-#Pinned versions: 0.2.1
-#test that import:
-
-fbscribelogger==0.1.6
-#Description: write to scribe from authenticated jobs on CI
 #Pinned versions: 0.1.6
 #test that import:

@ -109,7 +104,7 @@ networkx==2.8.8
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py

 numba==0.49.0 ; python_version < "3.9"
-numba==0.55.2 ; python_version == "3.9"
+numba==0.54.1 ; python_version == "3.9"
 numba==0.55.2 ; python_version == "3.10"
 #Description: Just-In-Time Compiler for Numerical Functions
 #Pinned versions: 0.54.1, 0.49.0, <=0.49.1
--- a/.ci/docker/ubuntu-rocm/Dockerfile
+++ b/.ci/docker/ubuntu-rocm/Dockerfile
@ -100,10 +100,10 @@ ARG TRITON
 # try to reach out to S3, which docker build runners don't have access
 COPY ./common/install_triton.sh install_triton.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/triton.txt triton.txt
+COPY ci_commit_pins/triton-rocm.txt triton-rocm.txt
 COPY triton_version.txt triton_version.txt
 RUN if [ -n "${TRITON}" ]; then bash ./install_triton.sh; fi
-RUN rm install_triton.sh common_utils.sh triton.txt triton_version.txt
+RUN rm install_triton.sh common_utils.sh triton-rocm.txt triton_version.txt

 # Install AOTriton
 COPY ./aotriton_version.txt aotriton_version.txt
--- a/.ci/pytorch/build.sh
+++ b/.ci/pytorch/build.sh
@ -285,8 +285,9 @@ else
    if [[ "$BUILD_ENVIRONMENT" != *rocm*  &&
          "$BUILD_ENVIRONMENT" != *xla* ]]; then
      if [[ "$BUILD_ENVIRONMENT" != *py3.8* ]]; then
-        # Install numpy-2.0.2 for builds which are backward compatible with 1.X
-        python -mpip install --pre numpy==2.0.2
+        # Install numpy-2.0 release candidate for builds
+        # Which should be backward compatible with Numpy-1.X
+        python -mpip install --pre numpy==2.0.0rc1
      fi

      WERROR=1 python setup.py clean
--- a/.ci/pytorch/test.sh
+++ b/.ci/pytorch/test.sh
@ -596,9 +596,6 @@ test_single_dynamo_benchmark() {

 test_inductor_micro_benchmark() {
  TEST_REPORTS_DIR=$(pwd)/test/test-reports
-  if [[ "${TEST_CONFIG}" == *cpu* ]]; then
-    test_inductor_set_cpu_affinity
-  fi
  python benchmarks/gpt_fast/benchmark.py --output "${TEST_REPORTS_DIR}/gpt_fast_benchmark.csv"
 }

@ -1482,7 +1479,7 @@ elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
  install_torchvision
  test_inductor_shard "${SHARD_NUMBER}"
  if [[ "${SHARD_NUMBER}" == 1 ]]; then
-    if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then
+    if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.8-gcc11-build ]]; then
      test_inductor_distributed
    fi
  fi
--- a/.ci/pytorch/win-test-helpers/build_pytorch.bat
+++ b/.ci/pytorch/win-test-helpers/build_pytorch.bat
@ -24,12 +24,6 @@ call %INSTALLER_DIR%\install_sccache.bat
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail

-if "%USE_XPU%"=="1" (
-  :: Install xpu support packages
-  call %INSTALLER_DIR%\install_xpu.bat
-  if errorlevel 1 exit /b 1
-)
-
 :: Miniconda has been installed as part of the Windows AMI with all the dependencies.
 :: We just need to activate it here
 call %INSTALLER_DIR%\activate_miniconda3.bat
@ -49,16 +43,6 @@ if "%VC_VERSION%" == "" (
 )
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail
-
-if "%USE_XPU%"=="1" (
-  :: Activate xpu environment - VS env is required for xpu
-  call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
-  if errorlevel 1 exit /b 1
-  :: Reduce build time. Only have MTL self-hosted runner now
-  SET TORCH_XPU_ARCH_LIST=xe-lpg
-  SET USE_KINETO=0
-)
-
@echo on
 popd

--- a/.ci/pytorch/win-test-helpers/installation-helpers/install_xpu.bat
+++ b/.ci/pytorch/win-test-helpers/installation-helpers/install_xpu.bat
@ -1,91 +0,0 @@
-@echo on
-REM Description: Install Intel Support Packages on Windows
-REM BKM reference: https://www.intel.com/content/www/us/en/developer/articles/tool/pytorch-prerequisites-for-intel-gpu/2-5.html
-
-set XPU_INSTALL_MODE=%~1
-if "%XPU_INSTALL_MODE%"=="" goto xpu_bundle_install_start
-if "%XPU_INSTALL_MODE%"=="bundle" goto xpu_bundle_install_start
-if "%XPU_INSTALL_MODE%"=="driver" goto xpu_driver_install_start
-if "%XPU_INSTALL_MODE%"=="all" goto xpu_driver_install_start
-
-:arg_error
-
-echo Illegal XPU installation mode. The value can be "bundle"/"driver"/"all"
-echo If keep the value as space, will use default "bundle" mode
-exit /b 1
-
-:xpu_driver_install_start
-:: TODO Need more testing for driver installation
-set XPU_DRIVER_LINK=https://downloadmirror.intel.com/830975/gfx_win_101.5972.exe
-curl -o xpu_driver.exe --retry 3 --retry-all-errors -k %XPU_DRIVER_LINK%
-echo "XPU Driver installing..."
-start /wait "Intel XPU Driver Installer" "xpu_driver.exe"
-if errorlevel 1 exit /b 1
-del xpu_driver.exe
-if "%XPU_INSTALL_MODE%"=="driver" goto xpu_install_end
-
-:xpu_bundle_install_start
-
-set XPU_BUNDLE_PARENT_DIR=C:\Program Files (x86)\Intel\oneAPI
-set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d1a91e2-e8b8-40a5-8c7f-5db768a6a60c/w_intel-for-pytorch-gpu-dev_p_0.5.3.37_offline.exe
-set XPU_PTI_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d1a91e2-e8b8-40a5-8c7f-5db768a6a60c/w_intel-pti-dev_p_0.9.0.37_offline.exe
-set XPU_BUNDLE_VERSION=0.5.3+31
-set XPU_PTI_VERSION=0.9.0+36
-set XPU_BUNDLE_PRODUCT_NAME=intel.oneapi.win.intel-for-pytorch-gpu-dev.product
-set XPU_PTI_PRODUCT_NAME=intel.oneapi.win.intel-pti-dev.product
-set XPU_BUNDLE_INSTALLED=0
-set XPU_PTI_INSTALLED=0
-set XPU_BUNDLE_UNINSTALL=0
-set XPU_PTI_UNINSTALL=0
-
-:: Check if XPU bundle is target version or already installed
-if exist "%XPU_BUNDLE_PARENT_DIR%\Installer\installer.exe" goto xpu_bundle_ver_check
-goto xpu_bundle_install
-
-:xpu_bundle_ver_check
-
-"%XPU_BUNDLE_PARENT_DIR%\Installer\installer.exe" --list-products > xpu_bundle_installed_ver.log
-
-for /f "tokens=1,2" %%a in (xpu_bundle_installed_ver.log) do (
-    if "%%a"=="%XPU_BUNDLE_PRODUCT_NAME%" (
-        echo %%a Installed Version: %%b
-        set XPU_BUNDLE_INSTALLED=1
-        if not "%XPU_BUNDLE_VERSION%"=="%%b" (
-            start /wait "Installer Title" "%XPU_BUNDLE_PARENT_DIR%\Installer\installer.exe" --action=remove --eula=accept --silent --product-id %XPU_BUNDLE_PRODUCT_NAME% --product-ver %%b --log-dir uninstall_bundle
-            set XPU_BUNDLE_UNINSTALL=1
-        )
-    )
-    if "%%a"=="%XPU_PTI_PRODUCT_NAME%" (
-        echo %%a Installed Version: %%b
-        set XPU_PTI_INSTALLED=1
-        if not "%XPU_PTI_VERSION%"=="%%b" (
-            start /wait "Installer Title" "%XPU_BUNDLE_PARENT_DIR%\Installer\installer.exe" --action=remove --eula=accept --silent --product-id %XPU_PTI_PRODUCT_NAME% --product-ver %%b --log-dir uninstall_bundle
-            set XPU_PTI_UNINSTALL=1
-        )
-    )
-)
-if errorlevel 1 exit /b 1
-if exist xpu_bundle_installed_ver.log del xpu_bundle_installed_ver.log
-if "%XPU_BUNDLE_INSTALLED%"=="0" goto xpu_bundle_install
-if "%XPU_BUNDLE_UNINSTALL%"=="1" goto xpu_bundle_install
-if "%XPU_PTI_INSTALLED%"=="0" goto xpu_pti_install
-if "%XPU_PTI_UNINSTALL%"=="1" goto xpu_pti_install
-goto xpu_install_end
-
-:xpu_bundle_install
-
-curl -o xpu_bundle.exe --retry 3 --retry-all-errors -k %XPU_BUNDLE_URL%
-echo "XPU Bundle installing..."
-start /wait "Intel Pytorch Bundle Installer" "xpu_bundle.exe" --action=install --eula=accept --silent --log-dir install_bundle
-if errorlevel 1 exit /b 1
-del xpu_bundle.exe
-
-:xpu_pti_install
-
-curl -o xpu_pti.exe --retry 3 --retry-all-errors -k %XPU_PTI_URL%
-echo "XPU PTI installing..."
-start /wait "Intel PTI Installer" "xpu_pti.exe" --action=install --eula=accept --silent --log-dir install_bundle
-if errorlevel 1 exit /b 1
-del xpu_pti.exe
-
-:xpu_install_end
--- a/.ci/pytorch/win-test.sh
+++ b/.ci/pytorch/win-test.sh
@ -40,9 +40,6 @@ python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==
 # Install Z3 optional dependency for Windows builds.
 python -m pip install z3-solver==4.12.2.0

-# Install tlparse for test\dynamo\test_structured_trace.py UTs.
-python -m pip install tlparse==0.3.25
-
 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/.circleci/scripts/binary_linux_test.sh
+++ b/.circleci/scripts/binary_linux_test.sh
@ -119,11 +119,6 @@ fi
 # Test the package
 /builder/check_binary.sh

-if [[ "\$GPU_ARCH_TYPE" != *s390x* && "\$GPU_ARCH_TYPE" != *xpu* && "\$GPU_ARCH_TYPE" != *rocm*  && "$PACKAGE_TYPE" != libtorch ]]; then
-  # Exclude s390, xpu, rocm and libtorch builds from smoke testing
-  python /builder/test/smoke_test/smoke_test.py --package=torchonly --torch-compile-check disabled
-fi
-
 # Clean temp files
 cd /builder && git clean -ffdx

--- a/.circleci/scripts/binary_populate_env.sh
+++ b/.circleci/scripts/binary_populate_env.sh
@ -90,7 +90,7 @@ fi
 if [[ "$PACKAGE_TYPE" =~ .*wheel.* && -n "$PYTORCH_BUILD_VERSION" && "$PYTORCH_BUILD_VERSION" =~ .*rocm.* && $(uname) == "Linux" ]]; then
    TRITON_REQUIREMENT="pytorch-triton-rocm==${TRITON_VERSION}; ${TRITON_CONSTRAINT}"
    if [[ -n "$PYTORCH_BUILD_VERSION" && "$PYTORCH_BUILD_VERSION" =~ .*dev.* ]]; then
-        TRITON_SHORTHASH=$(cut -c1-10 $PYTORCH_ROOT/.ci/docker/ci_commit_pins/triton.txt)
+        TRITON_SHORTHASH=$(cut -c1-10 $PYTORCH_ROOT/.ci/docker/ci_commit_pins/triton-rocm.txt)
        TRITON_REQUIREMENT="pytorch-triton-rocm==${TRITON_VERSION}+${TRITON_SHORTHASH}; ${TRITON_CONSTRAINT}"
    fi
    if [[ -z "${PYTORCH_EXTRA_INSTALL_REQUIREMENTS:-}" ]]; then
--- a/.circleci/scripts/binary_windows_build.sh
+++ b/.circleci/scripts/binary_windows_build.sh
@ -10,11 +10,6 @@ export SCCACHE_BUCKET=ossci-compiler-cache
 export SCCACHE_IGNORE_SERVER_IO_ERROR=1
 export VC_YEAR=2019

-if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
-    export VC_YEAR=2022
-    export USE_SCCACHE=0
-fi
-
 echo "Free space on filesystem before build:"
 df -h

--- a/.circleci/scripts/binary_windows_test.sh
+++ b/.circleci/scripts/binary_windows_test.sh
@ -6,10 +6,6 @@ source "${BINARY_ENV_FILE:-/c/w/env}"
 export CUDA_VERSION="${DESIRED_CUDA/cu/}"
 export VC_YEAR=2019

-if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
-    export VC_YEAR=2022
-fi
-
 pushd "$BUILDER_ROOT"

 ./windows/internal/smoke_test.bat
--- a/.github/actionlint.yaml
+++ b/.github/actionlint.yaml
@ -3,6 +3,8 @@ self-hosted-runner:
    # GitHub hosted x86 Linux runners
    - linux.20_04.4x
    - linux.20_04.16x
+    # Repo-specific LF hosted ARC runners
+    - linux.large.arc
    # Organization-wide AWS Linux Runners
    - linux.large
    - linux.2xlarge
--- a/.github/label_to_label.yml
+++ b/.github/label_to_label.yml
@ -31,10 +31,6 @@
  - "module: flex attention"
  then:
  - "module: higher order operators"
- any:
-  - "module: aotinductor"
-  then:
-  - "oncall: export"
 - any:
  - "module: dynamo"
  - "module: pt2-dispatcher"
--- a/.github/merge_rules.yaml
+++ b/.github/merge_rules.yaml
@ -107,8 +107,8 @@
  mandatory_checks_name:
  - EasyCLA
  - Lint
-  - pull / linux-focal-py3_9-clang9-xla / build
-  - pull / linux-focal-py3_9-clang9-xla / test (xla, 1, 1, linux.12xlarge)
+  - pull / linux-focal-py3_8-clang9-xla / build
+  - pull / linux-focal-py3_8-clang9-xla / test (xla, 1, 1, linux.12xlarge)

 - name: Documentation
  patterns:
--- a/.github/pytorch-probot.yml
+++ b/.github/pytorch-probot.yml
@ -9,7 +9,6 @@ ciflow_push_tags:
 - ciflow/inductor-rocm
 - ciflow/inductor-perf-compare
 - ciflow/inductor-micro-benchmark
- ciflow/inductor-micro-benchmark-cpu-x86
 - ciflow/inductor-cu124
 - ciflow/linux-aarch64
 - ciflow/mps
--- a/.github/requirements/pip-requirements-macOS.txt
+++ b/.github/requirements/pip-requirements-macOS.txt
@ -1,7 +1,6 @@
 boto3==1.19.12
 hypothesis==6.56.4
-expecttest==0.2.1
-fbscribelogger==0.1.6
+expecttest==0.1.6
 librosa>=0.6.2
 mpmath==1.3.0
 networkx==2.8.7
--- a/.github/scripts/build_triton_wheel.py
+++ b/.github/scripts/build_triton_wheel.py
@ -15,7 +15,9 @@ REPO_DIR = SCRIPT_DIR.parent.parent

 def read_triton_pin(device: str = "cuda") -> str:
    triton_file = "triton.txt"
-    if device == "xpu":
+    if device == "rocm":
+        triton_file = "triton-rocm.txt"
+    elif device == "xpu":
        triton_file = "triton-xpu.txt"
    with open(REPO_DIR / ".ci" / "docker" / "ci_commit_pins" / triton_file) as f:
        return f.read().strip()
--- a/.github/scripts/generate_binary_build_matrix.py
+++ b/.github/scripts/generate_binary_build_matrix.py
@ -325,7 +325,6 @@ def generate_wheels_matrix(
    os: str,
    arches: Optional[List[str]] = None,
    python_versions: Optional[List[str]] = None,
-    use_split_build: bool = False,
 ) -> List[Dict[str, str]]:
    package_type = "wheel"
    if os == "linux" or os == "linux-aarch64" or os == "linux-s390x":
@ -341,7 +340,7 @@ def generate_wheels_matrix(
        if os == "linux":
            arches += CPU_CXX11_ABI_ARCH + CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
        elif os == "windows":
-            arches += CUDA_ARCHES + XPU_ARCHES
+            arches += CUDA_ARCHES
        elif os == "linux-aarch64":
            # Only want the one arch as the CPU type is different and
            # uses different build/test scripts
@ -372,17 +371,7 @@ def generate_wheels_matrix(
            ) and python_version == "3.13":
                continue

-            if use_split_build and (
-                arch_version not in ["12.4", "12.1", "11.8", "cpu"] or os != "linux"
-            ):
-                raise RuntimeError(
-                    "Split build is only supported on linux with cuda 12.4, 12.1, 11.8, and cpu.\n"
-                    f"Currently attempting to build on arch version {arch_version} and os {os}.\n"
-                    "Please modify the matrix generation to exclude this combination."
-                )
-
            # 12.1 linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
-
            if (
                arch_version in ["12.4", "12.1", "11.8"]
                and os == "linux"
@ -396,7 +385,6 @@ def generate_wheels_matrix(
                        "desired_cuda": translate_desired_cuda(
                            gpu_arch_type, gpu_arch_version
                        ),
-                        "use_split_build": "True" if use_split_build else "False",
                        "devtoolset": (
                            "cxx11-abi" if arch_version == "cuda-aarch64" else ""
                        ),
@ -412,8 +400,7 @@ def generate_wheels_matrix(
                        ),
                    }
                )
-                # Special build building to use on Colab. PyThon 3.10 for 12.1 CUDA
-                if python_version == "3.10" and arch_version == "12.1":
+                if arch_version != "cuda-aarch64":
                    ret.append(
                        {
                            "python_version": python_version,
@ -422,16 +409,40 @@ def generate_wheels_matrix(
                            "desired_cuda": translate_desired_cuda(
                                gpu_arch_type, gpu_arch_version
                            ),
-                            "use_split_build": "True" if use_split_build else "False",
+                            "use_split_build": "True",
                            "devtoolset": "",
                            "container_image": WHEEL_CONTAINER_IMAGES[arch_version],
                            "package_type": package_type,
-                            "pytorch_extra_install_requirements": "",
-                            "build_name": f"{package_type}-py{python_version}-{gpu_arch_type}{gpu_arch_version}-full".replace(  # noqa: B950
+                            "pytorch_extra_install_requirements": (
+                                PYTORCH_EXTRA_INSTALL_REQUIREMENTS[arch_version]  # fmt: skip
+                                if os != "linux-aarch64"
+                                else ""
+                            ),
+                            "build_name": f"{package_type}-py{python_version}-{gpu_arch_type}{gpu_arch_version}-split".replace(  # noqa: B950
                                ".", "_"
                            ),
                        }
                    )
+                    # Special build building to use on Colab. PyThon 3.10 for 12.1 CUDA
+                    if python_version == "3.10" and arch_version == "12.1":
+                        ret.append(
+                            {
+                                "python_version": python_version,
+                                "gpu_arch_type": gpu_arch_type,
+                                "gpu_arch_version": gpu_arch_version,
+                                "desired_cuda": translate_desired_cuda(
+                                    gpu_arch_type, gpu_arch_version
+                                ),
+                                "use_split_build": "False",
+                                "devtoolset": "",
+                                "container_image": WHEEL_CONTAINER_IMAGES[arch_version],
+                                "package_type": package_type,
+                                "pytorch_extra_install_requirements": "",
+                                "build_name": f"{package_type}-py{python_version}-{gpu_arch_type}{gpu_arch_version}-full".replace(  # noqa: B950
+                                    ".", "_"
+                                ),
+                            }
+                        )
            else:
                ret.append(
                    {
@ -441,7 +452,6 @@ def generate_wheels_matrix(
                        "desired_cuda": translate_desired_cuda(
                            gpu_arch_type, gpu_arch_version
                        ),
-                        "use_split_build": "True" if use_split_build else "False",
                        "devtoolset": (
                            "cxx11-abi" if arch_version == "cpu-cxx11-abi" else ""
                        ),
@ -452,12 +462,11 @@ def generate_wheels_matrix(
                        ),
                        "pytorch_extra_install_requirements": (
                            PYTORCH_EXTRA_INSTALL_REQUIREMENTS["12.1"]  # fmt: skip
-                            if os != "linux" and gpu_arch_type != "xpu"
+                            if os != "linux"
                            else ""
                        ),
                    }
                )
-
    return ret


