[export] Preserve nn_module_stack for aliased nn modules

2025-10-22 22:25:10 +08:00 · 2025-09-30 18:02:43 -07:00
856 changed files with 5795 additions and 16093 deletions
--- a/.ci/aarch64_linux/aarch64_wheel_ci_build.py
+++ b/.ci/aarch64_linux/aarch64_wheel_ci_build.py
@ -13,6 +13,49 @@ def list_dir(path: str) -> list[str]:
    return check_output(["ls", "-1", path]).decode().split("\n")


+def build_ArmComputeLibrary() -> None:
+    """
+    Using ArmComputeLibrary for aarch64 PyTorch
+    """
+    print("Building Arm Compute Library")
+    acl_build_flags = [
+        "debug=0",
+        "neon=1",
+        "opencl=0",
+        "os=linux",
+        "openmp=1",
+        "cppthreads=0",
+        "arch=armv8a",
+        "multi_isa=1",
+        "fixed_format_kernels=1",
+        "build=native",
+    ]
+    acl_install_dir = "/acl"
+    acl_checkout_dir = os.getenv("ACL_SOURCE_DIR", "ComputeLibrary")
+    if os.path.isdir(acl_install_dir):
+        shutil.rmtree(acl_install_dir)
+    if not os.path.isdir(acl_checkout_dir) or not len(os.listdir(acl_checkout_dir)):
+        check_call(
+            [
+                "git",
+                "clone",
+                "https://github.com/ARM-software/ComputeLibrary.git",
+                "-b",
+                "v25.02",
+                "--depth",
+                "1",
+                "--shallow-submodules",
+            ]
+        )
+
+    check_call(
+        ["scons", "Werror=1", f"-j{os.cpu_count()}"] + acl_build_flags,
+        cwd=acl_checkout_dir,
+    )
+    for d in ["arm_compute", "include", "utils", "support", "src", "build"]:
+        shutil.copytree(f"{acl_checkout_dir}/{d}", f"{acl_install_dir}/{d}")
+
+
 def replace_tag(filename) -> None:
    with open(filename) as f:
        lines = f.readlines()
@ -313,13 +356,19 @@ if __name__ == "__main__":
        build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={branch[1 : branch.find('-')]} PYTORCH_BUILD_NUMBER=1 "

    if enable_mkldnn:
+        build_ArmComputeLibrary()
        print("build pytorch with mkldnn+acl backend")
-        build_vars += "USE_MKLDNN=ON USE_MKLDNN_ACL=ON "
-        build_vars += "ACL_ROOT_DIR=/acl "
+        build_vars += (
+            "USE_MKLDNN=ON USE_MKLDNN_ACL=ON "
+            "ACL_ROOT_DIR=/acl "
+            "LD_LIBRARY_PATH=/pytorch/build/lib:/acl/build:$LD_LIBRARY_PATH "
+            "ACL_INCLUDE_DIR=/acl/build "
+            "ACL_LIBRARY=/acl/build "
+        )
        if enable_cuda:
            build_vars += "BLAS=NVPL "
        else:
-            build_vars += "BLAS=OpenBLAS OpenBLAS_HOME=/opt/OpenBLAS "
+            build_vars += "BLAS=OpenBLAS OpenBLAS_HOME=/OpenBLAS "
    else:
        print("build pytorch without mkldnn backend")

--- a/.ci/aarch64_linux/build_aarch64_wheel.py
+++ b/.ci/aarch64_linux/build_aarch64_wheel.py
@ -299,6 +299,40 @@ def install_condaforge_python(host: RemoteHost, python_version="3.8") -> None:
        )


+def build_OpenBLAS(host: RemoteHost, git_clone_flags: str = "") -> None:
+    print("Building OpenBLAS")
+    host.run_cmd(
+        f"git clone https://github.com/xianyi/OpenBLAS -b v0.3.28 {git_clone_flags}"
+    )
+    make_flags = "NUM_THREADS=64 USE_OPENMP=1 NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=ARMV8"
+    host.run_cmd(
+        f"pushd OpenBLAS && make {make_flags} -j8 && sudo make {make_flags} install && popd && rm -rf OpenBLAS"
+    )
+
+
+def build_ArmComputeLibrary(host: RemoteHost, git_clone_flags: str = "") -> None:
+    print("Building Arm Compute Library")
+    acl_build_flags = " ".join(
+        [
+            "debug=0",
+            "neon=1",
+            "opencl=0",
+            "os=linux",
+            "openmp=1",
+            "cppthreads=0",
+            "arch=armv8a",
+            "multi_isa=1",
+            "fixed_format_kernels=1",
+            "build=native",
+        ]
+    )
+    host.run_cmd(
+        f"git clone https://github.com/ARM-software/ComputeLibrary.git -b v25.02 {git_clone_flags}"
+    )
+
+    host.run_cmd(f"cd ComputeLibrary && scons Werror=1 -j8 {acl_build_flags}")
+
+
 def embed_libgomp(host: RemoteHost, use_conda, wheel_name) -> None:
    host.run_cmd("pip3 install auditwheel")
    host.run_cmd(
@ -666,6 +700,7 @@ def start_build(
    configure_system(
        host, compiler=compiler, use_conda=use_conda, python_version=python_version
    )
+    build_OpenBLAS(host, git_clone_flags)

    if host.using_docker():
        print("Move libgfortant.a into a standard location")
@ -688,8 +723,6 @@ def start_build(
        f"git clone --recurse-submodules -b {branch} https://github.com/pytorch/pytorch {git_clone_flags}"
    )

-    host.run_cmd("pytorch/.ci/docker/common/install_openblas.sh")
-
    print("Building PyTorch wheel")
    build_opts = ""
    if pytorch_build_number is not None:
@ -710,18 +743,16 @@ def start_build(
    if host.using_docker():
        build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
    if enable_mkldnn:
-        host.run_cmd("pytorch/.ci/docker/common/install_acl.sh")
+        build_ArmComputeLibrary(host, git_clone_flags)
        print("build pytorch with mkldnn+acl backend")
        build_vars += " USE_MKLDNN=ON USE_MKLDNN_ACL=ON"
-        build_vars += " BLAS=OpenBLAS"
-        build_vars += " OpenBLAS_HOME=/opt/OpenBLAS"
-        build_vars += " ACL_ROOT_DIR=/acl"
        host.run_cmd(
-            f"cd $HOME/pytorch && {build_vars} python3 -m build --wheel --no-isolation{build_opts}"
+            f"cd $HOME/pytorch && export ACL_ROOT_DIR=$HOME/ComputeLibrary && "
+            f"{build_vars} python3 -m build --wheel --no-isolation{build_opts}"
        )
        print("Repair the wheel")
        pytorch_wheel_name = host.list_dir("pytorch/dist")[0]
-        ld_library_path = "/acl/build:$HOME/pytorch/build/lib"
+        ld_library_path = "$HOME/acl/build:$HOME/pytorch/build/lib"
        host.run_cmd(
            f"export LD_LIBRARY_PATH={ld_library_path} && auditwheel repair $HOME/pytorch/dist/{pytorch_wheel_name}"
        )
@ -877,7 +908,7 @@ def terminate_instances(instance_type: str) -> None:
 def parse_arguments():
    from argparse import ArgumentParser

-    parser = ArgumentParser("Build and test AARCH64 wheels using EC2")
+    parser = ArgumentParser("Builid and test AARCH64 wheels using EC2")
    parser.add_argument("--key-name", type=str)
    parser.add_argument("--debug", action="store_true")
    parser.add_argument("--build-only", action="store_true")
--- a/.ci/docker/ci_commit_pins/executorch.txt
+++ b/.ci/docker/ci_commit_pins/executorch.txt
@ -1 +1 @@
-deb42f2a8e48f5032b4a98ee781a15fa87a157cf
+e0dda9059d082537cee36be6c5e4fe3b18c880c0
--- a/.ci/docker/ci_commit_pins/nccl-cu12.txt
+++ b/.ci/docker/ci_commit_pins/nccl-cu12.txt
@ -1 +1 @@
-v2.27.5-1
+v2.28.3-1
--- a/.ci/docker/ci_commit_pins/nccl-cu13.txt
+++ b/.ci/docker/ci_commit_pins/nccl-cu13.txt
@ -1 +1 @@
-v2.27.7-1
+v2.28.3-1
--- a/.ci/docker/ci_commit_pins/triton.txt
+++ b/.ci/docker/ci_commit_pins/triton.txt
@ -1 +1 @@
-27664085f804afc83df26f740bb46c365854f2c4
+bbb06c0334a6772b92d24bde54956e675c8c6604
--- a/.ci/docker/common/install_acl.sh
+++ b/.ci/docker/common/install_acl.sh
@ -1,27 +1,16 @@
-#!/bin/bash
-# Script used only in CD pipeline
+set -euo pipefail