--- a/.github/scripts/generate_ci_workflows.py
+++ b/.github/scripts/generate_ci_workflows.py
@ -61,7 +61,6 @@ class BinaryBuildWorkflow:
    # Mainly for macos
    cross_compile_arm64: bool = False
    macos_runner: str = "macos-14-xlarge"
-    use_split_build: bool = False

    def __post_init__(self) -> None:
        if self.abi_version:
@ -76,11 +75,6 @@ class BinaryBuildWorkflow:
            GITHUB_DIR
            / f"workflows/generated-{self.build_environment}-{self.branches}.yml"
        )
-        if self.use_split_build:
-            output_file_path = (
-                GITHUB_DIR
-                / f"workflows/generated-{self.build_environment}-{self.branches}-split.yml"
-            )
        with open(output_file_path, "w") as output_file:
            GENERATED = "generated"  # Note that please keep the variable GENERATED otherwise phabricator will hide the whole file
            output_file.writelines([f"# @{GENERATED} DO NOT EDIT MANUALLY\n"])
@ -116,20 +110,6 @@ LINUX_BINARY_BUILD_WORFKLOWS = [
            isolated_workflow=True,
        ),
    ),
-    BinaryBuildWorkflow(
-        os=OperatingSystem.LINUX,
-        package_type="manywheel",
-        build_configs=generate_binary_build_matrix.generate_wheels_matrix(
-            OperatingSystem.LINUX,
-            use_split_build=True,
-            arches=["11.8", "12.1", "12.4", "cpu"],
-        ),
-        ciflow_config=CIFlowConfig(
-            labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},
-            isolated_workflow=True,
-        ),
-        use_split_build=True,
-    ),
    BinaryBuildWorkflow(
        os=OperatingSystem.LINUX,
        package_type="conda",
@ -182,21 +162,6 @@ LINUX_BINARY_SMOKE_WORKFLOWS = [
        ),
        branches="main",
    ),
-    BinaryBuildWorkflow(
-        os=OperatingSystem.LINUX,
-        package_type="manywheel",
-        build_configs=generate_binary_build_matrix.generate_wheels_matrix(
-            OperatingSystem.LINUX,
-            arches=["11.8", "12.1", "12.4"],
-            python_versions=["3.9"],
-            use_split_build=True,
-        ),
-        ciflow_config=CIFlowConfig(
-            labels={LABEL_CIFLOW_PERIODIC},
-        ),
-        branches="main",
-        use_split_build=True,
-    ),
    BinaryBuildWorkflow(
        os=OperatingSystem.LINUX,
        package_type="libtorch",
--- a/.github/scripts/github_utils.py
+++ b/.github/scripts/github_utils.py
@ -46,24 +46,16 @@ def gh_fetch_url_and_headers(
        with urlopen(Request(url, headers=headers, data=data_, method=method)) as conn:
            return conn.headers, reader(conn)
    except HTTPError as err:
-        if (
-            err.code == 403
-            and all(
-                key in err.headers
-                for key in ["X-RateLimit-Limit", "X-RateLimit-Remaining"]
-            )
-            and int(err.headers["X-RateLimit-Remaining"]) == 0
+        if err.code == 403 and all(
+            key in err.headers for key in ["X-RateLimit-Limit", "X-RateLimit-Used"]
        ):
            print(
-                f"""{url}
-                Rate limit exceeded:
+                f"""Rate limit exceeded:
                Used: {err.headers['X-RateLimit-Used']}
                Limit: {err.headers['X-RateLimit-Limit']}
                Remaining: {err.headers['X-RateLimit-Remaining']}
                Resets at: {err.headers['x-RateLimit-Reset']}"""
            )
-        else:
-            print(f"Error fetching {url} {err}")
        raise


--- a/.github/scripts/runner_determinator.py
+++ b/.github/scripts/runner_determinator.py
@ -137,14 +137,11 @@ def get_issue(gh: Github, repo: str, issue_num: int) -> Issue:


 def get_potential_pr_author(
-    github_token: str, repo: str, username: str, ref_type: str, ref_name: str
+    gh: Github, repo: str, username: str, ref_type: str, ref_name: str
 ) -> str:
    # If the trigger was a new tag added by a bot, this is a ciflow case
    # Fetch the actual username from the original PR. The PR number is
    # embedded in the tag name: ciflow/<name>/<pr-number>
-
-    gh = get_gh_client(github_token)
-
    if username == "pytorch-bot[bot]" and ref_type == "tag":
        split_tag = ref_name.split("/")
        if (
@ -166,32 +163,23 @@ def get_potential_pr_author(


 def is_exception_branch(branch: str) -> bool:
-    """
-    Branches that get opted out of all experiments and should always use Meta runners
-    """
    return branch.split("/")[0] in {"main", "nightly", "release", "landchecks"}


-def get_fleet(rollout_state: str, workflow_requestors: Iterable[str]) -> str:
-    """
-    Determines if the job should run on the LF fleet or the Meta fleet
-
-    Returns:
-        The appropriate label prefix for the runner, corresponding to the fleet to use.
-        This gets prefixed to the very start of the runner label.
-    """
-
+def get_workflow_type(issue: Issue, workflow_requestors: Iterable[str]) -> str:
    try:
-        if rollout_state[0] == "!":
+        first_comment = issue.get_comments()[0].body.strip("\n\t ")
+
+        if first_comment[0] == "!":
            log.info("LF Workflows are disabled for everyone. Using meta runners.")
            return WORKFLOW_LABEL_META
-        elif rollout_state[0] == "*":
+        elif first_comment[0] == "*":
            log.info("LF Workflows are enabled for everyone. Using LF runners.")
            return WORKFLOW_LABEL_LF
        else:
            all_opted_in_users = {
                usr_raw.strip("\n\t@ ").split(",")[0]
-                for usr_raw in rollout_state.split()
+                for usr_raw in first_comment.split()
            }
            opted_in_requestors = {
                usr for usr in workflow_requestors if usr in all_opted_in_users
@ -215,17 +203,11 @@ def get_fleet(rollout_state: str, workflow_requestors: Iterable[str]) -> str:


 def get_optin_feature(
-    rollout_state: str, workflow_requestors: Iterable[str], feature: str, fallback: str
+    issue: Issue, workflow_requestors: Iterable[str], feature: str, fallback: str
 ) -> str:
-    """
-    Used to dynamically opt in jobs to specific runner-type variants.
-
-    Returns:
-        The runner-type's variant name if the user has opted in to the feature, otherwise returns an empty string.
-        This variant name is prefixed to the runner-type in the label.
-    """
    try:
-        userlist = {u.lstrip("#").strip("\n\t@ ") for u in rollout_state.split()}
+        first_comment = issue.get_comments()[0].body.strip("\n\t ")
+        userlist = {u.lstrip("#").strip("\n\t@ ") for u in first_comment.split()}
        all_opted_in_users = set()
        for user in userlist:
            for i in user.split(","):
@ -253,17 +235,6 @@ def get_optin_feature(
        return fallback


-def get_rollout_state_from_issue(github_token: str, repo: str, issue_num: int) -> str:
-    """
-    Gets the first comment of the issue, which contains the desired rollout state.
-
-    The default issue we use - https://github.com/pytorch/test-infra/issues/5132
-    """
-    gh = get_gh_client(github_token)
-    issue = get_issue(gh, repo, issue_num)
-    return str(issue.get_comments()[0].body.strip("\n\t "))
-
-
 def main() -> None:
    args = parse_args()

@ -273,27 +244,25 @@ def main() -> None:
        runner_ami = RUNNER_AMI_LEGACY
    else:
        try:
-            rollout_state = get_rollout_state_from_issue(
-                args.github_token, args.github_issue_repo, args.github_issue
-            )
-
+            gh = get_gh_client(args.github_token)
+            # The default issue we use - https://github.com/pytorch/test-infra/issues/5132
+            issue = get_issue(gh, args.github_issue_repo, args.github_issue)
            username = get_potential_pr_author(
-                args.github_token,
+                gh,
                args.github_repo,
                args.github_actor,
                args.github_ref_type,
                args.github_branch,
            )
-
-            label_type = get_fleet(
-                rollout_state,
+            label_type = get_workflow_type(
+                issue,
                (
                    args.github_issue_owner,
                    username,
                ),
            )
            runner_ami = get_optin_feature(
-                rollout_state=rollout_state,
+                issue=issue,
                workflow_requestors=(
                    args.github_issue_owner,
                    username,
--- a/.github/scripts/s390x-ci/README.md
+++ b/.github/scripts/s390x-ci/README.md
@ -3,7 +3,7 @@
 ## Install prerequisites.

 ```
-$ sudo dnf install podman podman-docker jq
+$ sudo dnf install docker
 ```

 ## Add services.
@ -27,48 +27,23 @@ $ sudo systemctl enable --now qemu-user-static

 ## Rebuild the image

-First build s390x builder image `docker.io/pytorch/manylinuxs390x-builder`,
-using following commands:
-
-```
-$ cd ~
-$ git clone https://github.com/pytorch/pytorch
-$ cd pytorch
-$ git submodule update --init --recursive
-$ GPU_ARCH_TYPE=cpu-s390x "$(pwd)/.ci/docker/manywheel/build.sh" manylinuxs390x-builder
-$ docker image tag localhost/pytorch/manylinuxs390x-builder docker.io/pytorch/manylinuxs390x-builder:cpu-s390x
-$ docker image save -o ~/manywheel-s390x.tar docker.io/pytorch/manylinuxs390x-builder:cpu-s390x
-```
-
-Next step is to build `actions-runner` image using:
+In order to build or update the `iiilinuxibmcom/actions-runner` image, e.g. to get the
+latest OS security fixes, use the following commands:

 ```
 $ cd self-hosted-builder
 $ sudo docker build \
+      --build-arg repo=<owner>/<name> \
+      --build-arg token=<***> \
      --pull \
      -f actions-runner.Dockerfile \
-      -t iiilinuxibmcom/actions-runner.<name> \
+      -t iiilinuxibmcom/actions-runner \
      .
 ```

-If there are failures, ensure that selinux doesn't prevent it from working.
+If it fails, ensure that selinux doesn't prevent it from working.
 In worst case, selinux can be disabled with `setenforce 0`.

-Now prepare all necessary files for runner registration:
-
-```
-$ sudo mkdir -p /etc/actions-runner/<name>
-$ sudo chmod 700 /etc/actions-runner/<name>
-$ sudo /bin/cp <github_app_private_key_file> /etc/actions-runner/<name>/key_private.pem
-$ sudo echo <github_app_id> | sudo tee /etc/actions-runner/<name>/appid.env
-$ sudo echo <github_app_install_id> | sudo tee /etc/actions-runner/<name>/installid.env
-$ sudo echo NAME=<worker_name> | sudo tee    /etc/actions-runner/<name>/env
-$ sudo echo ORG=<github_org>   | sudo tee -a /etc/actions-runner/<name>/env
-$ cd self-hosted-builder
-$ sudo /bin/cp helpers/*.sh /usr/local/bin/
-$ sudo chmod 755 /usr/local/bin/app_token.sh /usr/local/bin/gh_token_generator.sh
-```
-
 ## Autostart the runner.

 ```
--- a/.github/scripts/s390x-ci/self-hosted-builder/actions-runner.Dockerfile
+++ b/.github/scripts/s390x-ci/self-hosted-builder/actions-runner.Dockerfile
@ -1,12 +1,12 @@
 # Self-Hosted IBM Z Github Actions Runner.

 # Temporary image: amd64 dependencies.
-FROM docker.io/amd64/ubuntu:23.10 as ld-prefix
+FROM docker.io/amd64/ubuntu:22.04 as ld-prefix
 ENV DEBIAN_FRONTEND=noninteractive
-RUN apt-get update && apt-get -y install ca-certificates libicu72 libssl3
+RUN apt-get update && apt-get -y install ca-certificates libicu70 libssl3

 # Main image.
-FROM docker.io/s390x/ubuntu:23.10
+FROM docker.io/s390x/ubuntu:22.04

 # Packages for pytorch building and testing.
 ENV DEBIAN_FRONTEND=noninteractive
@ -16,7 +16,6 @@ RUN apt-get update && apt-get -y install \
        gcc \
        git \
        jq \
-        zip \
        libxml2-dev \
        libxslt-dev \
        ninja-build \
@ -44,28 +43,24 @@ COPY fs/ /

 RUN chmod +x /usr/bin/actions-runner /usr/bin/entrypoint

-# install podman
-RUN apt -y install podman podman-docker
-
 # amd64 Github Actions Runner.
 RUN useradd -m actions-runner
 USER actions-runner
 WORKDIR /home/actions-runner
+RUN curl -L https://github.com/actions/runner/releases/download/v2.309.0/actions-runner-linux-x64-2.309.0.tar.gz | tar -xz

-# set up python virtual environment which is later used by runner.
-# build workflows use "python -m pip install ...",
-# and it doesn't work for non-root user
-RUN virtualenv --system-site-packages venv
+# repository
+ARG repo

-# copy prebuilt manywheel docker image for builds and tests
-# build command is:
-# GPU_ARCH_TYPE=cpu-s390x "$(pwd)/manywheel/build_docker.sh"
-# and save command is:
-# docker image save -o manywheel-s390x.tar pytorch/manylinuxs390x-builder:cpu-s390x
-#
-COPY --chown=actions-runner:actions-runner manywheel-s390x.tar /home/actions-runner/manywheel-s390x.tar
+# repository token
+ARG token

-RUN curl -L https://github.com/actions/runner/releases/download/v2.317.0/actions-runner-linux-x64-2.317.0.tar.gz | tar -xz
+RUN ./config.sh \
+        --unattended \
+        --url "https://github.com/${repo}" \
+        --token "${token}" \
+        --no-default-labels \
+        --labels self-hosted,linux.s390x

 ENTRYPOINT ["/usr/bin/entrypoint"]
 CMD ["/usr/bin/actions-runner"]
--- a/.github/scripts/s390x-ci/self-hosted-builder/actions-runner@.service
+++ b/.github/scripts/s390x-ci/self-hosted-builder/actions-runner@.service
@ -8,16 +8,12 @@ StartLimitIntervalSec=0
 Type=simple
 Restart=always
 ExecStartPre=-/usr/bin/docker rm --force actions-runner.%i
-ExecStartPre=-/usr/local/bin/gh_token_generator.sh /etc/actions-runner/%i/appid.env /etc/actions-runner/%i/installid.env /etc/actions-runner/%i/key_private.pem /etc/actions-runner/%i/ghtoken.env
 ExecStart=/usr/bin/docker run \
-              --env-file=/etc/actions-runner/%i/env \
-              --env-file=/etc/actions-runner/%i/ghtoken.env \
              --init \
              --interactive \
              --name=actions-runner.%i \
              --rm \
-              --privileged \
-              iiilinuxibmcom/actions-runner.%i
+              iiilinuxibmcom/actions-runner
 ExecStop=/bin/sh -c "docker exec actions-runner.%i kill -INT -- -1"
 ExecStop=/bin/sh -c "docker wait actions-runner.%i"
 ExecStop=/bin/sh -c "docker rm actions-runner.%i"
--- a/.github/scripts/s390x-ci/self-hosted-builder/fs/usr/bin/actions-runner
+++ b/.github/scripts/s390x-ci/self-hosted-builder/fs/usr/bin/actions-runner
@ -2,45 +2,5 @@

 set -e -u

-# first import docker image
-if [ -f ./manywheel-s390x.tar ] ; then
-        docker image load --input manywheel-s390x.tar
-        docker image tag docker.io/pytorch/manylinuxs390x-builder:cpu-s390x docker.io/pytorch/manylinuxs390x-builder:cpu-s390x-main
-        rm -f manywheel-s390x.tar
-fi
-
-token_file=registration-token.json
-
-# Generate registration token
-curl \
-        -X POST \
-        -H "Accept: application/vnd.github.v3+json" \
-        -H "Authorization: Bearer ${ACCESS_TOKEN}" \
-        "https://api.github.com/orgs/${ORG}/actions/runners/registration-token" \
-        -o "$token_file"
-
-unset ACCESS_TOKEN
-
-# register runner as ephemeral runner
-# it does one job, stops and unregisters
-registration_token=$(jq --raw-output .token "$token_file")
-
-./config.sh \
-        --unattended \
-        --ephemeral \
-        --url "https://github.com/${ORG}" \
-        --token "${registration_token}" \
-        --name "${NAME}" \
-        --no-default-labels \
-        --labels self-hosted,linux.s390x
-
-unset registration_token
-rm -f "$token_file"
-
-# enter into python virtual environment.
-# build workflows use "python -m pip install ...",
-# and it doesn't work for non-root user
-source venv/bin/activate
-
 # Run one job.
-./run.sh
+./run.sh --once
--- a/.github/scripts/s390x-ci/self-hosted-builder/helpers/app_token.sh
+++ b/.github/scripts/s390x-ci/self-hosted-builder/helpers/app_token.sh
@ -1,84 +0,0 @@
-#!/usr/bin/env bash
-#
-# Request an ACCESS_TOKEN to be used by a GitHub APP
-# Environment variable that need to be set up:
-# * APP_ID, the GitHub's app ID
-# * INSTALL_ID, the Github's app's installation ID
-# * APP_PRIVATE_KEY, the content of GitHub app's private key in PEM format.
-#
-# https://github.com/orgs/community/discussions/24743#discussioncomment-3245300
-#
-
-set -o pipefail
-
-_GITHUB_HOST=${GITHUB_HOST:="github.com"}
-
-# If URL is not github.com then use the enterprise api endpoint
-if [[ ${GITHUB_HOST} = "github.com" ]]; then
-  URI="https://api.${_GITHUB_HOST}"
-else
-  URI="https://${_GITHUB_HOST}/api/v3"
-fi
-
-API_VERSION=v3
-API_HEADER="Accept: application/vnd.github.${API_VERSION}+json"
-CONTENT_LENGTH_HEADER="Content-Length: 0"
-APP_INSTALLATIONS_URI="${URI}/app/installations"
-
-
-# JWT parameters based off
-# https://docs.github.com/en/developers/apps/building-github-apps/authenticating-with-github-apps#authenticating-as-a-github-app
-#
-# JWT token issuance and expiration parameters
-JWT_IAT_DRIFT=60
-JWT_EXP_DELTA=600
-
-JWT_JOSE_HEADER='{
-    "alg": "RS256",
-    "typ": "JWT"
-}'
-
-
-build_jwt_payload() {
-    now=$(date +%s)
-    iat=$((now - JWT_IAT_DRIFT))
-    jq -c \
-        --arg iat_str "${iat}" \
-        --arg exp_delta_str "${JWT_EXP_DELTA}" \
-        --arg app_id_str "${APP_ID}" \
-    '
-        ($iat_str | tonumber) as $iat
-        | ($exp_delta_str | tonumber) as $exp_delta
-        | ($app_id_str | tonumber) as $app_id
-        | .iat = $iat
-        | .exp = ($iat + $exp_delta)
-        | .iss = $app_id
-    ' <<< "{}" | tr -d '\n'
-}
-
-base64url() {
-    base64 | tr '+/' '-_' | tr -d '=\n'
-}
-
-rs256_sign() {
-    openssl dgst -binary -sha256 -sign <(echo "$1")
-}
-
-request_access_token() {
-    jwt_payload=$(build_jwt_payload)
-    encoded_jwt_parts=$(base64url <<<"${JWT_JOSE_HEADER}").$(base64url <<<"${jwt_payload}")
-    encoded_mac=$(echo -n "$encoded_jwt_parts" | rs256_sign "${APP_PRIVATE_KEY}" | base64url)
-    generated_jwt="${encoded_jwt_parts}.${encoded_mac}"
-
-    auth_header="Authorization: Bearer ${generated_jwt}"
-
-    app_installations_response=$(curl -sX POST \
-        -H "${auth_header}" \
-        -H "${API_HEADER}" \
-        --header "X-GitHub-Api-Version: 2022-11-28" \
-        --url "https://api.github.com/app/installations/${INSTALL_ID}/access_tokens" \
-    )
-    echo "$app_installations_response" | jq --raw-output '.token'
-}
-
-request_access_token
--- a/.github/scripts/s390x-ci/self-hosted-builder/helpers/gh_token_generator.sh
+++ b/.github/scripts/s390x-ci/self-hosted-builder/helpers/gh_token_generator.sh
@ -1,10 +0,0 @@
-#!/usr/bin/env bash
-
-SCRIPT_DIR=$(dirname "$0")
-APP_ID=$1
-INSTALL_ID=$2
-APP_PRIVATE_KEY=$3
-DST_FILE="$4"
-
-ACCESS_TOKEN="$(APP_ID="$(<"${APP_ID}")" INSTALL_ID="$(<"${INSTALL_ID}")" APP_PRIVATE_KEY="$(<"${APP_PRIVATE_KEY}")" "${SCRIPT_DIR}/app_token.sh")"
-echo "ACCESS_TOKEN=${ACCESS_TOKEN}" > "${DST_FILE}"
--- a/.github/templates/common.yml.j2
+++ b/.github/templates/common.yml.j2
@ -1,7 +1,7 @@
 {%- set upload_artifact_s3_action = "seemethere/upload-artifact-s3@v5" -%}
 {%- set download_artifact_s3_action = "seemethere/download-artifact-s3@v4" -%}
-{%- set upload_artifact_action = "actions/upload-artifact@v4.4.0" -%}
-{%- set download_artifact_action = "actions/download-artifact@v4.1.7" -%}
+{%- set upload_artifact_action = "actions/upload-artifact@v3" -%}
+{%- set download_artifact_action = "actions/download-artifact@v3" -%}

 {%- set timeout_minutes = 240 -%}

--- a/.github/templates/macos_binary_build_workflow.yml.j2
+++ b/.github/templates/macos_binary_build_workflow.yml.j2
@ -101,7 +101,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: !{{ config["build_name"] }}
--- a/.github/templates/upload.yml.j2
+++ b/.github/templates/upload.yml.j2
@ -45,7 +45,7 @@
  {%- if is_windows %}
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
  {%- endif %}

 {%- else %}
--- a/.github/templates/windows_binary_build_workflow.yml.j2
+++ b/.github/templates/windows_binary_build_workflow.yml.j2
@ -104,9 +104,9 @@ jobs:
      - get-label-type
 {%- if config["gpu_arch_type"] == "cuda" %}
 {%- if branches == "nightly" %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
 {%- else %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge.nonephemeral"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu.nonephemeral"
 {%- endif %}
 {%- else %}
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
--- a/.github/workflows/_binary-build-linux.yml
+++ b/.github/workflows/_binary-build-linux.yml
@ -283,7 +283,7 @@ jobs:
          # Ensure the working directory gets chowned back to the current user
          docker run --rm -v "${RUNNER_TEMP}/artifacts:/v" -w /v "${ALPINE_IMAGE}" chown -R "$(id -u):$(id -g)" .

-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: ${{ steps.filter.outputs.is-test-matrix-empty == 'False' }}
        with:
          name: ${{ inputs.build_name }}
--- a/.github/workflows/_binary-test-linux.yml
+++ b/.github/workflows/_binary-test-linux.yml
@ -210,7 +210,7 @@ jobs:

      - name: Download Build Artifacts
        if: ${{ steps.filter.outputs.is-test-matrix-empty == 'False' }}
-        uses: actions/download-artifact@v4.1.7
+        uses: actions/download-artifact@v3
        with:
          name: ${{ inputs.build_name }}
          path: "${{ runner.temp }}/artifacts/"
--- a/.github/workflows/_binary-upload.yml
+++ b/.github/workflows/_binary-upload.yml
@ -126,7 +126,7 @@ jobs:
        # NB: When the previous build job is skipped, there won't be any artifacts and
        # this step will fail. Binary build jobs can only be skipped on CI, not nightly
        continue-on-error: true
-        uses: actions/download-artifact@v4.1.7
+        uses: actions/download-artifact@v3
        with:
          name: ${{ inputs.build_name }}
          path: "${{ runner.temp }}/artifacts/"
--- a/.github/workflows/_ios-build-test.yml
+++ b/.github/workflows/_ios-build-test.yml
@ -292,7 +292,7 @@ jobs:
          bundler-cache: true

      - name: Download arm64 artifacts
-        uses: actions/download-artifact@v4.1.7
+        uses: actions/download-artifact@v3
        with:
          name: pytorch-ios-build-artifacts-arm64

--- a/.github/workflows/_runner-determinator.yml
+++ b/.github/workflows/_runner-determinator.yml
@ -196,14 +196,11 @@ jobs:


          def get_potential_pr_author(
-              github_token: str, repo: str, username: str, ref_type: str, ref_name: str
+              gh: Github, repo: str, username: str, ref_type: str, ref_name: str
          ) -> str:
              # If the trigger was a new tag added by a bot, this is a ciflow case
              # Fetch the actual username from the original PR. The PR number is
              # embedded in the tag name: ciflow/<name>/<pr-number>
-
-              gh = get_gh_client(github_token)
-
              if username == "pytorch-bot[bot]" and ref_type == "tag":
                  split_tag = ref_name.split("/")
                  if (
@ -225,32 +222,23 @@ jobs:


          def is_exception_branch(branch: str) -> bool:
-              """
-              Branches that get opted out of all experiments and should always use Meta runners
-              """
              return branch.split("/")[0] in {"main", "nightly", "release", "landchecks"}


-          def get_fleet(rollout_state: str, workflow_requestors: Iterable[str]) -> str:
-              """
-              Determines if the job should run on the LF fleet or the Meta fleet
-
-              Returns:
-                  The appropriate label prefix for the runner, corresponding to the fleet to use.
-                  This gets prefixed to the very start of the runner label.
-              """
-
+          def get_workflow_type(issue: Issue, workflow_requestors: Iterable[str]) -> str:
              try:
-                  if rollout_state[0] == "!":
+                  first_comment = issue.get_comments()[0].body.strip("\n\t ")
+
+                  if first_comment[0] == "!":
                      log.info("LF Workflows are disabled for everyone. Using meta runners.")
                      return WORKFLOW_LABEL_META
-                  elif rollout_state[0] == "*":
+                  elif first_comment[0] == "*":
                      log.info("LF Workflows are enabled for everyone. Using LF runners.")
                      return WORKFLOW_LABEL_LF
                  else:
                      all_opted_in_users = {
                          usr_raw.strip("\n\t@ ").split(",")[0]
-                          for usr_raw in rollout_state.split()
+                          for usr_raw in first_comment.split()
                      }
                      opted_in_requestors = {
                          usr for usr in workflow_requestors if usr in all_opted_in_users
@ -274,17 +262,11 @@ jobs:


          def get_optin_feature(
-              rollout_state: str, workflow_requestors: Iterable[str], feature: str, fallback: str
+              issue: Issue, workflow_requestors: Iterable[str], feature: str, fallback: str
          ) -> str:
-              """
-              Used to dynamically opt in jobs to specific runner-type variants.
-
-              Returns:
-                  The runner-type's variant name if the user has opted in to the feature, otherwise returns an empty string.
-                  This variant name is prefixed to the runner-type in the label.
-              """
              try:
-                  userlist = {u.lstrip("#").strip("\n\t@ ") for u in rollout_state.split()}
+                  first_comment = issue.get_comments()[0].body.strip("\n\t ")
+                  userlist = {u.lstrip("#").strip("\n\t@ ") for u in first_comment.split()}
                  all_opted_in_users = set()
                  for user in userlist:
                      for i in user.split(","):
@ -312,17 +294,6 @@ jobs:
                  return fallback


-          def get_rollout_state_from_issue(github_token: str, repo: str, issue_num: int) -> str:
-              """
-              Gets the first comment of the issue, which contains the desired rollout state.
-
-              The default issue we use - https://github.com/pytorch/test-infra/issues/5132
-              """
-              gh = get_gh_client(github_token)
-              issue = get_issue(gh, repo, issue_num)
-              return str(issue.get_comments()[0].body.strip("\n\t "))
-
-
          def main() -> None:
              args = parse_args()

@ -332,27 +303,25 @@ jobs:
                  runner_ami = RUNNER_AMI_LEGACY
              else:
                  try:
-                      rollout_state = get_rollout_state_from_issue(
-                          args.github_token, args.github_issue_repo, args.github_issue
-                      )
-
+                      gh = get_gh_client(args.github_token)
+                      # The default issue we use - https://github.com/pytorch/test-infra/issues/5132
+                      issue = get_issue(gh, args.github_issue_repo, args.github_issue)
                      username = get_potential_pr_author(
-                          args.github_token,
+                          gh,
                          args.github_repo,
                          args.github_actor,
                          args.github_ref_type,
                          args.github_branch,
                      )
-
-                      label_type = get_fleet(
-                          rollout_state,
+                      label_type = get_workflow_type(
+                          issue,
                          (
                              args.github_issue_owner,
                              username,
                          ),
                      )
                      runner_ami = get_optin_feature(
-                          rollout_state=rollout_state,
+                          issue=issue,
                          workflow_requestors=(
                              args.github_issue_owner,
                              username,
@ -377,7 +346,6 @@ jobs:

          if __name__ == "__main__":
              main()
-
          EOF

          cat runner_determinator.py
--- a/.github/workflows/_win-build.yml
+++ b/.github/workflows/_win-build.yml
@ -11,16 +11,6 @@ on:
        required: true
        type: string
        description: What CUDA version to build with, "cpu" for none.
-      use-xpu:
-        required: false
-        type: boolean
-        default: false
-        description: If set, build with XPU support.
-      vc-year:
-        required: false
-        type: string
-        default: "2019"
-        description: The Visual Studio year to use for building.
      build-with-debug:
        required: false
        type: boolean
@ -151,7 +141,7 @@ jobs:
          SCCACHE_REGION: us-east-1
          VC_PRODUCT: "BuildTools"
          VC_VERSION: ""
-          VC_YEAR: "${{ inputs.vc-year }}"
+          VC_YEAR: "2019"
          ALPINE_IMAGE: "308535385114.dkr.ecr.us-east-1.amazonaws.com/tool/alpine"
          AWS_DEFAULT_REGION: us-east-1
          PR_NUMBER: ${{ github.event.pull_request.number }}
@ -159,7 +149,6 @@ jobs:
          DEBUG: ${{ inputs.build-with-debug && '1' || '0' }}
          TORCH_CUDA_ARCH_LIST: "8.6"
          USE_CUDA: ${{ inputs.cuda-version != 'cpu' && '1' || '0' }}
-          USE_XPU: ${{ inputs.use-xpu == true && '1' || '0' }}
          OUR_GITHUB_JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
        run: |
          .ci/pytorch/win-build.sh
--- a/.github/workflows/build-triton-wheel.yml
+++ b/.github/workflows/build-triton-wheel.yml
@ -13,6 +13,7 @@ on:
      - .github/scripts/build_triton_wheel.py
      - .github/ci_commit_pins/triton.txt
      - .ci/docker/ci_commit_pins/triton.txt
+      - .ci/docker/ci_commit_pins/triton-rocm.txt
      - .ci/docker/ci_commit_pins/triton-xpu.txt
  pull_request:
    paths:
@ -20,6 +21,7 @@ on:
      - .github/scripts/build_triton_wheel.py
      - .github/ci_commit_pins/triton.txt
      - .ci/docker/ci_commit_pins/triton.txt
+      - .ci/docker/ci_commit_pins/triton-rocm.txt
      - .ci/docker/ci_commit_pins/triton-xpu.txt

 concurrency:
@ -29,7 +31,7 @@ concurrency:
 jobs:
  build-wheel:
    name: "Build Triton Wheel"
-    runs-on: [self-hosted, linux.4xlarge]
+    runs-on: [self-hosted, linux.2xlarge]
    strategy:
      fail-fast: false
      matrix:
@ -118,7 +120,7 @@ jobs:
          fi
          docker exec -t "${container_name}" chown -R 1000.1000 /artifacts

-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        with:
          name: pytorch-triton-wheel-${{ matrix.py_vers }}-${{ matrix.device }}
          if-no-files-found: error
@ -155,7 +157,7 @@ jobs:
          aws-region: us-east-1

      - name: Download Build Artifacts
-        uses: actions/download-artifact@v4.1.7
+        uses: actions/download-artifact@v3
        with:
          # Download all available artifacts
          path: ${{ runner.temp }}/artifacts-all
@ -251,7 +253,7 @@ jobs:
          docker exec -t "${container_name}" python /pytorch/.github/scripts/build_triton_wheel.py --build-conda --py-version="${PY_VERS}" $RELEASE
          docker exec -t "${container_name}" chown -R 1000.1000 /artifacts

-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        with:
          name: pytorch-triton-conda-${{ matrix.py_vers }}
          if-no-files-found: error
@ -271,7 +273,7 @@ jobs:
      - uses: actions/checkout@v3

      - name: Download Build Artifacts
-        uses: actions/download-artifact@v4.1.7
+        uses: actions/download-artifact@v3
        with:
          # Download all available artifacts
          path: ${{ runner.temp }}/artifacts-all
--- a/.github/workflows/create_release.yml
+++ b/.github/workflows/create_release.yml
@ -80,7 +80,7 @@ jobs:
      id-token: write
    needs: release
    steps:
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v2
        with:
          name: ${{ needs.release.outputs.pt_release_name }}
      - name: Configure AWS credentials(PyTorch account)
--- a/.github/workflows/docker-builds.yml
+++ b/.github/workflows/docker-builds.yml
@ -45,15 +45,15 @@ jobs:
          pytorch-linux-focal-cuda12.1-cudnn9-py3-gcc9-inductor-benchmarks,
          pytorch-linux-focal-cuda12.1-cudnn9-py3.12-gcc9-inductor-benchmarks,
          pytorch-linux-focal-cuda11.8-cudnn9-py3-gcc9,
-          pytorch-linux-focal-py3.9-clang10,
+          pytorch-linux-focal-py3.8-clang10,
          pytorch-linux-focal-py3.11-clang10,
          pytorch-linux-focal-py3.12-clang10,
          pytorch-linux-focal-rocm-n-1-py3,
          pytorch-linux-focal-rocm-n-py3,
-          pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-clang12,
+          pytorch-linux-jammy-cuda11.8-cudnn9-py3.8-clang12,
          pytorch-linux-focal-py3-clang9-android-ndk-r21e,
-          pytorch-linux-jammy-py3.9-gcc11,
-          pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks,
+          pytorch-linux-jammy-py3.8-gcc11,
+          pytorch-linux-jammy-py3.8-gcc11-inductor-benchmarks,