-set -eux
-
-ACL_VERSION=${ACL_VERSION:-"v25.02"}
-ACL_INSTALL_DIR="/acl"
+readonly version=v25.02
+readonly src_host=https://github.com/ARM-software
+readonly src_repo=ComputeLibrary

 # Clone ACL
-git clone https://github.com/ARM-software/ComputeLibrary.git -b "${ACL_VERSION}" --depth 1 --shallow-submodules
+[[ ! -d ${src_repo} ]] && git clone ${src_host}/${src_repo}.git
+cd ${src_repo}
+
+git checkout $version

-ACL_CHECKOUT_DIR="ComputeLibrary"
 # Build with scons
-pushd $ACL_CHECKOUT_DIR
 scons -j8  Werror=0 debug=0 neon=1 opencl=0 embed_kernels=0 \
  os=linux arch=armv8a build=native multi_isa=1 \
  fixed_format_kernels=1 openmp=1 cppthreads=0
-popd
-
-# Install ACL
-sudo mkdir -p ${ACL_INSTALL_DIR}
-for d in arm_compute include utils support src build
-do
-  sudo cp -r ${ACL_CHECKOUT_DIR}/${d} ${ACL_INSTALL_DIR}/${d}
-done
-
-rm -rf $ACL_CHECKOUT_DIR
--- a/.ci/docker/common/install_onnx.sh
+++ b/.ci/docker/common/install_onnx.sh
@ -19,8 +19,8 @@ pip_install \
  transformers==4.36.2

 pip_install coloredlogs packaging
-pip_install onnxruntime==1.23.0
-pip_install onnxscript==0.5.3
+pip_install onnxruntime==1.22.1
+pip_install onnxscript==0.4.0

 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/
--- a/.ci/docker/common/install_openblas.sh
+++ b/.ci/docker/common/install_openblas.sh
@ -3,10 +3,8 @@

 set -ex

-OPENBLAS_VERSION=${OPENBLAS_VERSION:-"v0.3.30"}
-
-# Clone OpenBLAS
-git clone https://github.com/OpenMathLib/OpenBLAS.git -b "${OPENBLAS_VERSION}" --depth 1 --shallow-submodules
+cd /
+git clone https://github.com/OpenMathLib/OpenBLAS.git -b "${OPENBLAS_VERSION:-v0.3.30}" --depth 1 --shallow-submodules

 OPENBLAS_CHECKOUT_DIR="OpenBLAS"
 OPENBLAS_BUILD_FLAGS="
@ -19,7 +17,5 @@ CFLAGS=-O3
 BUILD_BFLOAT16=1
 "