          pytorch-linux-jammy-py3.12-halide,
          pytorch-linux-jammy-xpu-2024.0-py3,
          pytorch-linux-jammy-py3-clang15-asan,
--- a/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
@ -58,7 +58,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -82,7 +81,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cpu-aarch64
      build_environment: linux-aarch64-binary-manywheel
@ -105,7 +103,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cpu-aarch64
    secrets:
@ -128,7 +125,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -153,7 +149,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cuda-aarch64
    secrets:
@ -175,7 +170,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -199,7 +193,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      build_name: manywheel-py3_10-cpu-aarch64
      build_environment: linux-aarch64-binary-manywheel
@ -222,7 +215,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      build_name: manywheel-py3_10-cpu-aarch64
    secrets:
@ -245,7 +237,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -270,7 +261,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      build_name: manywheel-py3_10-cuda-aarch64
    secrets:
@ -292,7 +282,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -316,7 +305,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      build_name: manywheel-py3_11-cpu-aarch64
      build_environment: linux-aarch64-binary-manywheel
@ -339,7 +327,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      build_name: manywheel-py3_11-cpu-aarch64
    secrets:
@ -362,7 +349,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -387,7 +373,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      build_name: manywheel-py3_11-cuda-aarch64
    secrets:
@ -409,7 +394,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -433,7 +417,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      build_name: manywheel-py3_12-cpu-aarch64
      build_environment: linux-aarch64-binary-manywheel
@ -456,7 +439,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cpu-aarch64-main
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      build_name: manywheel-py3_12-cpu-aarch64
    secrets:
@ -479,7 +461,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
@ -504,7 +485,6 @@ jobs:
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: pytorch/manylinuxaarch64-builder:cuda12.4-main
      DESIRED_DEVTOOLSET: cxx11-abi
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      build_name: manywheel-py3_12-cuda-aarch64
    secrets:
--- a/.github/workflows/generated-linux-binary-libtorch-cxx11-abi-nightly.yml
+++ b/.github/workflows/generated-linux-binary-libtorch-cxx11-abi-nightly.yml
@ -366,7 +366,7 @@ jobs:
    steps:
      - name: Setup ROCm
        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-rocm6_1-shared-with-deps-cxx11-abi
@ -476,7 +476,7 @@ jobs:
    steps:
      - name: Setup ROCm
        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-rocm6_2-shared-with-deps-cxx11-abi
--- a/.github/workflows/generated-linux-binary-libtorch-pre-cxx11-nightly.yml
+++ b/.github/workflows/generated-linux-binary-libtorch-pre-cxx11-nightly.yml
@ -366,7 +366,7 @@ jobs:
    steps:
      - name: Setup ROCm
        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-rocm6_1-shared-with-deps-pre-cxx11
@ -476,7 +476,7 @@ jobs:
    steps:
      - name: Setup ROCm
        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-rocm6_2-shared-with-deps-pre-cxx11
--- a/.github/workflows/generated-linux-binary-manywheel-main-split.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-main-split.yml
@ -1,182 +0,0 @@
-# @generated DO NOT EDIT MANUALLY
-
-# Template is at:    .github/templates/linux_binary_build_workflow.yml.j2
-# Generation script: .github/scripts/generate_ci_workflows.py
-name: linux-binary-manywheel
-
-
-on:
-  push:
-    branches:
-      - main
-    tags:
-      - 'ciflow/periodic/*'
-  workflow_dispatch:
-
-env:
-  # Needed for conda builds
-  ALPINE_IMAGE: "308535385114.dkr.ecr.us-east-1.amazonaws.com/tool/alpine"
-  ANACONDA_USER: pytorch
-  AWS_DEFAULT_REGION: us-east-1
-  BINARY_ENV_FILE: /tmp/env
-  BUILD_ENVIRONMENT: linux-binary-manywheel
-  BUILDER_ROOT: /builder
-  GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-  PR_NUMBER: ${{ github.event.pull_request.number }}
-  PYTORCH_FINAL_PACKAGE_DIR: /artifacts
-  PYTORCH_ROOT: /pytorch
-  SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
-  SKIP_ALL_TESTS: 0
-concurrency:
-  group: linux-binary-manywheel-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
-  cancel-in-progress: true
-
-jobs:
-  get-label-type:
-    name: get-label-type
-    uses: ./.github/workflows/_runner-determinator.yml
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-  manywheel-py3_9-cuda11_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      BUILDER_ROOT: /builder
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu118
-      GPU_ARCH_VERSION: 11.8
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: pytorch/manylinux-builder:cuda11.8-main
-      use_split_build: True
-      DESIRED_PYTHON: "3.9"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_9-cuda11_8
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu11==11.8.89; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu11==11.8.89; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu11==11.8.87; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu11==9.1.0.70; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu11==11.11.3.6; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu11==10.9.0.58; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu11==10.3.0.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu11==11.4.1.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu11==11.7.5.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu11==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu11==11.8.86; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_9-cuda11_8-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_9-cuda11_8-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      BUILDER_ROOT: /builder
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu118
-      GPU_ARCH_VERSION: 11.8
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: pytorch/manylinux-builder:cuda11.8-main
-      use_split_build: True
-      DESIRED_PYTHON: "3.9"
-      build_name: manywheel-py3_9-cuda11_8
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-
-  manywheel-py3_9-cuda12_1-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      BUILDER_ROOT: /builder
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu121
-      GPU_ARCH_VERSION: 12.1
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.1-main
-      use_split_build: True
-      DESIRED_PYTHON: "3.9"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_9-cuda12_1
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.1.0.70; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.1.3.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.0.2.54; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.2.106; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.4.5.107; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.1.0.106; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_9-cuda12_1-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_9-cuda12_1-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      BUILDER_ROOT: /builder
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu121
-      GPU_ARCH_VERSION: 12.1
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.1-main
-      use_split_build: True
-      DESIRED_PYTHON: "3.9"
-      build_name: manywheel-py3_9-cuda12_1
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-
-  manywheel-py3_9-cuda12_4-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      BUILDER_ROOT: /builder
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu124
-      GPU_ARCH_VERSION: 12.4
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.4-main
-      use_split_build: True
-      DESIRED_PYTHON: "3.9"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_9-cuda12_4
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.1.0.70; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.4.5.8; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.2.1.3; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.5.147; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.6.1.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.3.1.170; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_9-cuda12_4-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_9-cuda12_4-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      BUILDER_ROOT: /builder
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu124
-      GPU_ARCH_VERSION: 12.4
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.4-main
-      use_split_build: True
-      DESIRED_PYTHON: "3.9"
-      build_name: manywheel-py3_9-cuda12_4
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-linux-binary-manywheel-main.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-main.yml
@ -54,7 +54,6 @@ jobs:
      GPU_ARCH_VERSION: 11.8
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: pytorch/manylinux-builder:cuda11.8-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_9-cuda11_8
@ -78,7 +77,6 @@ jobs:
      GPU_ARCH_VERSION: 11.8
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: pytorch/manylinux-builder:cuda11.8-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cuda11_8
      build_environment: linux-binary-manywheel
@ -87,6 +85,53 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}

+  manywheel-py3_9-cuda11_8-split-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      BUILDER_ROOT: /builder
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu118
+      GPU_ARCH_VERSION: 11.8
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/manylinux-builder:cuda11.8-main
+      use_split_build: True
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_9-cuda11_8-split
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu11==11.8.89; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu11==11.8.89; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu11==11.8.87; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu11==9.1.0.70; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu11==11.11.3.6; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu11==10.9.0.58; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu11==10.3.0.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu11==11.4.1.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu11==11.7.5.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu11==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu11==11.8.86; platform_system == 'Linux' and platform_machine == 'x86_64'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda11_8-split-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cuda11_8-split-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      BUILDER_ROOT: /builder
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu118
+      GPU_ARCH_VERSION: 11.8
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/manylinux-builder:cuda11.8-main
+      use_split_build: True
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda11_8-split
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.4xlarge.nvidia.gpu
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
  manywheel-py3_9-cuda12_1-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -101,7 +146,6 @@ jobs:
      GPU_ARCH_VERSION: 12.1
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.1-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_9-cuda12_1
@ -125,7 +169,6 @@ jobs:
      GPU_ARCH_VERSION: 12.1
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.1-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cuda12_1
      build_environment: linux-binary-manywheel
@ -134,6 +177,53 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}

+  manywheel-py3_9-cuda12_1-split-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      BUILDER_ROOT: /builder
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu121
+      GPU_ARCH_VERSION: 12.1
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.1-main
+      use_split_build: True
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_9-cuda12_1-split
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.1.0.70; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.1.3.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.0.2.54; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.2.106; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.4.5.107; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.1.0.106; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda12_1-split-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cuda12_1-split-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      BUILDER_ROOT: /builder
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu121
+      GPU_ARCH_VERSION: 12.1
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.1-main
+      use_split_build: True
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_1-split
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.4xlarge.nvidia.gpu
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
  manywheel-py3_9-cuda12_4-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -148,7 +238,6 @@ jobs:
      GPU_ARCH_VERSION: 12.4
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.4-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_9-cuda12_4
@ -172,7 +261,6 @@ jobs:
      GPU_ARCH_VERSION: 12.4
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.4-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cuda12_4
      build_environment: linux-binary-manywheel
@ -180,3 +268,50 @@ jobs:
      runs_on: linux.4xlarge.nvidia.gpu
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
+
+  manywheel-py3_9-cuda12_4-split-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      BUILDER_ROOT: /builder
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu124
+      GPU_ARCH_VERSION: 12.4
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.4-main
+      use_split_build: True
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_9-cuda12_4-split
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.1.0.70; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.4.5.8; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.2.1.3; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.5.147; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.6.1.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.3.1.170; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda12_4-split-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cuda12_4-split-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      BUILDER_ROOT: /builder
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu124
+      GPU_ARCH_VERSION: 12.4
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/manylinux-builder:cuda12.4-main
+      use_split_build: True
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_4-split
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.4xlarge.nvidia.gpu
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-linux-binary-manywheel-nightly-split.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-nightly-split.yml
--- a/.github/workflows/generated-linux-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-nightly.yml
--- a/.github/workflows/generated-linux-s390x-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-s390x-binary-manywheel-nightly.yml
@ -58,7 +58,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      runs_on: linux.s390x
      ALPINE_IMAGE: "docker.io/s390x/alpine"
@ -82,7 +81,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cpu-s390x
      build_environment: linux-s390x-binary-manywheel
@ -105,7 +103,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.9"
      build_name: manywheel-py3_9-cpu-s390x
    secrets:
@ -127,7 +124,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      runs_on: linux.s390x
      ALPINE_IMAGE: "docker.io/s390x/alpine"
@ -151,7 +147,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      build_name: manywheel-py3_10-cpu-s390x
      build_environment: linux-s390x-binary-manywheel
@ -174,7 +169,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.10"
      build_name: manywheel-py3_10-cpu-s390x
    secrets:
@ -196,7 +190,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      runs_on: linux.s390x
      ALPINE_IMAGE: "docker.io/s390x/alpine"
@ -220,7 +213,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      build_name: manywheel-py3_11-cpu-s390x
      build_environment: linux-s390x-binary-manywheel
@ -243,7 +235,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.11"
      build_name: manywheel-py3_11-cpu-s390x
    secrets:
@ -265,7 +256,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      runs_on: linux.s390x
      ALPINE_IMAGE: "docker.io/s390x/alpine"
@ -289,7 +279,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      build_name: manywheel-py3_12-cpu-s390x
      build_environment: linux-s390x-binary-manywheel
@ -312,7 +301,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.12"
      build_name: manywheel-py3_12-cpu-s390x
    secrets:
@ -334,7 +322,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.13"
      runs_on: linux.s390x
      ALPINE_IMAGE: "docker.io/s390x/alpine"
@ -358,7 +345,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.13"
      build_name: manywheel-py3_13-cpu-s390x
      build_environment: linux-s390x-binary-manywheel
@ -381,7 +367,6 @@ jobs:
      DESIRED_CUDA: cpu
      GPU_ARCH_TYPE: cpu-s390x
      DOCKER_IMAGE: pytorch/manylinuxs390x-builder:cpu-s390x-main
-      use_split_build: False
      DESIRED_PYTHON: "3.13"
      build_name: manywheel-py3_13-cpu-s390x
    secrets:
--- a/.github/workflows/generated-macos-arm64-binary-conda-nightly.yml
+++ b/.github/workflows/generated-macos-arm64-binary-conda-nightly.yml
@ -117,7 +117,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_9-cpu
@ -232,7 +232,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_10-cpu
@ -347,7 +347,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_11-cpu
@ -462,7 +462,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_12-cpu
--- a/.github/workflows/generated-macos-arm64-binary-libtorch-cxx11-abi-nightly.yml
+++ b/.github/workflows/generated-macos-arm64-binary-libtorch-cxx11-abi-nightly.yml
@ -49,7 +49,7 @@ jobs:
      DESIRED_DEVTOOLSET: cxx11-abi
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -121,7 +121,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cpu-shared-with-deps-cxx11-abi
--- a/.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml
+++ b/.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml
@ -118,7 +118,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: wheel-py3_9-cpu
@ -234,7 +234,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: wheel-py3_10-cpu
@ -350,7 +350,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: wheel-py3_11-cpu
@ -466,7 +466,7 @@ jobs:
          # shellcheck disable=SC1091
          source "${RUNNER_TEMP}/anaconda/bin/activate"
          "${PYTORCH_ROOT}/.circleci/scripts/binary_macos_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: wheel-py3_12-cpu
--- a/.github/workflows/generated-windows-binary-conda-nightly.yml
+++ b/.github/workflows/generated-windows-binary-conda-nightly.yml
@ -132,7 +132,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_9-cpu
@ -213,7 +213,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_9-cpu
@ -378,7 +378,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_9-cuda11_8
@ -403,7 +403,7 @@ jobs:
    needs:
      - conda-py3_9-cuda11_8-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -460,7 +460,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_9-cuda11_8
@ -626,7 +626,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_9-cuda12_1
@ -651,7 +651,7 @@ jobs:
    needs:
      - conda-py3_9-cuda12_1-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -708,7 +708,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_9-cuda12_1
@ -874,7 +874,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_9-cuda12_4
@ -899,7 +899,7 @@ jobs:
    needs:
      - conda-py3_9-cuda12_4-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -956,7 +956,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_9-cuda12_4
@ -1121,7 +1121,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_10-cpu
@ -1202,7 +1202,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_10-cpu
@ -1367,7 +1367,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_10-cuda11_8
@ -1392,7 +1392,7 @@ jobs:
    needs:
      - conda-py3_10-cuda11_8-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -1449,7 +1449,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_10-cuda11_8
@ -1615,7 +1615,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_10-cuda12_1
@ -1640,7 +1640,7 @@ jobs:
    needs:
      - conda-py3_10-cuda12_1-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -1697,7 +1697,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_10-cuda12_1
@ -1863,7 +1863,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_10-cuda12_4
@ -1888,7 +1888,7 @@ jobs:
    needs:
      - conda-py3_10-cuda12_4-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -1945,7 +1945,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_10-cuda12_4
@ -2110,7 +2110,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_11-cpu
@ -2191,7 +2191,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_11-cpu
@ -2356,7 +2356,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_11-cuda11_8
@ -2381,7 +2381,7 @@ jobs:
    needs:
      - conda-py3_11-cuda11_8-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -2438,7 +2438,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_11-cuda11_8
@ -2604,7 +2604,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_11-cuda12_1
@ -2629,7 +2629,7 @@ jobs:
    needs:
      - conda-py3_11-cuda12_1-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -2686,7 +2686,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_11-cuda12_1
@ -2852,7 +2852,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_11-cuda12_4
@ -2877,7 +2877,7 @@ jobs:
    needs:
      - conda-py3_11-cuda12_4-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -2934,7 +2934,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_11-cuda12_4
@ -3099,7 +3099,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_12-cpu
@ -3180,7 +3180,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_12-cpu
@ -3345,7 +3345,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_12-cuda11_8
@ -3370,7 +3370,7 @@ jobs:
    needs:
      - conda-py3_12-cuda11_8-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -3427,7 +3427,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_12-cuda11_8
@ -3593,7 +3593,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_12-cuda12_1
@ -3618,7 +3618,7 @@ jobs:
    needs:
      - conda-py3_12-cuda12_1-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -3675,7 +3675,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_12-cuda12_1
@ -3841,7 +3841,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: conda-py3_12-cuda12_4
@ -3866,7 +3866,7 @@ jobs:
    needs:
      - conda-py3_12-cuda12_4-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -3923,7 +3923,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: conda-py3_12-cuda12_4
--- a/.github/workflows/generated-windows-binary-libtorch-debug-main.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-debug-main.yml
@ -51,7 +51,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -129,7 +129,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cpu-shared-with-deps-debug
@ -169,7 +169,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -214,7 +214,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cpu-shared-with-deps-debug
--- a/.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml
@ -58,7 +58,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -136,7 +136,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cpu-shared-with-deps-debug
@ -176,7 +176,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -221,7 +221,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cpu-shared-with-deps-debug
@ -290,7 +290,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cpu-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -316,7 +316,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -394,7 +394,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cuda11_8-shared-with-deps-debug
@ -419,7 +419,7 @@ jobs:
    needs:
      - libtorch-cuda11_8-shared-with-deps-debug-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -435,7 +435,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -480,7 +480,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cuda11_8-shared-with-deps-debug
@ -550,7 +550,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cuda11_8-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -576,7 +576,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -654,7 +654,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cuda12_1-shared-with-deps-debug
@ -679,7 +679,7 @@ jobs:
    needs:
      - libtorch-cuda12_1-shared-with-deps-debug-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -695,7 +695,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -740,7 +740,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cuda12_1-shared-with-deps-debug
@ -810,7 +810,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cuda12_1-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -836,7 +836,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -914,7 +914,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cuda12_4-shared-with-deps-debug
@ -939,7 +939,7 @@ jobs:
    needs:
      - libtorch-cuda12_4-shared-with-deps-debug-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -955,7 +955,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -1000,7 +1000,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cuda12_4-shared-with-deps-debug
@ -1070,7 +1070,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cuda12_4-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-windows-binary-libtorch-release-main.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-release-main.yml
@ -51,7 +51,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -129,7 +129,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cpu-shared-with-deps-release
@ -169,7 +169,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -214,7 +214,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cpu-shared-with-deps-release
--- a/.github/workflows/generated-windows-binary-libtorch-release-nightly.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-release-nightly.yml
@ -58,7 +58,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -136,7 +136,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cpu-shared-with-deps-release
@ -176,7 +176,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -221,7 +221,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cpu-shared-with-deps-release
@ -290,7 +290,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cpu-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -316,7 +316,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -394,7 +394,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cuda11_8-shared-with-deps-release
@ -419,7 +419,7 @@ jobs:
    needs:
      - libtorch-cuda11_8-shared-with-deps-release-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -435,7 +435,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -480,7 +480,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cuda11_8-shared-with-deps-release
@ -550,7 +550,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cuda11_8-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -576,7 +576,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -654,7 +654,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cuda12_1-shared-with-deps-release
@ -679,7 +679,7 @@ jobs:
    needs:
      - libtorch-cuda12_1-shared-with-deps-release-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -695,7 +695,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -740,7 +740,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cuda12_1-shared-with-deps-release
@ -810,7 +810,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cuda12_1-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -836,7 +836,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -914,7 +914,7 @@ jobs:
        shell: bash
        run: |
          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
-      - uses: actions/upload-artifact@v4.4.0
+      - uses: actions/upload-artifact@v3
        if: always()
        with:
          name: libtorch-cuda12_4-shared-with-deps-release
@ -939,7 +939,7 @@ jobs:
    needs:
      - libtorch-cuda12_4-shared-with-deps-release-build
      - get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.8xlarge.nvidia.gpu"
    timeout-minutes: 240
    env:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@ -955,7 +955,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
    steps:
      - name: Display EC2 information
        shell: bash
@ -1000,7 +1000,7 @@ jobs:
          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
-      - uses: actions/download-artifact@v4.1.7
+      - uses: actions/download-artifact@v3
        name: Download Build Artifacts
        with:
          name: libtorch-cuda12_4-shared-with-deps-release
@ -1070,7 +1070,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.8"
      build_name: libtorch-cuda12_4-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-windows-binary-wheel-nightly.yml
+++ b/.github/workflows/generated-windows-binary-wheel-nightly.yml
--- a/.github/workflows/inductor-micro-benchmark-x86.yml
+++ b/.github/workflows/inductor-micro-benchmark-x86.yml
@ -1,40 +0,0 @@
-name: inductor-micro-benchmark-x86
-
-on:
-  schedule:
-    - cron: 0 7 * * *
-  push:
-    tags:
-      - ciflow/inductor-micro-benchmark-cpu-x86/*
-  workflow_dispatch:
-
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
-    uses: ./.github/workflows/_linux-build.yml
-    with:
-      build-environment: linux-jammy-py3.9-gcc11
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
-      # Use metal host for benchmark jobs
-      test-matrix: |
-        { include: [
-          { config: "inductor-micro-benchmark-cpu-x86", shard: 1, num_shards: 1, runner: "linux.24xl.spr-metal" },
-        ]}
-
-  linux-jammy-cpu-py3_9-gcc11-inductor-micro-benchmark-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
-    with:
-      build-environment: linux-jammy-py3.9-gcc11
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
-      use-gha: anything-non-empty-to-use-gha
-      timeout-minutes: 720
--- a/.github/workflows/inductor-perf-test-nightly-x86.yml
+++ b/.github/workflows/inductor-perf-test-nightly-x86.yml
@ -48,12 +48,12 @@ concurrency:
 permissions: read-all

 jobs:
-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  linux-jammy-cpu-py3_8-gcc11-inductor-build:
+    name: linux-jammy-cpu-py3.8-gcc11-inductor
    uses: ./.github/workflows/_linux-build.yml
    with:
-      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      build-environment: linux-jammy-py3.8-gcc11-build
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11-inductor-benchmarks
      test-matrix: |
        { include: [
          { config: "inductor_huggingface_perf_cpu_x86", shard: 1, num_shards: 3, runner: "linux.24xl.spr-metal" },
@ -74,32 +74,32 @@ jobs:
      HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}


-  linux-jammy-cpu-py3_9-gcc11-inductor-test-nightly:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  linux-jammy-cpu-py3_8-gcc11-inductor-test-nightly:
+    name: linux-jammy-cpu-py3.8-gcc11-inductor
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: linux-jammy-cpu-py3_8-gcc11-inductor-build
    if: github.event.schedule == '0 7 * * *'
    with:
-      build-environment: linux-jammy-py3.9-gcc11-build
+      build-environment: linux-jammy-py3.8-gcc11-build
      dashboard-tag: training-false-inference-true-default-true-dynamic-true-aotinductor-true
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.linux-jammy-cpu-py3_8-gcc11-inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cpu-py3_8-gcc11-inductor-build.outputs.test-matrix }}
      use-gha: anything-non-empty-to-use-gha
      timeout-minutes: 720
    secrets:
      HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}


-  linux-jammy-cpu-py3_9-gcc11-inductor-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  linux-jammy-cpu-py3_8-gcc11-inductor-test:
+    name: linux-jammy-cpu-py3.8-gcc11-inductor
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: linux-jammy-cpu-py3_8-gcc11-inductor-build
    if: github.event_name == 'workflow_dispatch'
    with:
-      build-environment: linux-jammy-py3.9-gcc11-build
+      build-environment: linux-jammy-py3.8-gcc11-build
      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-aotinductor-${{ inputs.aotinductor }}
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.linux-jammy-cpu-py3_8-gcc11-inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cpu-py3_8-gcc11-inductor-build.outputs.test-matrix }}
      use-gha: anything-non-empty-to-use-gha
      timeout-minutes: 720
    secrets:
--- a/.github/workflows/inductor.yml
+++ b/.github/workflows/inductor.yml
@ -149,13 +149,13 @@ jobs:
    secrets:
      HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}

-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  linux-jammy-cpu-py3_8-gcc11-inductor-build:
+    name: linux-jammy-cpu-py3.8-gcc11-inductor
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
-      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      build-environment: linux-jammy-py3.8-gcc11-build
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11-inductor-benchmarks
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      test-matrix: |
        { include: [
@ -204,13 +204,13 @@ jobs:
    secrets:
      HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}

-  linux-jammy-cpu-py3_9-gcc11-inductor-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  linux-jammy-cpu-py3_8-gcc11-inductor-test:
+    name: linux-jammy-cpu-py3.8-gcc11-inductor
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: linux-jammy-cpu-py3_8-gcc11-inductor-build
    with:
-      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      build-environment: linux-jammy-py3.8-gcc11-build
+      docker-image: ${{ needs.linux-jammy-cpu-py3_8-gcc11-inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cpu-py3_8-gcc11-inductor-build.outputs.test-matrix }}
    secrets:
      HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
--- a/.github/workflows/lint.yml
+++ b/.github/workflows/lint.yml
@ -223,7 +223,7 @@ jobs:
          cache: pip
      - name: Install dependencies
        run: |
-          pip install pytest-rerunfailures==11.1.* pytest-flakefinder==1.1.* pytest-xdist==3.3.* expecttest==0.2.* fbscribelogger==0.1.* numpy==1.24.*
+          pip install pytest-rerunfailures==11.1.* pytest-flakefinder==1.1.* pytest-xdist==3.3.* expecttest==0.1.* numpy==1.24.*
          pip install torch --pre --index-url https://download.pytorch.org/whl/nightly/cpu/
      - name: Run run_test.py (nonretryable)
        run: |
--- a/.github/workflows/nightly.yml
+++ b/.github/workflows/nightly.yml
@ -32,8 +32,8 @@ jobs:
    needs: get-label-type
    with:
      runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
-      build-environment: linux-jammy-py3.9-gcc11
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11
+      build-environment: linux-jammy-py3.8-gcc11
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11

  docs-push:
    name: docs push
@ -43,7 +43,7 @@ jobs:
      - get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.9-gcc11
+      build-environment: linux-jammy-py3.8-gcc11
      docker-image: ${{ needs.docs-build.outputs.docker-image }}
      push: ${{ github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' || startsWith(github.event.ref, 'refs/tags/v') }}
      run-doxygen: true
--- a/.github/workflows/periodic.yml
+++ b/.github/workflows/periodic.yml
@ -104,14 +104,14 @@ jobs:
      docker-image: ${{ needs.linux-focal-cuda12_4-py3_10-gcc9-build.outputs.docker-image }}
      test-matrix: ${{ needs.linux-focal-cuda12_4-py3_10-gcc9-build.outputs.test-matrix }}

-  parallelnative-linux-jammy-py3_9-gcc11-build:
-    name: parallelnative-linux-jammy-py3.9-gcc11
+  parallelnative-linux-jammy-py3_8-gcc11-build:
+    name: parallelnative-linux-jammy-py3.8-gcc11
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: parallelnative-linux-jammy-py3.9-gcc11
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11
+      build-environment: parallelnative-linux-jammy-py3.8-gcc11
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11
      test-matrix: |
        { include: [
          { config: "default", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
@ -119,16 +119,16 @@ jobs:
          { config: "default", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
        ]}

-  parallelnative-linux-jammy-py3_9-gcc11-test:
-    name: parallelnative-linux-jammy-py3.9-gcc11
+  parallelnative-linux-jammy-py3_8-gcc11-test:
+    name: parallelnative-linux-jammy-py3.8-gcc11
    uses: ./.github/workflows/_linux-test.yml
    needs:
-      - parallelnative-linux-jammy-py3_9-gcc11-build
+      - parallelnative-linux-jammy-py3_8-gcc11-build
      - target-determination
    with:
-      build-environment: parallelnative-linux-jammy-py3.9-gcc11
-      docker-image: ${{ needs.parallelnative-linux-jammy-py3_9-gcc11-build.outputs.docker-image }}
-      test-matrix: ${{ needs.parallelnative-linux-jammy-py3_9-gcc11-build.outputs.test-matrix }}
+      build-environment: parallelnative-linux-jammy-py3.8-gcc11
+      docker-image: ${{ needs.parallelnative-linux-jammy-py3_8-gcc11-build.outputs.docker-image }}
+      test-matrix: ${{ needs.parallelnative-linux-jammy-py3_8-gcc11-build.outputs.test-matrix }}

  linux-focal-cuda11_8-py3_9-gcc9-build:
    name: linux-focal-cuda11.8-py3.9-gcc9
--- a/.github/workflows/pull.yml
+++ b/.github/workflows/pull.yml
@ -43,14 +43,14 @@ jobs:
      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
      curr_branch: ${{ github.head_ref || github.ref_name }}

-  linux-jammy-py3_9-gcc11-build:
-    name: linux-jammy-py3.9-gcc11
+  linux-jammy-py3_8-gcc11-build:
+    name: linux-jammy-py3.8-gcc11
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.9-gcc11
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11
+      build-environment: linux-jammy-py3.8-gcc11
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11
      test-matrix: |
        { include: [
          { config: "default", shard: 1, num_shards: 4, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
@ -65,49 +65,49 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-py3_9-gcc11-test:
-    name: linux-jammy-py3.9-gcc11
+  linux-jammy-py3_8-gcc11-test:
+    name: linux-jammy-py3.8-gcc11
    uses: ./.github/workflows/_linux-test.yml
    needs:
-      - linux-jammy-py3_9-gcc11-build
+      - linux-jammy-py3_8-gcc11-build
      - target-determination
    with:
-      build-environment: linux-jammy-py3.9-gcc11
-      docker-image: ${{ needs.linux-jammy-py3_9-gcc11-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_9-gcc11-build.outputs.test-matrix }}
+      build-environment: linux-jammy-py3.8-gcc11
+      docker-image: ${{ needs.linux-jammy-py3_8-gcc11-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-py3_8-gcc11-build.outputs.test-matrix }}
    secrets: inherit

  linux-docs:
    name: linux-docs
    uses: ./.github/workflows/_docs.yml
-    needs: linux-jammy-py3_9-gcc11-build
+    needs: linux-jammy-py3_8-gcc11-build
    with:
-      build-environment: linux-jammy-py3.9-gcc11
-      docker-image: ${{ needs.linux-jammy-py3_9-gcc11-build.outputs.docker-image }}
+      build-environment: linux-jammy-py3.8-gcc11
+      docker-image: ${{ needs.linux-jammy-py3_8-gcc11-build.outputs.docker-image }}
    secrets: inherit

-  linux-jammy-py3_9-gcc11-no-ops:
-    name: linux-jammy-py3.9-gcc11-no-ops
+  linux-jammy-py3_8-gcc11-no-ops:
+    name: linux-jammy-py3.8-gcc11-no-ops
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.9-gcc11-no-ops
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11
+      build-environment: linux-jammy-py3.8-gcc11-no-ops
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11
      test-matrix: |
        { include: [
          { config: "default", shard: 1, num_shards: 1 },
        ]}
    secrets: inherit

-  linux-jammy-py3_9-gcc11-pch:
-    name: linux-jammy-py3.9-gcc11-pch
+  linux-jammy-py3_8-gcc11-pch:
+    name: linux-jammy-py3.8-gcc11-pch
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.9-gcc11-pch
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11
+      build-environment: linux-jammy-py3.8-gcc11-pch
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11
      test-matrix: |
        { include: [
          { config: "default", shard: 1, num_shards: 1 },
@ -148,13 +148,13 @@ jobs:
      sync-tag: asan-test
    secrets: inherit

-  linux-focal-py3_9-clang10-onnx-build:
-    name: linux-focal-py3.9-clang10-onnx
+  linux-focal-py3_8-clang10-onnx-build:
+    name: linux-focal-py3.8-clang10-onnx
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-focal-py3.9-clang10-onnx
+      build-environment: linux-focal-py3.8-clang10-onnx
      docker-image-name: pytorch-linux-focal-py3-clang10-onnx
      test-matrix: |
        { include: [
@ -163,26 +163,26 @@ jobs:
        ]}
    secrets: inherit

-  linux-focal-py3_9-clang10-onnx-test:
-    name: linux-focal-py3.9-clang10-onnx
+  linux-focal-py3_8-clang10-onnx-test:
+    name: linux-focal-py3.8-clang10-onnx
    uses: ./.github/workflows/_linux-test.yml
    needs:
-      - linux-focal-py3_9-clang10-onnx-build
+      - linux-focal-py3_8-clang10-onnx-build
      - target-determination
    with:
-      build-environment: linux-focal-py3.9-clang10-onnx
-      docker-image: ${{ needs.linux-focal-py3_9-clang10-onnx-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-focal-py3_9-clang10-onnx-build.outputs.test-matrix }}
+      build-environment: linux-focal-py3.8-clang10-onnx
+      docker-image: ${{ needs.linux-focal-py3_8-clang10-onnx-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-focal-py3_8-clang10-onnx-build.outputs.test-matrix }}
    secrets: inherit

-  linux-focal-py3_9-clang10-build:
-    name: linux-focal-py3.9-clang10
+  linux-focal-py3_8-clang10-build:
+    name: linux-focal-py3.8-clang10
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
-      build-environment: linux-focal-py3.9-clang10
-      docker-image-name: pytorch-linux-focal-py3.9-clang10
+      build-environment: linux-focal-py3.8-clang10
+      docker-image-name: pytorch-linux-focal-py3.8-clang10
      test-matrix: |
        { include: [
          { config: "default", shard: 1, num_shards: 4, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
@ -195,16 +195,16 @@ jobs:
          { config: "dynamo", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
          { config: "dynamo", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
        ]}
-  linux-focal-py3_9-clang10-test:
-    name: linux-focal-py3.9-clang10
+  linux-focal-py3_8-clang10-test:
+    name: linux-focal-py3.8-clang10
    uses: ./.github/workflows/_linux-test.yml
    needs:
-      - linux-focal-py3_9-clang10-build
+      - linux-focal-py3_8-clang10-build
      - target-determination
    with:
-      build-environment: linux-focal-py3.9-clang10
-      docker-image: ${{ needs.linux-focal-py3_9-clang10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-focal-py3_9-clang10-build.outputs.test-matrix }}
+      build-environment: linux-focal-py3.8-clang10
+      docker-image: ${{ needs.linux-focal-py3_8-clang10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-focal-py3_8-clang10-build.outputs.test-matrix }}
    secrets: inherit

  linux-focal-py3_11-clang10-build:
@ -347,14 +347,14 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-cuda-11_8-cudnn9-py3_9-clang12-build:
-    name: linux-jammy-cuda11.8-cudnn9-py3.9-clang12
+  linux-jammy-cuda-11_8-cudnn9-py3_8-clang12-build:
+    name: linux-jammy-cuda11.8-cudnn9-py3.8-clang12
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-cuda11.8-cudnn9-py3.9-clang12
-      docker-image-name: pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-clang12
+      build-environment: linux-jammy-cuda11.8-cudnn9-py3.8-clang12
+      docker-image-name: pytorch-linux-jammy-cuda11.8-cudnn9-py3.8-clang12
      test-matrix: |
        { include: [
          { config: "default", shard: 1, num_shards: 1 },
@ -376,13 +376,13 @@ jobs:
        ]}
    secrets: inherit

-  linux-focal-py3_9-clang9-xla-build:
-    name: linux-focal-py3_9-clang9-xla
+  linux-focal-py3_8-clang9-xla-build:
+    name: linux-focal-py3_8-clang9-xla
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-focal-py3.9-clang9-xla
+      build-environment: linux-focal-py3.8-clang9-xla
      docker-image-name: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/xla_base:v1.1-lite
      test-matrix: |
        { include: [
@ -390,14 +390,14 @@ jobs:
        ]}
    secrets: inherit

-  linux-focal-py3_9-clang9-xla-test:
-    name: linux-focal-py3_9-clang9-xla
+  linux-focal-py3_8-clang9-xla-test:
+    name: linux-focal-py3_8-clang9-xla
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-focal-py3_9-clang9-xla-build
+    needs: linux-focal-py3_8-clang9-xla-build
    with:
-      build-environment: linux-focal-py3.9-clang9-xla