-make -j8 ${OPENBLAS_BUILD_FLAGS} -C $OPENBLAS_CHECKOUT_DIR
-sudo make install -C $OPENBLAS_CHECKOUT_DIR
-
-rm -rf $OPENBLAS_CHECKOUT_DIR
+make -j8 ${OPENBLAS_BUILD_FLAGS} -C ${OPENBLAS_CHECKOUT_DIR}
+make -j8 ${OPENBLAS_BUILD_FLAGS} install -C ${OPENBLAS_CHECKOUT_DIR}
--- a/.ci/docker/common/patch_libstdc.sh
+++ b/.ci/docker/common/patch_libstdc.sh
@ -1,9 +0,0 @@
-#!/bin/bash
-set -xe
-# Script used in Linux x86 and aarch64 CD pipeline
-
-# Workaround for exposing statically linked libstdc++ CXX11 ABI symbols.
-# see: https://github.com/pytorch/pytorch/issues/133437
-LIBNONSHARED=$(gcc -print-file-name=libstdc++_nonshared.a)
-nm -g $LIBNONSHARED | grep " T " | grep recursive_directory_iterator | cut -c 20-  > weaken-symbols.txt
-objcopy --weaken-symbols weaken-symbols.txt $LIBNONSHARED $LIBNONSHARED
--- a/.ci/docker/manywheel/Dockerfile_2_28
+++ b/.ci/docker/manywheel/Dockerfile_2_28
@ -130,8 +130,7 @@ ENV LD_LIBRARY_PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib64:/op
 RUN for cpython_version in "cp312-cp312" "cp313-cp313" "cp313-cp313t"; do \
    /opt/python/${cpython_version}/bin/python -m pip install setuptools wheel; \
    done;
-ADD ./common/patch_libstdc.sh patch_libstdc.sh
-RUN bash ./patch_libstdc.sh && rm patch_libstdc.sh
+

 # cmake-3.18.4 from pip; force in case cmake3 already exists
 RUN yum install -y python3-pip && \
--- a/.ci/docker/manywheel/Dockerfile_2_28_aarch64
+++ b/.ci/docker/manywheel/Dockerfile_2_28_aarch64
@ -62,13 +62,6 @@ ARG OPENBLAS_VERSION
 ADD ./common/install_openblas.sh install_openblas.sh
 RUN bash ./install_openblas.sh && rm install_openblas.sh

-# Install Arm Compute Library
-FROM base as arm_compute
-# use python3.9 to install scons
-RUN python3.9 -m pip install scons==4.7.0
-RUN ln -sf /opt/python/cp39-cp39/bin/scons /usr/local/bin
-COPY ./common/install_acl.sh install_acl.sh
-RUN bash ./install_acl.sh && rm install_acl.sh
 FROM base as final

 # remove unnecessary python versions
@ -77,7 +70,4 @@ RUN rm -rf /opt/python/cp26-cp26mu /opt/_internal/cpython-2.6.9-ucs4
 RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6
 RUN rm -rf /opt/python/cp34-cp34m /opt/_internal/cpython-3.4.6
 COPY --from=openblas     /opt/OpenBLAS/  /opt/OpenBLAS/
-COPY --from=arm_compute /acl /acl
-ENV LD_LIBRARY_PATH=/opt/OpenBLAS/lib:/acl/build/:$LD_LIBRARY_PATH
-ADD ./common/patch_libstdc.sh patch_libstdc.sh
-RUN bash ./patch_libstdc.sh && rm patch_libstdc.sh
+ENV LD_LIBRARY_PATH=/opt/OpenBLAS/lib:$LD_LIBRARY_PATH
--- a/.ci/docker/manywheel/Dockerfile_cuda_aarch64
+++ b/.ci/docker/manywheel/Dockerfile_cuda_aarch64
@ -86,15 +86,6 @@ FROM base as nvpl
 ADD ./common/install_nvpl.sh install_nvpl.sh
 RUN bash ./install_nvpl.sh && rm install_nvpl.sh

-# Install Arm Compute Library
-FROM base as arm_compute
-# use python3.9 to install scons
-RUN python3.9 -m pip install scons==4.7.0
-RUN ln -sf /opt/python/cp39-cp39/bin/scons /usr/local/bin
-COPY ./common/install_acl.sh install_acl.sh
-RUN bash ./install_acl.sh && rm install_acl.sh
-FROM base as final
-
 FROM final as cuda_final
 ARG BASE_CUDA_VERSION
 RUN rm -rf /usr/local/cuda-${BASE_CUDA_VERSION}
@ -102,9 +93,5 @@ COPY --from=cuda     /usr/local/cuda-${BASE_CUDA_VERSION}  /usr/local/cuda-${BAS
 COPY --from=magma    /usr/local/cuda-${BASE_CUDA_VERSION}  /usr/local/cuda-${BASE_CUDA_VERSION}
 COPY --from=nvpl /opt/nvpl/lib/  /usr/local/lib/
 COPY --from=nvpl /opt/nvpl/include/  /usr/local/include/
-COPY --from=arm_compute /acl /acl
 RUN ln -sf /usr/local/cuda-${BASE_CUDA_VERSION} /usr/local/cuda
 ENV PATH=/usr/local/cuda/bin:$PATH
-ENV LD_LIBRARY_PATH=/acl/build/:$LD_LIBRARY_PATH
-ADD ./common/patch_libstdc.sh patch_libstdc.sh
-RUN bash ./patch_libstdc.sh && rm patch_libstdc.sh
--- a/.ci/docker/manywheel/build.sh
+++ b/.ci/docker/manywheel/build.sh
@ -28,7 +28,6 @@ fi
 MANY_LINUX_VERSION=${MANY_LINUX_VERSION:-}
 DOCKERFILE_SUFFIX=${DOCKERFILE_SUFFIX:-}
 OPENBLAS_VERSION=${OPENBLAS_VERSION:-}
-ACL_VERSION=${ACL_VERSION:-}