-      docker-image: ${{ needs.linux-focal-py3_9-clang9-xla-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-focal-py3_9-clang9-xla-build.outputs.test-matrix }}
+      build-environment: linux-focal-py3.8-clang9-xla
+      docker-image: ${{ needs.linux-focal-py3_8-clang9-xla-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-focal-py3_8-clang9-xla-build.outputs.test-matrix }}
    secrets: inherit

  win-vs2019-cpu-py3-build:
@ -488,14 +488,14 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-py3_9-gcc11-mobile-lightweight-dispatch-build:
-    name: linux-jammy-py3.9-gcc11-mobile-lightweight-dispatch-build
+  linux-jammy-py3_8-gcc11-mobile-lightweight-dispatch-build:
+    name: linux-jammy-py3.8-gcc11-mobile-lightweight-dispatch-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.9-gcc111-mobile-lightweight-dispatch-build
-      docker-image-name: pytorch-linux-jammy-py3.9-gcc11
+      build-environment: linux-jammy-py3.8-gcc111-mobile-lightweight-dispatch-build
+      docker-image-name: pytorch-linux-jammy-py3.8-gcc11
      build-generates-artifacts: false
      test-matrix: |
        { include: [
--- a/.github/workflows/runner-determinator-validator.yml
+++ b/.github/workflows/runner-determinator-validator.yml
@ -1,40 +0,0 @@
-name: Validate Runner Determinator Script is in Sync
-
-on:
-  # Run on PRs when the runner-determinator script is updated to ensure it's copies are kept in sync
-  pull_request:
-    paths:
-      - .github/workflows/_runner-determinator.yml
-      - .github/workflows/runner-determinator-validator.yml
-      - .github/scripts/runner_determinator.py
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-jobs:
-  check-runner-determinator:
-    runs-on: ubuntu-latest
-
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v2
-
-    - name: Run Hardcode runner-determinator script
-      id: hardcode-script
-      run: |
-        # Extract the script content from _runner-determinator.yml and skip the first 10 spaces of each line
-        script_content=$(awk '/cat <<EOF > runner_determinator.py/{flag=1;next}/EOF$/{flag=0}flag{print substr($0, 11)}' .github/workflows/_runner-determinator.yml)
-
-        # Write the extracted script content to runner_determinator.py
-        echo "$script_content" > runner_determinator_workflow.py
-
-    - name: Compare runner-determinator script embedded in workflow with checked in script
-      run: |
-        # Compare the extracted runner_determinator script with the existing one
-        # If this check fails, then make sure the contents of .github/scripts/runner_determinator.py is in sync with the
-        # version embedded into .github/workflows/_runner-determinator.yml
-        diff runner_determinator_workflow.py .github/scripts/runner_determinator.py
-      # Fail the job if the scripts are not identical
-      continue-on-error: false
--- a/.github/workflows/slow.yml
+++ b/.github/workflows/slow.yml
@ -102,30 +102,30 @@ jobs:
      docker-image: ${{ needs.linux-focal-cuda12_1-py3_10-gcc9-sm86-build.outputs.docker-image }}
      test-matrix: ${{ needs.linux-focal-cuda12_1-py3_10-gcc9-sm86-build.outputs.test-matrix }}

-  linux-focal-py3_9-clang10-build:
-    name: linux-focal-py3.9-clang10
+  linux-focal-py3_8-clang10-build:
+    name: linux-focal-py3.8-clang10
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-focal-py3.9-clang10
-      docker-image-name: pytorch-linux-focal-py3.9-clang10
+      build-environment: linux-focal-py3.8-clang10
+      docker-image-name: pytorch-linux-focal-py3.8-clang10
      test-matrix: |
        { include: [
          { config: "slow", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
          { config: "slow", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
        ]}

-  linux-focal-py3_9-clang10-test:
-    name: linux-focal-py3.9-clang10
+  linux-focal-py3_8-clang10-test:
+    name: linux-focal-py3.8-clang10
    uses: ./.github/workflows/_linux-test.yml
    needs:
-      - linux-focal-py3_9-clang10-build
+      - linux-focal-py3_8-clang10-build
      - target-determination
    with:
-      build-environment: linux-focal-py3.9-clang10
-      docker-image: ${{ needs.linux-focal-py3_9-clang10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-focal-py3_9-clang10-build.outputs.test-matrix }}
+      build-environment: linux-focal-py3.8-clang10
+      docker-image: ${{ needs.linux-focal-py3_8-clang10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-focal-py3_8-clang10-build.outputs.test-matrix }}

  linux-focal-rocm6_1-py3_8-build:
    name: linux-focal-rocm6.1-py3.8
--- a/.github/workflows/stale.yml
+++ b/.github/workflows/stale.yml
@ -21,7 +21,7 @@ on:
 jobs:
  stale:
    if: ${{ github.repository == 'pytorch/pytorch' }}
-    runs-on: linux.large
+    runs-on: linux.large.arc
    permissions:
      contents: read
      pull-requests: write
--- a/.github/workflows/upload-test-stats.yml
+++ b/.github/workflows/upload-test-stats.yml
@ -2,7 +2,7 @@ name: Upload test stats

 on:
  workflow_run:
-    workflows: [pull, trunk, periodic, inductor, unstable, slow, unstable-periodic, inductor-periodic, rocm, inductor-micro-benchmark, inductor-micro-benchmark-x86, inductor-cu124, inductor-rocm]
+    workflows: [pull, trunk, periodic, inductor, unstable, slow, unstable-periodic, inductor-periodic, rocm, inductor-micro-benchmark, inductor-cu124, inductor-rocm]
    types:
      - completed

--- a/.github/workflows/xpu.yml
+++ b/.github/workflows/xpu.yml
@ -49,12 +49,3 @@ jobs:
      build-environment: linux-jammy-xpu-py3.9
      docker-image: ${{ needs.linux-jammy-xpu-py3_9-build.outputs.docker-image }}
      test-matrix: ${{ needs.linux-jammy-xpu-py3_9-build.outputs.test-matrix }}
-
-  windows-xpu-build:
-    name: win-vs2022-xpu-py3
-    uses: ./.github/workflows/_win-build.yml
-    with:
-      build-environment: win-vs2022-xpu-py3
-      cuda-version: cpu
-      use-xpu: true
-      vc-year: '2022'
--- a/.lintrunner.toml
+++ b/.lintrunner.toml
@ -138,7 +138,7 @@ init_command = [
    '--dry-run={{DRYRUN}}',
    'numpy==1.24.3 ; python_version == "3.8"',
    'numpy==1.26.0 ; python_version >= "3.9"',
-    'expecttest==0.2.1',
+    'expecttest==0.1.6',
    'mypy==1.10.0',
    'sympy==1.12.1 ; python_version == "3.8"',
    'sympy==1.13.0 ; python_version >= "3.9"',
@ -1482,7 +1482,7 @@ init_command = [
    'black==23.12.1',
    'usort==1.0.8.post1',
    'isort==5.13.2',
-    'ruff==0.6.3',  # sync with RUFF
+    'ruff==0.6.0',  # sync with RUFF
 ]
 is_formatter = true

@ -1567,7 +1567,7 @@ init_command = [
    'python3',
    'tools/linter/adapters/pip_init.py',
    '--dry-run={{DRYRUN}}',
-    'ruff==0.6.3',  # sync with PYFMT
+    'ruff==0.6.0',  # sync with PYFMT
 ]
 is_formatter = true

--- a/BUILD.bazel
+++ b/BUILD.bazel
@ -332,7 +332,6 @@ intern_build_aten_ops(
        "@fbgemm",
        "@mkl",
        "@sleef",
-        "@mkl_dnn//:mkl-dnn",
    ],
 )

--- a/1
+++ b/1
@ -57,6 +57,7 @@ nn/qat/ @jerryzh168
 # Docker
 /.ci/docker/ @jeffdaily
 /.ci/docker/ci_commit_pins/triton.txt @desertfire @Chillee @eellison @shunting314 @bertmaher @jeffdaily @jataylo @jithunnair-amd @pruthvistony
+/.ci/docker/ci_commit_pins/triton-rocm.txt @jeffdaily @jataylo @jithunnair-amd @pruthvistony
 /.ci/docker/ci_commit_pins/triton-xpu.txt @EikanWang @gujinghui

 # Github Actions
--- a/RELEASE.md
+++ b/RELEASE.md
@ -50,7 +50,6 @@ Following is the Release Compatibility Matrix for PyTorch releases:

 | PyTorch version | Python | Stable CUDA | Experimental CUDA | Stable ROCm |
 | --- | --- | --- | --- | --- |
-| 2.5 | >=3.9, <=3.12, (3.13 experimental) | CUDA 11.8, CUDA 12.1, CUDA 12.4, CUDNN 9.1.0.70  | None | ROCm 6.2 |
 | 2.4 | >=3.8, <=3.12 | CUDA 11.8, CUDA 12.1, CUDNN 9.1.0.70  | CUDA 12.4, CUDNN 9.1.0.70 | ROCm 6.1 |
 | 2.3 | >=3.8, <=3.11, (3.12 experimental) | CUDA 11.8, CUDNN 8.7.0.84 | CUDA 12.1, CUDNN 8.9.2.26 | ROCm 6.0 |
 | 2.2 | >=3.8, <=3.11, (3.12 experimental) | CUDA 11.8, CUDNN 8.7.0.84 | CUDA 12.1, CUDNN 8.9.2.26 | ROCm 5.7 |
--- a/aten/src/ATen/Context.h
+++ b/aten/src/ATen/Context.h
@ -52,7 +52,7 @@ class TORCH_API Context {
    } else if (device_type == at::kIPU) {
      return at::detail::getIPUHooks().getDefaultIPUGenerator(device.index());
    } else if (device_type == at::kPrivateUse1) {
-      return at::detail::getPrivateUse1Hooks().getDefaultGenerator(
+      return at::GetPrivateUse1HooksInterface()->getDefaultGenerator(
          device.index());
    } else {
      AT_ERROR(c10::DeviceTypeName(device_type), " device type not enabled.");
@ -91,7 +91,7 @@ class TORCH_API Context {
    } else if (device_type == at::kXPU) {
      return at::detail::getXPUHooks().getDeviceFromPtr(data);
    } else if (device_type == at::kPrivateUse1) {
-      return at::detail::getPrivateUse1Hooks().getDeviceFromPtr(data);
+      return at::GetPrivateUse1HooksInterface()->getDeviceFromPtr(data);
    } else {
      AT_ERROR(c10::DeviceTypeName(device_type), " device type not enabled.");
    }
@ -182,7 +182,7 @@ class TORCH_API Context {
  void lazyInitPrivateUse1() {
    c10::call_once(thp_init, [&] {
      if (isPrivateUse1HooksRegistered()) {
-        at::detail::getPrivateUse1Hooks().initPrivateUse1();
+        at::GetPrivateUse1HooksInterface()->initPrivateUse1();
      }
    });
  }
--- a/aten/src/ATen/EmptyTensor.cpp
+++ b/aten/src/ATen/EmptyTensor.cpp
@ -21,7 +21,7 @@ c10::Allocator* GetCPUAllocatorMaybePinned(bool pin_memory) {
    } else if (at::globalContext().hasXPU()) {
      return at::detail::getXPUHooks().getPinnedMemoryAllocator();
    } else if(at::isPrivateUse1HooksRegistered()) {
-      return at::detail::getPrivateUse1Hooks().getPinnedMemoryAllocator();
+      return at::GetPrivateUse1HooksInterface()->getPinnedMemoryAllocator();
    } else {
      TORCH_CHECK(false, "Need to provide pin_memory allocator to use pin memory.")
    }
--- a/aten/src/ATen/FunctionalTensorWrapper.cpp
+++ b/aten/src/ATen/FunctionalTensorWrapper.cpp
@ -707,12 +707,7 @@ bool are_all_mutations_under_no_grad_or_inference_mode(const Tensor& functional_
 }

 bool isFunctionalTensor(const at::Tensor& tensor) {
-   return tensor.unsafeGetTensorImpl()->key_set().has(c10::DispatchKey::Functionalize);
-}
-
-bool isBaseTensor(const at::Tensor& tensor) {
-  TORCH_INTERNAL_ASSERT_DEBUG_ONLY(isFunctionalTensor(tensor));
-  return unsafeGetFunctionalWrapper(tensor)->isBaseTensor();
+  return tensor.unsafeGetTensorImpl()->key_set().has(c10::DispatchKey::Functionalize);
 }

 bool isFunctionalTensor(const std::optional<Tensor>& t) {
--- a/aten/src/ATen/FunctionalTensorWrapper.h
+++ b/aten/src/ATen/FunctionalTensorWrapper.h
@ -165,12 +165,6 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
    was_storage_changed_ = true;
  }

-  // A FunctionalTensor is considered a base if its not a view of another
-  // tensor.
-  bool isBaseTensor() const {
-    return view_metas_.empty();
-  }
-
  c10::SymInt get_storage_size(bool before) {
    return functional_storage_impl()->get_storage_size(before);
  }
@ -296,8 +290,6 @@ TORCH_API inline FunctionalTensorWrapper* unsafeGetFunctionalWrapper(
  return functional_impl;
 }

-TORCH_API bool isBaseTensor(const at::Tensor& tensor);
-
 TORCH_API bool isFunctionalTensor(const at::Tensor& tensor);
 TORCH_API bool isFunctionalTensor(const std::optional<Tensor>& t);
 TORCH_API bool isFunctionalTensor(
--- a/aten/src/ATen/ThreadLocalState.cpp
+++ b/aten/src/ATen/ThreadLocalState.cpp
@ -1,7 +1,6 @@
 #include <ATen/ThreadLocalState.h>

-#if !defined(CAFFE2_IS_XPLAT_BUILD) && !defined(C10_MOBILE) && !defined(BUILD_LITE_INTERPRETER)
-#include <ATen/autocast_mode.h>
+#if !defined(CAFFE2_IS_XPLAT_BUILD) && !defined(C10_MOBILE)
 #include <ATen/core/grad_mode.h>
 #endif

@ -19,13 +18,7 @@ ThreadLocalState::ThreadLocalState()
      torch_dispatch_mode_state_(c10::impl::TorchDispatchModeTLS::get_state()), python_dispatcher_state_(c10::impl::PythonDispatcherTLS::get_state()),
      python_torch_function_state_(at::impl::PythonTorchFunctionTLS::get_state()),
      saved_tensors_default_hooks_state_(at::SavedTensorDefaultHooks::get_tls_state()), functionalization_reapply_views_state_(at::functionalization::impl::getFunctionalizationReapplyViewsTLS()),
-      saved_objects_(at::impl::ThreadLocalPythonObjects::get_state()) {
-#if !defined(CAFFE2_IS_XPLAT_BUILD) && !defined(C10_MOBILE) && !defined(BUILD_LITE_INTERPRETER)
-  for(uint8_t i=0; i<autocast_dtypes_.size(); i++) {
-     autocast_dtypes_[i] = at::autocast::get_autocast_dtype(static_cast<at::DeviceType>(i));
-  }
-#endif
-}
+      saved_objects_(at::impl::ThreadLocalPythonObjects::get_state()) {}

 void ThreadLocalState::set_grad_mode(bool enabled) {
  autograd_tls_.set_grad_mode(enabled);
@ -61,11 +54,6 @@ void ThreadLocalState::setThreadLocalState(
  at::functionalization::impl::setFunctionalizationReapplyViewsTLS(state.functionalization_reapply_views_state_);

  at::impl::ThreadLocalPythonObjects::set_state(state.saved_objects_);
-#if !defined(CAFFE2_IS_XPLAT_BUILD) && !defined(C10_MOBILE) && !defined(BUILD_LITE_INTERPRETER)
-  for(uint8_t i=0; i<state.autocast_dtypes_.size(); i++) {
-     at::autocast::set_autocast_dtype(static_cast<at::DeviceType>(i), state.autocast_dtypes_[i]);
-  }
-#endif
 }

 } // namespace at
--- a/aten/src/ATen/ThreadLocalState.h
+++ b/aten/src/ATen/ThreadLocalState.h
@ -78,13 +78,6 @@ class TORCH_API ThreadLocalState {
  // TLS for arbitrary python objects that is registered via hooks
  at::impl::ThreadLocalPythonObjects saved_objects_;

-#if !defined(CAFFE2_IS_XPLAT_BUILD) && !defined(C10_MOBILE) && \
-    !defined(BUILD_LITE_INTERPRETER)
-  // TLS for autocast dtypes
-  std::array<at::ScalarType, at::COMPILE_TIME_MAX_DEVICE_TYPES>
-      autocast_dtypes_;
-#endif
-
  friend class ThreadLocalStateGuard;
 };

--- a/aten/src/ATen/autocast_mode.cpp
+++ b/aten/src/ATen/autocast_mode.cpp
@ -69,7 +69,7 @@ thread_local std::array<at::ScalarType, at::COMPILE_TIME_MAX_DEVICE_TYPES>
        at::ScalarType::Undefined, // Vulkan
        at::ScalarType::Undefined, // Metal
        at::kHalf, // XPU
-        at::kHalf, // MPS
+        at::ScalarType::Undefined, // MPS
        at::ScalarType::Undefined, // Meta (tensors with no data)
        at::kBFloat16, // HPU / HABANA
        at::ScalarType::Undefined, // SX-Aurora / NEC
@ -206,118 +206,6 @@ TORCH_LIBRARY_IMPL(aten, Autocast, m) {
         TORCH_FN((&at::autocast::binary_cross_entropy_banned)));
 }

-TORCH_LIBRARY_IMPL(_, AutocastMPS, m) {
-  m.fallback(torch::CppFunction::makeFallthrough());
-}
-
-TORCH_LIBRARY_IMPL(aten, AutocastMPS, m) {
-  // lower_precision_fp
-  KERNEL_MPS2(_convolution, deprecated, lower_precision_fp)
-  KERNEL_MPS(_convolution, lower_precision_fp)
-  KERNEL_MPS(conv1d, lower_precision_fp)
-  KERNEL_MPS(conv2d, lower_precision_fp)
-  KERNEL_MPS(conv_tbc, lower_precision_fp)
-  KERNEL_MPS(conv_transpose1d, lower_precision_fp)
-  KERNEL_MPS2(conv_transpose2d, input, lower_precision_fp)
-  KERNEL_MPS(convolution, lower_precision_fp)
-  KERNEL_MPS(_mps_convolution, lower_precision_fp)
-  KERNEL_MPS(prelu, lower_precision_fp)
-  KERNEL_MPS(addmm, lower_precision_fp)
-  KERNEL_MPS(addmv, lower_precision_fp)
-  KERNEL_MPS(addr, lower_precision_fp)
-  KERNEL_MPS(matmul, lower_precision_fp)
-  KERNEL_MPS(einsum, lower_precision_fp)
-  KERNEL_MPS(mm, lower_precision_fp)
-  KERNEL_MPS(mv, lower_precision_fp)
-  KERNEL_MPS(linear, lower_precision_fp)
-  KERNEL_MPS(addbmm, lower_precision_fp)
-  KERNEL_MPS(baddbmm, lower_precision_fp)
-  KERNEL_MPS(bmm, lower_precision_fp)
-  KERNEL_MPS(chain_matmul, lower_precision_fp)
-  KERNEL_MPS(linalg_multi_dot, lower_precision_fp)
-  KERNEL_MPS(lstm_cell, lower_precision_fp)
-
-  // fp32
-  KERNEL_MPS(acos, fp32)
-  KERNEL_MPS(asin, fp32)
-  KERNEL_MPS(cosh, fp32)
-  KERNEL_MPS(erfinv, fp32)
-  KERNEL_MPS(exp, fp32)
-  KERNEL_MPS(expm1, fp32)
-  KERNEL_MPS(log, fp32)
-  KERNEL_MPS(log10, fp32)
-  KERNEL_MPS(log2, fp32)
-  KERNEL_MPS(log1p, fp32)
-  KERNEL_MPS(reciprocal, fp32)
-  KERNEL_MPS(rsqrt, fp32)
-  KERNEL_MPS(sinh, fp32)
-  KERNEL_MPS(tan, fp32)
-  KERNEL_MPS2(pow, Tensor_Scalar, fp32)
-  KERNEL_MPS2(pow, Tensor_Tensor, fp32)
-  KERNEL_MPS2(pow, Scalar, fp32)
-  KERNEL_MPS(softplus, fp32)
-  KERNEL_MPS(layer_norm, fp32)
-  KERNEL_MPS(native_layer_norm, fp32)
-  KERNEL_MPS(group_norm, fp32)
-  KERNEL_MPS2(frobenius_norm, dim, fp32)
-  KERNEL_MPS(nuclear_norm, fp32)
-  KERNEL_MPS2(nuclear_norm, dim, fp32)
-  KERNEL_MPS(batch_norm, fp32)
-  KERNEL_MPS(cosine_similarity, fp32)
-  KERNEL_MPS(poisson_nll_loss, fp32)
-  KERNEL_MPS(cosine_embedding_loss, fp32)
-  KERNEL_MPS(nll_loss, fp32)
-  KERNEL_MPS(nll_loss2d, fp32)
-  KERNEL_MPS(hinge_embedding_loss, fp32)
-  KERNEL_MPS(kl_div, fp32)
-  KERNEL_MPS(l1_loss, fp32)
-  KERNEL_MPS(smooth_l1_loss, fp32)
-  KERNEL_MPS(huber_loss, fp32)
-  KERNEL_MPS(mse_loss, fp32)
-  KERNEL_MPS(margin_ranking_loss, fp32)
-  KERNEL_MPS(multilabel_margin_loss, fp32)
-  KERNEL_MPS(soft_margin_loss, fp32)
-  KERNEL_MPS(triplet_margin_loss, fp32)
-  KERNEL_MPS(multi_margin_loss, fp32)
-  KERNEL_MPS(binary_cross_entropy_with_logits, fp32)
-  KERNEL_MPS(dist, fp32)
-  KERNEL_MPS(pdist, fp32)
-  KERNEL_MPS(cdist, fp32)
-  KERNEL_MPS(renorm, fp32)
-  KERNEL_MPS(logsumexp, fp32)
-
-  // fp32_set_opt_dtype
-  KERNEL_MPS(prod, fp32)
-  KERNEL_MPS2(prod, dim_int, fp32)
-  KERNEL_MPS2(prod, dim_Dimname, fp32)
-  KERNEL_MPS2(softmax, int, fp32)
-  KERNEL_MPS2(softmax, Dimname, fp32)
-  KERNEL_MPS2(log_softmax, int, fp32)
-  KERNEL_MPS2(log_softmax, Dimname, fp32)
-  KERNEL_MPS(cumprod, fp32)
-  KERNEL_MPS2(cumprod, dimname, fp32)
-  KERNEL_MPS(cumsum, fp32)
-  KERNEL_MPS2(cumsum, dimname, fp32)
-  KERNEL_MPS(linalg_vector_norm, fp32)
-  KERNEL_MPS(linalg_matrix_norm, fp32)
-  KERNEL_MPS2(linalg_matrix_norm, str_ord, fp32)
-  KERNEL_MPS(sum, fp32)
-  KERNEL_MPS2(sum, dim_IntList, fp32)
-  KERNEL_MPS2(sum, dim_DimnameList, fp32)
-  //
-  // promote
-  KERNEL_MPS(addcdiv, promote)
-  KERNEL_MPS(addcmul, promote)
-  KERNEL_MPS(atan2, promote)
-  KERNEL_MPS(bilinear, promote)
-  KERNEL_MPS(cross, promote)
-  KERNEL_MPS(dot, promote)
-  KERNEL_MPS(grid_sampler, promote)
-  KERNEL_MPS(index_put, promote)
-  KERNEL_MPS(tensordot, promote)
-  KERNEL_MPS(scatter_add, promote)
-}
-
 TORCH_LIBRARY_IMPL(_, AutocastCPU, m) {
  m.fallback(torch::CppFunction::makeFallthrough());
 }
--- a/aten/src/ATen/autocast_mode.h
+++ b/aten/src/ATen/autocast_mode.h
@ -145,8 +145,6 @@ inline bool is_autocast_eligible(
      return tensor.is_xla() && tensor.is_floating_point();
    case c10::DeviceType::PrivateUse1:
      return tensor.is_privateuseone() && tensor.is_floating_point();
-    case c10::DeviceType::MPS:
-      return tensor.is_mps() && tensor.is_floating_point();
    default:
      return false;
  }
@ -170,8 +168,6 @@ inline DispatchKey get_autocast_dispatch_key_from_device_type(
      return DispatchKey::AutocastXLA;
    case c10::DeviceType::PrivateUse1:
      return DispatchKey::AutocastPrivateUse1;
-    case c10::DeviceType::MPS:
-      return DispatchKey::AutocastMPS;
    default:
      throw std::runtime_error(
          "unknown device type for autocast in get_autocast_dispatch_key_from_device_type");
@ -182,7 +178,7 @@ inline bool is_autocast_available(c10::DeviceType device_type) {
  if (device_type == at::kCPU || device_type == at::kCUDA ||
      device_type == at::kXPU || device_type == at::kIPU ||
      device_type == at::kHPU || device_type == at::kXLA ||
-      device_type == at::kPrivateUse1 || device_type == at::kMPS) {
+      device_type == at::kPrivateUse1) {
    return true;
  } else {
    return false;
@ -749,27 +745,6 @@ copy pasted in from VariableTypeEverything.cpp with appropriate substitutions.
      REDISPATCH_SIGNATURE,                                  \
      POLICY)

-// KERNEL_MPS registration for AutocastMPS
-#define KERNEL_MPS(OP, POLICY)            \
-  m.impl(                                 \
-      TORCH_SELECTIVE_NAME("aten::" #OP), \
-      &WrapFunction<                      \
-          CastPolicy::POLICY,             \
-          DeviceType::MPS,                \
-          decltype(ATEN_FN(OP)),          \
-          decltype(ATEN_FN(OP)),          \
-          &ATEN_FN(OP)>::type::call);
-
-#define KERNEL_MPS2(OP, OVERLOAD, POLICY)               \
-  m.impl(                                               \
-      TORCH_SELECTIVE_NAME("aten::" #OP "." #OVERLOAD), \
-      &WrapFunction<                                    \
-          CastPolicy::POLICY,                           \
-          DeviceType::MPS,                              \
-          decltype(ATEN_FN2(OP, OVERLOAD)),             \
-          decltype(ATEN_FN2(OP, OVERLOAD)),             \
-          &ATEN_FN2(OP, OVERLOAD)>::type::call);
-
 // Op lists for different policies.
 // To make sure other backends can reuse the policy op list.
 #define AT_FORALL_LOWER_PRECISION_FP(_)  \
--- a/aten/src/ATen/core/function_schema.cpp
+++ b/aten/src/ATen/core/function_schema.cpp
@ -197,411 +197,4 @@ bool FunctionSchema::may_contain_alias(const SchemaArgument& lhs, const SchemaAr
    return rhsWildcard || canAliasTypeSetsAlias(lhsContainedTypes, rhsContainedTypes);
  }
 }
-
-std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema) {
-  // eventually this should look almost identical to python arg parser, but
-  // it is simpler for now to work directly on this schema
-
-  out << schema.name();
-  if (!schema.overload_name().empty()) {
-    out << "." << schema.overload_name();
-  }
-  out << "(";
-
-  bool seen_kwarg_only = false;
-  for (const auto i : c10::irange(schema.arguments().size())) {
-    if (i > 0) out << ", ";
-    if (schema.arguments()[i].kwarg_only() && !seen_kwarg_only) {
-      out << "*, ";
-      seen_kwarg_only = true;
-    }
-    out << schema.arguments()[i];
-  }
-
-  if(schema.is_vararg()) {
-    if(!schema.arguments().empty())
-      out << ", ";
-    out << "...";
-  }
-
-  out << ") -> ";
-
-  const auto& returns = schema.returns();
-
-  /*
-   * We should skip parenthesis if we return a single item and it's not varret,
-   * or we return nothing but varret.
-   *
-   * Need special handling for schema
-   *   aten::items.str(Dict(str, t) self) -> (str,t)[]
-   * Even though this schema returns a single item, we need add parenthesis.
-   * The is necessary so the printed schema can be parsed by the C++ SchemaParser
-   * Without the extra parenthesis, the parser sees the first parenthesis in '(str,t)' and mistakenly
-   * treat the return type as a tuple. An alternative is to enhance the Lexer
-   * to lookahead multiple tokens to accurately decide if the return type is
-   * a tuple.
-   */
-  bool need_paren = !(
-    (returns.size() == 1 && !schema.is_varret()) ||
-    (returns.empty() && schema.is_varret()));
-
-  if (returns.size() == 1 && !schema.is_varret()) {
-    std::stringstream return_ss;
-    return_ss << returns.at(0);
-    auto return_str = return_ss.str();
-
-    // enclosing the single return item with parenthesis if the return type
-    // starts with a left parenthesis.
-    //
-    // There are 2 cases
-    // 1. something like 'aten::items.str(Dict(str, t) self) -> ((str, t)[])'.
-    // without the extra parenthesis, the c++ schem parser can not parse it.
-    // 2. something like '-> ((str, str))'. Need extra parenthesis so the return
-    // type is a single tuple rather than two strings.
-    // PR (https://github.com/pytorch/pytorch/pull/23204) has more context about
-    // this. test_serialize_and_deserialize (https://github.com/pytorch/pytorch/blob/master/test/test_function_schema.py#L15)
-    // also covers this case.
-    if (!return_str.empty() && return_str.front() == '(') {
-      need_paren = true;
-    }
-  }
-
-  if (need_paren) {
-    out << "(";
-  }
-  for (const auto i : c10::irange(returns.size())) {
-    if (i > 0) {
-      out << ", ";
-    }
-    out << returns.at(i);
-  }
-  if (schema.is_varret()) {
-    if (!returns.empty()) {
-      out << ", ";
-    }
-    out << "...";
-  }
-  if (need_paren) {
-    out << ")";
-  }
-  return out;
-}
-
-static size_t findFirstOutArg(const std::vector<Argument>& args) {
-  // find the start of out args in the schema
-  for (const auto out_start_idx : c10::irange(args.size())) {
-    if (args.at(out_start_idx).is_out()) {
-      return out_start_idx;
-    }
-  }
-  return args.size();
-}
-
-bool Argument::isBackwardCompatibleWith(
-      const Argument& old,
-      std::ostream* why_not) const {
-    const Argument* lhs = this;
-    const Argument* rhs = &old;
-    if (!(lhs->name() == rhs->name()
-        && lhs->N() == rhs->N()
-          && (lhs->alias_info() == rhs->alias_info()
-              || (lhs->alias_info() != nullptr && rhs->alias_info() != nullptr
-                  && *lhs->alias_info() == *rhs->alias_info())))) {
-      return false;
-    }
-    if (lhs->kwarg_only() && !rhs->kwarg_only()) {
-      return false;
-    }
-    if (!rhs->type()->isSubtypeOfExt(*lhs->type(), why_not)) {
-      return false;
-    }
-    if (rhs->default_value().has_value() &&
-        lhs->default_value() != rhs->default_value()) {
-      return false;
-    }
-    return true;
-}
-
-bool Argument::isForwardCompatibleWith(
-    const Argument& old,
-    std::ostream* why_not) const {
-  const Argument* lhs = this;
-  const Argument* rhs = &old;
-  if (!(lhs->name() == rhs->name()
-      && lhs->N() == rhs->N()
-        && (lhs->alias_info() == rhs->alias_info()
-            || (lhs->alias_info() != nullptr && rhs->alias_info() != nullptr
-                && *lhs->alias_info() == *rhs->alias_info())))) {
-    return false;
-  }
-  if (lhs->kwarg_only() && !rhs->kwarg_only()) {
-    return false;
-  }
-  if (!lhs->type()->isSubtypeOfExt(rhs->type(), why_not)) {
-    return false;
-  }
-  if (rhs->default_value().has_value() &&
-      lhs->default_value() != rhs->default_value()) {
-    return false;
-  }
-  if (lhs->default_value().has_value() && !rhs->default_value().has_value()) {
-    return false;
-  }
-  return true;
-}
-
-std::string FunctionSchema::formatTypeMismatchMsg(
-    const Argument& expected,
-    const std::string& actual_type,
-    std::optional<size_t> position,
-    std::optional<std::string> value) const {
-  std::string position_str;
-  if (position) {
-    position_str = c10::str("Position: ", *position, "\n");
-  }
-  std::string value_str;
-  if (value) {
-    value_str = c10::str("Value: ", *value, "\n");
-  }
-  return c10::str(
-      name(),
-      "() ",
-      expected.formatTypeMismatchMsg(actual_type),
-      position_str,
-      value_str,
-      "Declaration: ",
-      *this);
-}
-
-bool FunctionSchema::isBackwardCompatibleWith(
-    const FunctionSchema& old,
-    std::ostream* why_not) const {
-  if (!(name() == old.name()
-        && overload_name() == old.overload_name()
-        // we are conservative on is_vararg and is_varret,
-        // since they are only used by internal operators
-        && is_vararg() == old.is_vararg()
-        && is_varret() == old.is_varret()
-        && returns().size() == old.returns().size()
-        && arguments().size() >= old.arguments().size())) {
-    return false;
-  }
-  for (const auto i : c10::irange(returns().size())) {
-    // Backwards compatibility requires covariance on argument types
-    // (i.e. more generic), and contravariance on return types (i.e.
-    //  more specific).
-    if (!old.returns().at(i).isBackwardCompatibleWith(
-          returns().at(i),
-          why_not)) {
-      return false;
-    }
-  }
-
-  // we want to test both out and default args separately
-  size_t old_out_start_idx = findFirstOutArg(old.arguments());
-  size_t new_out_start_idx = findFirstOutArg(arguments());
-
-  // make sure among the default args, they are backward compatible
-  for (const auto i : c10::irange(old_out_start_idx)) {
-    if (!arguments().at(i).isBackwardCompatibleWith(
-          old.arguments().at(i), why_not)) {
-      return false;
-    }
-  }
-
-  // Validate that all new arguments provided has a default value
-  for (const auto i : c10::irange(old_out_start_idx, new_out_start_idx)) {
-    if (!arguments().at(i).default_value()) {
-      if (why_not) {
-        *why_not
-            << "Function schema not backward compatible since the new argument '"
-            << arguments().at(i).name() << "' of type "
-            << arguments().at(i).type()->str()
-            << " did not provide a default value.";
-      }
-      return false;
-    }
-  }
-
-  // now compare the out args
-  for (const auto i : c10::irange(old_out_start_idx, old.arguments().size())) {
-    if (!arguments()
-             .at(i - old_out_start_idx + new_out_start_idx)
-             .isBackwardCompatibleWith(old.arguments().at(i), why_not)) {
-      return false;
-    }
-  }
-
-  return true;
-}
-
-bool FunctionSchema::isForwardCompatibleWith(
-    const FunctionSchema& old,
-    std::ostringstream& why_not) const {
-  if (!(name() == old.name() &&
-        overload_name() == old.overload_name()
-        // we are conservative on is_vararg and is_varret,
-        // since they are only used by internal operators
-        && is_vararg() == old.is_vararg() && is_varret() == old.is_varret() &&
-        returns().size() == old.returns().size())) {
-    return false;
-  }
-
-  // we want to test both out and default args separately
-  size_t old_out_start_idx = findFirstOutArg(old.arguments());
-  size_t new_out_start_idx = findFirstOutArg(arguments());
-
-  if (old.arguments().size() - old_out_start_idx !=
-      arguments().size() - new_out_start_idx) {
-    if (why_not) {
-      why_not << "Function schema should have the "
-              << "same number of out arguments";
-    }
-    return false;
-  }
-
-  // make sure among the default args, they are forward compatible
-  for (size_t i = 0; i < std::min(old_out_start_idx, new_out_start_idx); i++) {
-    if (!arguments().at(i).isForwardCompatibleWith(old.arguments().at(i))) {
-      if (why_not) {
-        why_not
-            << "'" << arguments().at(i).name() << "'"
-            << " is not forward compatible with the older version of the schema";
-      }
-      return false;
-    }
-  }
-
-  // Validate that all new arguments provided has a default value
-  for (size_t i = old_out_start_idx; i < new_out_start_idx; ++i) {
-    if (!arguments().at(i).default_value()) {
-      if (why_not) {
-        why_not
-            << "Function schema is not forward compatible since the new argument '"
-            << arguments().at(i).name() << "' of type "
-            << arguments().at(i).type()->str()
-            << " did not provide a default value.";
-      }
-      return false;
-    }
-
-    auto default_val = arguments().at(i).default_value().value();
-    if (default_val.isList() || default_val.isGenericDict()) {
-      if (why_not) {
-        why_not
-            << "Function schema is not forward compatible since the new argument '"
-            << arguments().at(i).name() << "' of type "
-            << arguments().at(i).type()->str() << " has a container type "
-            << "as its default value.";
-      }
-      return false;
-    }
-  }
-
-  // now compare the out args
-  for (size_t i = old_out_start_idx; i < old.arguments().size(); i++) {
-    if (!arguments()
-             .at(i - old_out_start_idx + new_out_start_idx)
-             .isForwardCompatibleWith(old.arguments().at(i))) {
-      if (why_not) {
-        why_not << "Out argument '"
-                << "'" << arguments().at(i).name()
-                << " is not FC with the older version of the schema";
-      }
-      return false;
-    }
-  }
-
-  return true;
-}
-
-std::string FunctionSchema::findErrorInKwargs(const std::vector<std::string>& kwargs) const {
-  // First check if any of the kwargs are unknown, i.e. don't match the name of
-  // any argument in the schema.
-  for (const auto& kwarg : kwargs) {
-    if (!std::count_if(
-            arguments().begin(),
-            arguments().end(),
-            [&kwarg](const Argument& argument) {
-              return argument.name() == kwarg;
-            })) {
-      return c10::str(
-          "Unknown keyword argument '",
-          kwarg,
-          "' for operator '",
-          name(),
-          "'. Schema: ",
-          *this);
-    }
-  }
-  // If there are unconsumed kwargs but none of them were unknown, the first
-  // positional argument present in the kwargs is duplicated.
-  for (const auto& argument : arguments()) {
-    if (std::find(kwargs.begin(), kwargs.end(), argument.name()) != kwargs.end()) {
-      AT_ASSERT(!argument.default_value());
-      return c10::str(
-          "Argument '",
-          argument.name(),
-          "' specified both as positional and ",
-          "keyword argument. Schema: ",
-          *this);
-    }
-  }
-  return "";
-}
-
-
-FunctionSchema FunctionSchema::cloneWithRemappedTypes(
-    const std::function<TypePtr(TypePtr)> type_map) const {
-  auto update_args = [&](const std::vector<Argument>& args) {
-    std::vector<Argument> new_args;
-    new_args.reserve(args.size());
-    for(const Argument& arg : args) {
-      new_args.emplace_back(arg.cloneWithType(type_map(arg.type())));
-    }
-    return new_args;
-  };
-  return FunctionSchema(
-      name(),
-      overload_name(),
-      update_args(arguments()),
-      update_args(returns()),
-      is_vararg(),
-      is_varret());
-}
-
-// covariant subtyping of list of Arguments
-static bool isSubtypeOfList(
-    ArrayRef<Argument> child,
-    ArrayRef<Argument> parent,
-    std::ostream* why_not) {
-  if (child.size() != parent.size()) {
-    return false;
-  }
-  for (const auto i : c10::irange(child.size())) {
-    const Argument& c = child[i];
-    const Argument& p = parent[i];
-    if (c.name() != p.name()) {
-      return false;
-    }
-    if (!c.type()->isSubtypeOfExt(*p.type(), why_not)) {
-      return false;
-    }
-  }
-  return true;
-}
-
-bool FunctionSchema::isSubtypeOf(
-    const FunctionSchema& rhs,
-    bool as_method,
-    std::ostream* why_not) const {
-  size_t start = as_method ? 1 : 0;
-  // functions are contravariant in arguments but covariant in returns
-  return isSubtypeOfList(
-             ArrayRef<Argument>(rhs.arguments()).slice(start),
-             ArrayRef<Argument>(arguments()).slice(start),
-             why_not) &&
-      isSubtypeOfList(returns(), rhs.returns(), why_not);
-}
-
 } // namespace c10
--- a/aten/src/ATen/core/function_schema.h
+++ b/aten/src/ATen/core/function_schema.h
@ -25,7 +25,7 @@ using AliasTypeSet = std::vector<TypePtr>;

 bool operator==(const Argument& lhs, const Argument& rhs);

-struct TORCH_API Argument {
+struct Argument {
  Argument(
      std::string name = "",
      const TypePtr& type = nullptr,
@ -622,7 +622,7 @@ inline std::ostream& operator<<(std::ostream& out, const Argument& arg) {
  return out;
 }

-TORCH_API std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema);
+inline std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema);

 inline std::string toString(const FunctionSchema& schema) {
  std::ostringstream str;
--- a/aten/src/ATen/core/function_schema_inl.h
+++ b/aten/src/ATen/core/function_schema_inl.h
@ -2,8 +2,328 @@
 #include <ostream>
 #include <sstream>

+// note: windows build doesn't find symbols in operator files unless
+// this is a header file
+
 namespace c10 {

+inline std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema) {
+  // eventually this should look almost identical to python arg parser, but
+  // it is simpler for now to work directly on this schema
+
+  out << schema.name();
+  if (!schema.overload_name().empty()) {
+    out << "." << schema.overload_name();
+  }
+  out << "(";
+
+  bool seen_kwarg_only = false;
+  for (const auto i : c10::irange(schema.arguments().size())) {
+    if (i > 0) out << ", ";
+    if (schema.arguments()[i].kwarg_only() && !seen_kwarg_only) {
+      out << "*, ";
+      seen_kwarg_only = true;
+    }
+    out << schema.arguments()[i];
+  }
+
+  if(schema.is_vararg()) {
+    if(!schema.arguments().empty())
+      out << ", ";
+    out << "...";
+  }
+
+  out << ") -> ";
+
+  const auto& returns = schema.returns();
+
+  /*
+   * We should skip parenthesis if we return a single item and it's not varret,
+   * or we return nothing but varret.
+   *
+   * Need special handling for schema
+   *   aten::items.str(Dict(str, t) self) -> (str,t)[]
+   * Even though this schema returns a single item, we need add parenthesis.
+   * The is necessary so the printed schema can be parsed by the C++ SchemaParser
+   * Without the extra parenthesis, the parser sees the first parenthesis in '(str,t)' and mistakenly
+   * treat the return type as a tuple. An alternative is to enhance the Lexer
+   * to lookahead multiple tokens to accurately decide if the return type is
+   * a tuple.
+   */
+  bool need_paren = !(
+    (returns.size() == 1 && !schema.is_varret()) ||
+    (returns.empty() && schema.is_varret()));
+
+  if (returns.size() == 1 && !schema.is_varret()) {
+    std::stringstream return_ss;
+    return_ss << returns.at(0);
+    auto return_str = return_ss.str();
+
+    // enclosing the single return item with parenthesis if the return type
+    // starts with a left parenthesis.
+    //
+    // There are 2 cases
+    // 1. something like 'aten::items.str(Dict(str, t) self) -> ((str, t)[])'.
+    // without the extra parenthesis, the c++ schem parser can not parse it.
+    // 2. something like '-> ((str, str))'. Need extra parenthesis so the return
+    // type is a single tuple rather than two strings.
+    // PR (https://github.com/pytorch/pytorch/pull/23204) has more context about
+    // this. test_serialize_and_deserialize (https://github.com/pytorch/pytorch/blob/master/test/test_function_schema.py#L15)
+    // also covers this case.
+    if (!return_str.empty() && return_str.front() == '(') {
+      need_paren = true;
+    }
+  }
+
+  if (need_paren) {
+    out << "(";
+  }
+  for (const auto i : c10::irange(returns.size())) {
+    if (i > 0) {
+      out << ", ";
+    }
+    out << returns.at(i);
+  }
+  if (schema.is_varret()) {
+    if (!returns.empty()) {
+      out << ", ";
+    }
+    out << "...";
+  }
+  if (need_paren) {
+    out << ")";
+  }
+  return out;
+}
+
+inline size_t findFirstOutArg(const std::vector<Argument>& args) {
+  // find the start of out args in the schema
+  for (const auto out_start_idx : c10::irange(args.size())) {
+    if (args.at(out_start_idx).is_out()) {
+      return out_start_idx;
+    }
+  }
+  return args.size();
+}
+
+inline bool Argument::isBackwardCompatibleWith(
+      const Argument& old,
+      std::ostream* why_not) const {
+    const Argument* lhs = this;
+    const Argument* rhs = &old;
+    if (!(lhs->name() == rhs->name()
+        && lhs->N() == rhs->N()
+          && (lhs->alias_info() == rhs->alias_info()
+              || (lhs->alias_info() != nullptr && rhs->alias_info() != nullptr
+                  && *lhs->alias_info() == *rhs->alias_info())))) {
+      return false;
+    }
+    if (lhs->kwarg_only() && !rhs->kwarg_only()) {
+      return false;
+    }
+    if (!rhs->type()->isSubtypeOfExt(*lhs->type(), why_not)) {
+      return false;
+    }
+    if (rhs->default_value().has_value() &&
+        lhs->default_value() != rhs->default_value()) {
+      return false;
+    }
+    return true;
+}
+
+inline bool Argument::isForwardCompatibleWith(
+    const Argument& old,
+    std::ostream* why_not) const {
+  const Argument* lhs = this;
+  const Argument* rhs = &old;
+  if (!(lhs->name() == rhs->name()
+      && lhs->N() == rhs->N()
+        && (lhs->alias_info() == rhs->alias_info()
+            || (lhs->alias_info() != nullptr && rhs->alias_info() != nullptr
+                && *lhs->alias_info() == *rhs->alias_info())))) {
+    return false;
+  }
+  if (lhs->kwarg_only() && !rhs->kwarg_only()) {
+    return false;
+  }
+  if (!lhs->type()->isSubtypeOfExt(rhs->type(), why_not)) {
+    return false;
+  }
+  if (rhs->default_value().has_value() &&
+      lhs->default_value() != rhs->default_value()) {
+    return false;
+  }
+  if (lhs->default_value().has_value() && !rhs->default_value().has_value()) {
+    return false;
+  }
+  return true;
+}
+
+inline std::string FunctionSchema::formatTypeMismatchMsg(
+    const Argument& expected,
+    const std::string& actual_type,
+    std::optional<size_t> position,
+    std::optional<std::string> value) const {
+  std::string position_str;
+  if (position) {
+    position_str = c10::str("Position: ", *position, "\n");
+  }
+  std::string value_str;
+  if (value) {
+    value_str = c10::str("Value: ", *value, "\n");
+  }
+  return c10::str(
+      name(),
+      "() ",
+      expected.formatTypeMismatchMsg(actual_type),
+      position_str,
+      value_str,
+      "Declaration: ",
+      *this);
+}
+
+inline bool FunctionSchema::isBackwardCompatibleWith(
+    const FunctionSchema& old,
+    std::ostream* why_not) const {
+  if (!(name() == old.name()
+        && overload_name() == old.overload_name()
+        // we are conservative on is_vararg and is_varret,
+        // since they are only used by internal operators
+        && is_vararg() == old.is_vararg()
+        && is_varret() == old.is_varret()
+        && returns().size() == old.returns().size()
+        && arguments().size() >= old.arguments().size())) {
+    return false;
+  }
+  for (const auto i : c10::irange(returns().size())) {
+    // Backwards compatibility requires covariance on argument types
+    // (i.e. more generic), and contravariance on return types (i.e.
+    //  more specific).
+    if (!old.returns().at(i).isBackwardCompatibleWith(
+          returns().at(i),
+          why_not)) {
+      return false;
+    }
+  }
+
+  // we want to test both out and default args separately
+  size_t old_out_start_idx = findFirstOutArg(old.arguments());
+  size_t new_out_start_idx = findFirstOutArg(arguments());
+
+  // make sure among the default args, they are backward compatible
+  for (const auto i : c10::irange(old_out_start_idx)) {
+    if (!arguments().at(i).isBackwardCompatibleWith(
+          old.arguments().at(i), why_not)) {
+      return false;
+    }
+  }
+
+  // Validate that all new arguments provided has a default value
+  for (const auto i : c10::irange(old_out_start_idx, new_out_start_idx)) {
+    if (!arguments().at(i).default_value()) {
+      if (why_not) {
+        *why_not
+            << "Function schema not backward compatible since the new argument '"
+            << arguments().at(i).name() << "' of type "
+            << arguments().at(i).type()->str()
+            << " did not provide a default value.";
+      }
+      return false;
+    }
+  }
+
+  // now compare the out args
+  for (const auto i : c10::irange(old_out_start_idx, old.arguments().size())) {
+    if (!arguments()
+             .at(i - old_out_start_idx + new_out_start_idx)
+             .isBackwardCompatibleWith(old.arguments().at(i), why_not)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+inline bool FunctionSchema::isForwardCompatibleWith(
+    const FunctionSchema& old,
+    std::ostringstream& why_not) const {
+  if (!(name() == old.name() &&
+        overload_name() == old.overload_name()
+        // we are conservative on is_vararg and is_varret,
+        // since they are only used by internal operators
+        && is_vararg() == old.is_vararg() && is_varret() == old.is_varret() &&
+        returns().size() == old.returns().size())) {
+    return false;
+  }
+
+  // we want to test both out and default args separately
+  size_t old_out_start_idx = findFirstOutArg(old.arguments());
+  size_t new_out_start_idx = findFirstOutArg(arguments());
+
+  if (old.arguments().size() - old_out_start_idx !=
+      arguments().size() - new_out_start_idx) {
+    if (why_not) {
+      why_not << "Function schema should have the "
+              << "same number of out arguments";
+    }
+    return false;
+  }
+
+  // make sure among the default args, they are forward compatible
+  for (size_t i = 0; i < std::min(old_out_start_idx, new_out_start_idx); i++) {
+    if (!arguments().at(i).isForwardCompatibleWith(old.arguments().at(i))) {
+      if (why_not) {
+        why_not
+            << "'" << arguments().at(i).name() << "'"
+            << " is not forward compatible with the older version of the schema";
+      }
+      return false;
+    }
+  }
+
+  // Validate that all new arguments provided has a default value
+  for (size_t i = old_out_start_idx; i < new_out_start_idx; ++i) {
+    if (!arguments().at(i).default_value()) {
+      if (why_not) {
+        why_not
+            << "Function schema is not forward compatible since the new argument '"
+            << arguments().at(i).name() << "' of type "
+            << arguments().at(i).type()->str()
+            << " did not provide a default value.";
+      }
+      return false;
+    }
+
+    auto default_val = arguments().at(i).default_value().value();
+    if (default_val.isList() || default_val.isGenericDict()) {
+      if (why_not) {
+        why_not
+            << "Function schema is not forward compatible since the new argument '"
+            << arguments().at(i).name() << "' of type "
+            << arguments().at(i).type()->str() << " has a container type "
+            << "as its default value.";
+      }
+      return false;
+    }
+  }
+
+  // now compare the out args
+  for (size_t i = old_out_start_idx; i < old.arguments().size(); i++) {
+    if (!arguments()
+             .at(i - old_out_start_idx + new_out_start_idx)
+             .isForwardCompatibleWith(old.arguments().at(i))) {
+      if (why_not) {
+        why_not << "Out argument '"
+                << "'" << arguments().at(i).name()
+                << " is not FC with the older version of the schema";
+      }
+      return false;
+    }
+  }
+
+  return true;
+}
+
 template<typename T>
 inline void FunctionSchema::checkArg(
    const IValue& value,
@ -21,6 +341,41 @@ inline void FunctionSchema::checkArg(
  }
 }

+inline std::string FunctionSchema::findErrorInKwargs(const std::vector<std::string>& kwargs) const {
+  // First check if any of the kwargs are unknown, i.e. don't match the name of
+  // any argument in the schema.
+  for (const auto& kwarg : kwargs) {
+    if (!std::count_if(
+            arguments().begin(),
+            arguments().end(),
+            [&kwarg](const Argument& argument) {
+              return argument.name() == kwarg;
+            })) {
+      return c10::str(
+          "Unknown keyword argument '",
+          kwarg,
+          "' for operator '",
+          name(),
+          "'. Schema: ",
+          *this);
+    }
+  }
+  // If there are unconsumed kwargs but none of them were unknown, the first
+  // positional argument present in the kwargs is duplicated.
+  for (const auto& argument : arguments()) {
+    if (std::find(kwargs.begin(), kwargs.end(), argument.name()) != kwargs.end()) {
+      AT_ASSERT(!argument.default_value());
+      return c10::str(
+          "Argument '",
+          argument.name(),
+          "' specified both as positional and ",
+          "keyword argument. Schema: ",
+          *this);
+    }
+  }
+  return "";
+}
+
 template <typename T>
 inline void FunctionSchema::checkAndNormalizeInputs(
    std::vector<IValue>& inputs,
@ -72,4 +427,57 @@ inline void FunctionSchema::checkAndNormalizeInputs(
  }
 }

+inline FunctionSchema FunctionSchema::cloneWithRemappedTypes(
+    const std::function<TypePtr(TypePtr)> type_map) const {
+  auto update_args = [&](const std::vector<Argument>& args) {
+    std::vector<Argument> new_args;
+    new_args.reserve(args.size());
+    for(const Argument& arg : args) {
+      new_args.emplace_back(arg.cloneWithType(type_map(arg.type())));
+    }
+    return new_args;
+  };
+  return FunctionSchema(
+      name(),
+      overload_name(),
+      update_args(arguments()),
+      update_args(returns()),
+      is_vararg(),
+      is_varret());
+}
+
+// covariant subtyping of list of Arguments
+inline bool isSubtypeOfList(
+    ArrayRef<Argument> child,
+    ArrayRef<Argument> parent,
+    std::ostream* why_not) {
+  if (child.size() != parent.size()) {
+    return false;
+  }
+  for (const auto i : c10::irange(child.size())) {
+    const Argument& c = child[i];
+    const Argument& p = parent[i];
+    if (c.name() != p.name()) {
+      return false;
+    }
+    if (!c.type()->isSubtypeOfExt(*p.type(), why_not)) {
+      return false;
+    }
+  }
+  return true;
+}
+
+inline bool FunctionSchema::isSubtypeOf(
+    const FunctionSchema& rhs,
+    bool as_method,
+    std::ostream* why_not) const {
+  size_t start = as_method ? 1 : 0;
+  // functions are contravariant in arguments but covariant in returns
+  return isSubtypeOfList(
+             ArrayRef<Argument>(rhs.arguments()).slice(start),
+             ArrayRef<Argument>(arguments()).slice(start),
+             why_not) &&
+      isSubtypeOfList(returns(), rhs.returns(), why_not);
+}
+
 } // namespace c10
--- a/aten/src/ATen/core/interned_strings.h
+++ b/aten/src/ATen/core/interned_strings.h
@ -228,7 +228,6 @@ namespace c10 {
  _(aten, is_autocast_cpu_enabled)   \
  _(aten, is_autocast_xla_enabled)   \
  _(aten, get_autocast_dtype)        \
-  _(aten, is_autocast_mps_enabled)   \
  FORALL_ATEN_BASE_SYMBOLS(_)        \
  _(onnx, Add)                       \
  _(onnx, Concat)                    \
--- a/aten/src/ATen/cpu/Utils.cpp
+++ b/aten/src/ATen/cpu/Utils.cpp
@ -9,7 +9,7 @@
 #endif

 namespace at::cpu {
-bool is_avx2_supported() {
+bool is_cpu_support_avx2() {
 #if !defined(__s390x__) && !defined(__powerpc__)
  return cpuinfo_initialize() && cpuinfo_has_x86_avx2();
 #else
@ -17,7 +17,7 @@ bool is_avx2_supported() {
 #endif
 }

-bool is_avx512_supported() {
+bool is_cpu_support_avx512() {
 #if !defined(__s390x__) && !defined(__powerpc__)
  return cpuinfo_initialize() && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512vl() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq();
 #else
@ -25,7 +25,7 @@ bool is_avx512_supported() {
 #endif
 }

-bool is_avx512_vnni_supported() {
+bool is_cpu_support_avx512_vnni() {
 #if !defined(__s390x__) && !defined(__powerpc__)
  return cpuinfo_initialize() && cpuinfo_has_x86_avx512vnni();
 #else
@ -33,15 +33,7 @@ bool is_avx512_vnni_supported() {
 #endif
 }

-bool is_avx512_bf16_supported() {
-#if !defined(__s390x__) && !defined(__powerpc__)
-  return cpuinfo_initialize() && cpuinfo_has_x86_avx512bf16();
-#else
-  return false;
-#endif
-}
-
-bool is_amx_tile_supported() {
+bool is_cpu_support_amx_tile() {
 #if !defined(__s390x__) && !defined(__powerpc__)
  return cpuinfo_initialize() && cpuinfo_has_x86_amx_tile();
 #else
@ -50,7 +42,7 @@ bool is_amx_tile_supported() {
 }

 bool init_amx() {
-  if (!is_amx_tile_supported()) {
+  if (!is_cpu_support_amx_tile()) {
    return false;
  }

--- a/aten/src/ATen/cpu/Utils.h
+++ b/aten/src/ATen/cpu/Utils.h
@ -6,17 +6,14 @@

 namespace at::cpu {

-TORCH_API bool is_avx2_supported();
-TORCH_API bool is_avx512_supported();
+TORCH_API bool is_cpu_support_avx2();
+TORCH_API bool is_cpu_support_avx512();

 // Detect if CPU support Vector Neural Network Instruction.
-TORCH_API bool is_avx512_vnni_supported();
-
-// Detect if CPU supports AVX512_BF16 ISA
-TORCH_API bool is_avx512_bf16_supported();
+TORCH_API bool is_cpu_support_avx512_vnni();

 // Detect if CPU support Advanced Matrix Extension.
-TORCH_API bool is_amx_tile_supported();
+TORCH_API bool is_cpu_support_amx_tile();

 // Enable the system to use AMX instructions.
 TORCH_API bool init_amx();
--- a/aten/src/ATen/cpu/vec/vec256/vec256_convert.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_convert.h
@ -43,26 +43,6 @@ struct VecConvert<BFloat16, 1, float, 1> {
  }
 };

-template <>
-struct VecConvert<BFloat16, 1, float, 2> {
-  static inline VectorizedN<BFloat16, 1> apply(
-      const VectorizedN<float, 2>& src) {
-    VectorizedN<BFloat16, 1> result;
-    result[0] = convert_float_bfloat16(src[0], src[1]);
-    return result;
-  }
-};
-
-template <>
-struct VecConvert<float, 2, BFloat16, 1> {
-  static inline VectorizedN<float, 2> apply(
-      const VectorizedN<BFloat16, 1>& src) {
-    VectorizedN<float, 2> result;
-    std::tie(result[0], result[1]) = convert_bfloat16_float(src[0]);
-    return result;
-  }
-};
-
 template <>
 struct VecConvert<Half, 1, float, 1> {
  static inline VectorizedN<Half, 1> apply(const VectorizedN<float, 1>& src) {
@ -72,24 +52,6 @@ struct VecConvert<Half, 1, float, 1> {
  }
 };

-template <>
-struct VecConvert<Half, 1, float, 2> {
-  static inline VectorizedN<Half, 1> apply(const VectorizedN<float, 2>& src) {
-    VectorizedN<Half, 1> result;
-    result[0] = convert_float_half(src[0], src[1]);
-    return result;
-  }
-};
-
-template <>
-struct VecConvert<float, 2, Half, 1> {
-  static inline VectorizedN<float, 2> apply(const VectorizedN<Half, 1>& src) {
-    VectorizedN<float, 2> result;
-    std::tie(result[0], result[1]) = convert_half_float(src[0]);
-    return result;
-  }
-};
-
 template <>
 inline Vectorized<double> convert_to_fp_of_same_size<double>(
    const Vectorized<int64_t>& src);
--- a/aten/src/ATen/cpu/vec/vec256/vec256_float.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_float.h
@ -636,21 +636,6 @@ inline void transpose_mxn<float, 8, 8>(
  _mm256_storeu_ps(&dst[7 * ld_dst], th);
 }

-template<>
-inline void transpose_mxn<float, 16, 16>(
-    const float* src,
-    int64_t ld_src,
-    float* dst,
-    int64_t ld_dst) {
-  transpose_mxn<float, 8, 8>(
-          src , ld_src, dst, ld_dst);
-  transpose_mxn<float, 8, 8>(
-          src + 8, ld_src, dst + 8 * ld_dst, ld_dst);
-  transpose_mxn<float, 8, 8>(
-          src + 8 * ld_src, ld_src, dst + 8, ld_dst);
-  transpose_mxn<float, 8, 8>(
-          src + 8 * ld_src + 8, ld_src, dst + 8 * ld_dst + 8, ld_dst);
-}
 #endif

 }} // namespace at::vec::CPU_CAPABILITY
--- a/aten/src/ATen/cpu/vec/vec256/vec256_mask.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_mask.h
@ -9,218 +9,49 @@ inline namespace CPU_CAPABILITY {

 #if defined(CPU_CAPABILITY_AVX2) && !defined(_MSC_VER)

-template <typename T, int dst_n, typename mask_t, int mask_n>
-struct VecMaskLoad<
-    T,
-    dst_n,
-    mask_t,
-    mask_n,
-    typename std::enable_if_t<
-        (mask_n == dst_n * 2 && dst_n >= 1) &&
-            (std::is_same_v<T, float> || std::is_same_v<T, int32_t>),
-        void>> {
-  static inline VectorizedN<T, dst_n> apply(
-      const T* ptr,
-      const VecMask<mask_t, mask_n>& vec_mask) {
-    VectorizedN<mask_t, 2> tmp_vec;
-    VectorizedN<T, dst_n> result;
-    for (int i = 0; i < dst_n; i++) {
-      tmp_vec[0] = vec_mask[2 * i];
-      tmp_vec[1] = vec_mask[2 * i + 1];
-      auto int64_mask = VecMask<mask_t, 2>(tmp_vec).template cast<int64_t, 2>();
-      auto int_mask = int64_mask.template cast<int, 1>()[0];
-      if constexpr (std::is_same_v<T, float>) {
-        result[i] = Vectorized<T>(
-            _mm256_maskload_ps(ptr + i * Vectorized<T>::size(), int_mask));
-      } else {
-        result[i] = Vectorized<T>(
-            _mm256_maskload_epi32(ptr + i * Vectorized<T>::size(), int_mask));
-      }
-    }
-    return result;
-  }
-};
-
-template <typename T, int dst_n, typename mask_t>
-struct VecMaskLoad<
-    T,
-    dst_n,
-    mask_t,
-    dst_n,
-    typename std::enable_if_t<
-        std::is_same_v<T, float> || std::is_same_v<T, int32_t>,
-        void>> {
-  static inline VectorizedN<T, dst_n> apply(
-      const T* ptr,
-      const VecMask<mask_t, dst_n>& vec_mask) {
-    VectorizedN<T, dst_n> result;
-#ifndef _MSC_VER
-#pragma unroll
-#endif
-    for (int i = 0; i < dst_n; i++) {
-      auto tmp_mask = VecMask<mask_t, 1>(vec_mask[i]);
-      auto int_mask = tmp_mask.template cast<int, 1>()[0];
-      if constexpr (std::is_same_v<T, float>) {
-        result[i] = Vectorized<T>(
-            _mm256_maskload_ps(ptr + i * Vectorized<T>::size(), int_mask));
-      } else {
-        result[i] = Vectorized<T>(
-            _mm256_maskload_epi32(ptr + i * Vectorized<T>::size(), int_mask));
-      }
-    }
-    return result;
-  }
-};
-
 template <typename T, typename mask_t>
 struct VecMaskLoad<
    T,
-    2,
+    1,
    mask_t,
    1,
    typename std::enable_if_t<
-        std::is_same_v<T, int64_t> || std::is_same_v<T, double>>> {
-  static inline VectorizedN<T, 2> apply(
+        std::is_same_v<T, float> || std::is_same_v<T, int32_t> ||
+            std::is_same_v<T, uint32_t>,
+        void>> {
+  static inline VectorizedN<T, 1> apply(
      const T* ptr,
      const VecMask<mask_t, 1>& vec_mask) {
-    auto int64_mask = vec_mask.template cast<int64_t, 2>();
-    auto result = at::vec::VectorizedN<T, 2>();
-    if constexpr (std::is_same_v<T, double>) {
-      result[0] = _mm256_maskload_pd(ptr, int64_mask[0]);
-      result[1] = _mm256_maskload_pd(
-          ptr + at::vec::Vectorized<T>::size(), int64_mask[1]);
+    auto int_mask = vec_mask.template cast<int, 1>()[0];
+    if constexpr (std::is_same_v<T, float>) {
+      return Vectorized<T>(_mm256_maskload_ps(ptr, int_mask));
    } else {
-      result[0] = _mm256_maskload_epi64(
-          reinterpret_cast<const long long*>(ptr), int64_mask[0]);
-      result[1] = _mm256_maskload_epi64(
-          reinterpret_cast<const long long*>(
-              ptr + at::vec::Vectorized<T>::size()),
-          int64_mask[1]);
+      return Vectorized<T>(_mm256_maskload_epi32(ptr, int_mask));
    }
-    return result;
  }
 };

 // TODO: add specialization of VecMaskLoad for bfloat16/half and int8/uint8

-template <int N>
-struct VecMaskCast<float, N, int, N> {
-  static inline VecMask<float, N> apply(const VecMask<int, N>& vec_mask) {
-    VectorizedN<float, N> result;
-#ifndef _MSC_VER
-#pragma unroll
-#endif
-    for (int i = 0; i < N; ++i) {
-      result[i] = _mm256_castsi256_ps(vec_mask[i]);
-    }
-    return result;
-  }
-};
-
-template <int N>
-struct VecMaskCast<int, N, float, N> {
-  static inline VecMask<int, N> apply(const VecMask<float, N>& vec_mask) {
-    VectorizedN<int, N> result;
-#ifndef _MSC_VER
-#pragma unroll
-#endif
-    for (int i = 0; i < N; ++i) {
-      result[i] = _mm256_castps_si256(vec_mask[i]);
-    }
-    return result;
-  }
-};
-
-template <int N>
-struct VecMaskCast<int64_t, N, double, N> {
-  static inline VecMask<int64_t, N> apply(const VecMask<double, N>& vec_mask) {
-    VectorizedN<int64_t, N> result;
-#ifndef _MSC_VER
-#pragma unroll
-#endif
-    for (int i = 0; i < N; ++i) {
-      result[i] = _mm256_castpd_si256(vec_mask[i]);
-    }
-    return result;
-  }
-};
-
-template <int N>
-struct VecMaskCast<double, N, int64_t, N> {
-  static inline VecMask<double, N> apply(const VecMask<int64_t, N>& vec_mask) {
-    VectorizedN<double, N> result;
-#ifndef _MSC_VER
-#pragma unroll
-#endif
-    for (int i = 0; i < N; ++i) {
-      result[i] = _mm256_castsi256_pd(vec_mask[i]);
-    }
-    return result;
-  }
-};
-
-template <int dst_n, typename mask_t, int mask_n>
-struct VecMaskCast<
-    int64_t,
-    dst_n,
-    mask_t,
-    mask_n,
-    typename std::enable_if_t<
-        (dst_n == 2 * mask_n) &&
-            (std::is_same_v<mask_t, float> || std::is_same_v<mask_t, int>),
-        void>> {
-  static inline VecMask<int64_t, dst_n> apply(
-      const VecMask<mask_t, mask_n>& vec_mask) {
-    VectorizedN<int64_t, dst_n> result;
-    auto int_mask = vec_mask.template cast<int, mask_n>();
-#ifndef _MSC_VER
-#pragma unroll
-#endif
-    for (int i = 0; i < mask_n; ++i) {
-      auto int64_vec =
-          convert<int64_t, 2, int, 1>(VectorizedN<int, 1>(int_mask[i]));
-      result[2 * i] = int64_vec[0];
-      result[2 * i + 1] = int64_vec[1];
-    }
-    return VecMask<int64_t, dst_n>(result);
-  }
-};
-
-template <typename dst_t, int dst_n, int mask_n>
-struct VecMaskCast<
-    dst_t,
-    dst_n,
-    int64_t,
-    mask_n,
-    typename std::enable_if_t<
-        (mask_n == 2 * dst_n) &&
-            (std::is_same_v<dst_t, float> || std::is_same_v<dst_t, int>),
-        void>> {
-  static inline VecMask<dst_t, dst_n> apply(
-      const VecMask<int64_t, mask_n>& vec_mask) {
-    VectorizedN<int, dst_n> result;
-    VectorizedN<int64_t, 2> int64_vec;
-    for (int i = 0; i < dst_n; ++i) {
-      int64_vec[0] = vec_mask[2 * i];
-      int64_vec[1] = vec_mask[2 * i + 1];
-      result[i] = convert<int, 1, int64_t, 2>(int64_vec);
-    }
-    return VecMask<int, dst_n>(result).template cast<dst_t, dst_n>();
+template <>
+struct VecMaskCast<float, 1, int, 1> {
+  static inline VecMask<float, 1> apply(const VecMask<int, 1>& vec_mask) {
+    return Vectorized<float>(_mm256_castsi256_ps(vec_mask[0]));
  }
 };

 template <>
-struct VecMaskCast<double, 2, float, 1> {
-  static inline VecMask<double, 2> apply(const VecMask<float, 1>& vec_mask) {
-    auto int64_mask = VecMaskCast<int64_t, 2, float, 1>::apply(vec_mask);
-    return VecMaskCast<double, 2, int64_t, 2>::apply(int64_mask);
+struct VecMaskCast<int, 1, float, 1> {
+  static inline VecMask<int, 1> apply(const VecMask<float, 1>& vec_mask) {
+    return Vectorized<int>(_mm256_castps_si256(vec_mask[0]));
  }
 };
-template <>
-struct VecMaskCast<float, 1, double, 2> {
-  static inline VecMask<float, 1> apply(const VecMask<double, 2>& vec_mask) {
-    auto int64_mask = VecMaskCast<int64_t, 2, double, 2>::apply(vec_mask);
-    return VecMaskCast<float, 1, int64_t, 2>::apply(int64_mask);
+
+template <typename dst_t>
+struct VecMaskCast<dst_t, 1, int64_t, 2> {
+  static inline VecMask<dst_t, 1> apply(const VecMask<int64_t, 2>& vec_mask) {
+    auto int_vec = convert<int, 1, int64_t, 2>(VectorizedN<int64_t, 2>(vec_mask));
+    return VecMask<int, 1>(int_vec).cast<dst_t, 1>();
  }
 };

@ -240,42 +71,6 @@ inline bool VecMask<int, 1>::all_masked() const {
  return mask == 0xff;
 }

-template <int N>
-struct VecMaskCheck<int64_t, N> {
-  static inline bool all_zero(const VectorizedN<int64_t, N>& vec_mask) {
-    bool all_zero = true;
-    for (int i = 0; i < N; ++i) {
-      all_zero = all_zero && (_mm256_testz_si256(vec_mask[i], vec_mask[i]) > 0);
-      if (!all_zero) {
-        return all_zero;
-      }
-    }
-    return all_zero;
-  }
-
-  static inline bool is_masked(const VectorizedN<int64_t, N>& vec_mask, int i) {
-    for (int j = 0; j < N; ++j) {
-      if (i < (j + 1) * 4) {
-        return _mm256_movemask_pd(_mm256_castsi256_pd(vec_mask[j])) &
-            (1 << (i - j * 4));
-      }
-    }
-    return false;
-  }
-
-  static inline bool all_masked(const VectorizedN<int64_t, N>& vec_mask) {
-    bool all_masked = true;
-    for (int i = 0; i < N; ++i) {
-      all_masked = all_masked &&
-          (_mm256_movemask_pd(_mm256_castsi256_pd(vec_mask[i])) == 0x0f);
-      if (!all_masked) {
-        return all_masked;
-      }
-    }
-    return all_masked;
-  }
-};
-
 #define VEC_MASK_METHOD_WITH_CAST_TO_INT(                   \
    T, N, return_type, method, args_def, args)              \
  template <>                                               \
--- a/aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h
+++ b/aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h
@ -1348,20 +1348,41 @@ static inline void _transpose_mxn_half_32_32(__m512i r[], __m512i d[]) {
 // Code referred to FBGEMM:
 // https://github.com/pytorch/FBGEMM/blob/39a423e4ad1a04b77fea81c7d09c3e6f8984fae9/src/UtilsAvx512.cc#LL19C6-L19C6
 template<>
-inline void transpose_mxn<BFloat16>(const BFloat16* src, int64_t ld_src, BFloat16* dst, int64_t ld_dst, int M, int N) {
+inline void transpose_mxn<BFloat16, 32, 32>(
+    const BFloat16* src,
+    int64_t ld_src,
+    BFloat16* dst,
+    int64_t ld_dst) {
+  // Load from memory
+  __m512i r[32];
+#ifndef __msvc_cl__
+#pragma unroll(32)
+#endif
+  for (int i = 0; i < 32; ++i) {
+    r[i] = _mm512_loadu_si512(reinterpret_cast<const __m512i*>(src + i* ld_src));
+  }
+
+  __m512i d[32];
+  _transpose_mxn_half_32_32(r, d);
+
+  // Store to dst
+#ifndef __msvc_cl__
+#pragma unroll(32)
+#endif
+  for (int i = 0; i < 32; ++i) {
+    _mm512_storeu_si512(dst + i* ld_dst, d[i]);
+  }
+}
+
+template <typename T, int M, int N,
+          typename std::enable_if_t<std::is_same<T, BFloat16>::value && ((M < 32 && M != 16) || (N < 32 && N != 16)), int> = 0>
+inline void transpose_mxn(const BFloat16* src, int64_t ld_src, BFloat16* dst, int64_t ld_dst) {
  // load from src
-  TORCH_CHECK(M <= 32 && N <= 32, "transpose_mxn<BFloat16> expects M, N <= 32.");
+  __mmask32 src_mask = (1 << N) - 1;
  __m512i r[32];
  int i;
-  if (N == 32) {
-    for (i = 0; i < M; ++i) {
-      r[i] = _mm512_loadu_si512(&src[i * ld_src]);
-    }
-  } else {
-    __mmask32 src_mask = (1 << N) - 1;
-    for (i = 0; i < M; ++i) {
-      r[i] = _mm512_maskz_loadu_epi16(src_mask, &src[i * ld_src]);
-    }
+  for (i = 0; i < M; ++i) {
+    r[i] = _mm512_maskz_loadu_epi16(src_mask, &src[i * ld_src]);
  }
  for (; i < 32; ++i) {
    r[i] = _mm512_setzero_si512();
@ -1371,39 +1392,48 @@ inline void transpose_mxn<BFloat16>(const BFloat16* src, int64_t ld_src, BFloat1
  _transpose_mxn_half_32_32(r, d);

  // store to dst
-  if (M == 32) {
-    for (i = 0; i < N; ++i) {
-      _mm512_storeu_si512(&dst[i * ld_dst], d[i]);
-    }
-  } else {
-    __mmask32 dst_mask = (1 << M) - 1;
-    for (i = 0; i < N; ++i) {
-      _mm512_mask_storeu_epi16(&dst[i * ld_dst], dst_mask, d[i]);
-    }
+  __mmask32 dst_mask = (1 << M) - 1;
+  for (i = 0; i < N; ++i) {
+    _mm512_mask_storeu_epi16(&dst[i * ld_dst], dst_mask, d[i]);
  }
 }

-template <typename T, int M, int N,
-          typename std::enable_if_t<std::is_same<T, BFloat16>::value && ((M <= 32 && M != 16) || (N <= 32 && N != 16)), int> = 0>
-inline void transpose_mxn(const BFloat16* src, int64_t ld_src, BFloat16* dst, int64_t ld_dst) {
-  transpose_mxn<BFloat16>(src, ld_src, dst, ld_dst, M, N);
-}
-
 template<>
-inline void transpose_mxn<Half>(const Half* src, int64_t ld_src, Half* dst, int64_t ld_dst, int M, int N) {
-  TORCH_CHECK(M <= 32 && N <= 32, "transpose_mxn<Half> expects M, N <= 32.");
+inline void transpose_mxn<Half, 32, 32>(
+    const Half* src,
+    int64_t ld_src,
+    Half* dst,
+    int64_t ld_dst) {
+  // Load from memory
+  __m512i r[32];
+#ifndef __msvc_cl__
+#pragma unroll(32)
+#endif
+  for (int i = 0; i < 32; ++i) {
+    r[i] = _mm512_loadu_si512(reinterpret_cast<const __m512i*>(src + i* ld_src));
+  }
+
+  __m512i d[32];
+  _transpose_mxn_half_32_32(r, d);
+
+  // Store to dst
+#ifndef __msvc_cl__
+#pragma unroll(32)
+#endif
+  for (int i = 0; i < 32; ++i) {
+    _mm512_storeu_si512(dst + i* ld_dst, d[i]);
+  }
+}
+
+template <typename T, int M, int N,
+          typename std::enable_if_t<std::is_same<T, Half>::value && ((M < 32 && M != 16) || (N < 32 && N != 16)), int> = 0>
+inline void transpose_mxn(const Half* src, int64_t ld_src, Half* dst, int64_t ld_dst) {
  // load from src
+  __mmask32 src_mask = (1 << N) - 1;
  __m512i r[32];
  int i;
-  if (N == 32) {
-    for (i = 0; i < M; ++i) {
-      r[i] = _mm512_loadu_si512(&src[i * ld_src]);
-    }
-  } else {
-    __mmask32 src_mask = (1 << N) - 1;
-    for (i = 0; i < M; ++i) {
-      r[i] = _mm512_maskz_loadu_epi16(src_mask, &src[i * ld_src]);
-    }
+  for (i = 0; i < M; ++i) {
+    r[i] = _mm512_maskz_loadu_epi16(src_mask, &src[i * ld_src]);
  }
  for (; i < 32; ++i) {
    r[i] = _mm512_setzero_si512();
@ -1413,24 +1443,12 @@ inline void transpose_mxn<Half>(const Half* src, int64_t ld_src, Half* dst, int6
  _transpose_mxn_half_32_32(r, d);

  // store to dst
-  if (M == 32) {
-    for (i = 0; i < N; ++i) {
-      _mm512_storeu_si512(&dst[i * ld_dst], d[i]);
-    }
-  } else {
-    __mmask32 dst_mask = (1 << M) - 1;
-    for (i = 0; i < N; ++i) {
-      _mm512_mask_storeu_epi16(&dst[i * ld_dst], dst_mask, d[i]);
-    }
+  __mmask32 dst_mask = (1 << M) - 1;
+  for (i = 0; i < N; ++i) {
+    _mm512_mask_storeu_epi16(&dst[i * ld_dst], dst_mask, d[i]);
  }
 }

-template <typename T, int M, int N,
-          typename std::enable_if_t<std::is_same<T, Half>::value && ((M <= 32 && M != 16) || (N <= 32 && N != 16)), int> = 0>
-inline void transpose_mxn(const Half* src, int64_t ld_src, Half* dst, int64_t ld_dst) {
-  transpose_mxn<Half>(src, ld_src, dst, ld_dst, M, N);
-}
-
 template <>
 class Vectorized<Half>: public Vectorized16<Half> {
 public:
--- a/Show More
+++ b/Show More