 case ${image} in
    manylinux2_28-builder:cpu)
@ -42,6 +41,7 @@ case ${image} in
        GPU_IMAGE=arm64v8/almalinux:8
        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13 --build-arg NINJA_VERSION=1.12.1"
        MANY_LINUX_VERSION="2_28_aarch64"
+        OPENBLAS_VERSION="v0.3.30"
        ;;
    manylinuxs390x-builder:cpu-s390x)
        TARGET=final
@ -119,8 +119,7 @@ tmp_tag=$(basename "$(mktemp -u)" | tr '[:upper:]' '[:lower:]')
 DOCKER_BUILDKIT=1 docker build  \
    ${DOCKER_GPU_BUILD_ARG} \
    --build-arg "GPU_IMAGE=${GPU_IMAGE}" \
-    --build-arg "OPENBLAS_VERSION=${OPENBLAS_VERSION:-}" \
-    --build-arg "ACL_VERSION=${ACL_VERSION:-}" \
+    --build-arg "OPENBLAS_VERSION=${OPENBLAS_VERSION}" \
    --target "${TARGET}" \
    -t "${tmp_tag}" \
    $@ \
--- a/.ci/docker/requirements-ci.txt
+++ b/.ci/docker/requirements-ci.txt
@ -52,10 +52,10 @@ flatbuffers==24.12.23
 #Pinned versions: 24.12.23
 #test that import:

-hypothesis==6.56.4
+hypothesis==5.35.1
 # Pin hypothesis to avoid flakiness: https://github.com/pytorch/pytorch/issues/31136
 #Description: advanced library for generating parametrized tests
-#Pinned versions: 6.56.4
+#Pinned versions: 5.35.1
 #test that import: test_xnnpack_integration.py, test_pruning_op.py, test_nn.py

 junitparser==2.1.1
@ -98,7 +98,7 @@ librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
 #Pinned versions:
 #test that import:

-mypy==1.16.0 ; platform_system == "Linux"
+mypy==1.16.0 ; platform_system != "Windows"
 # Pin MyPy version because new errors are likely to appear with each release
 # Skip on Windows as lots of type annotations are POSIX specific
 #Description: linter
@ -169,7 +169,7 @@ optree==0.13.0

 pillow==11.0.0
 #Description:  Python Imaging Library fork
-#Pinned versions: 11.0.0
+#Pinned versions: 10.3.0
 #test that import:

 protobuf==5.29.5
@ -217,7 +217,7 @@ pytest-subtests==0.13.1
 #Pinned versions:
 #test that import:

-xdoctest==1.3.0
+xdoctest==1.1.0
 #Description: runs doctests in pytest
 #Pinned versions: 1.1.0
 #test that import:
@ -268,7 +268,7 @@ scipy==1.14.1 ; python_version >= "3.12"
 #test that import:

 # needed by torchgen utils
-typing-extensions==4.12.2
+typing-extensions>=4.10.0
 #Description: type hints for python
 #Pinned versions:
 #test that import:
@ -341,7 +341,7 @@ onnx==1.18.0
 #Pinned versions:
 #test that import:

-onnxscript==0.5.3
+onnxscript==0.4.0
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@ -361,10 +361,9 @@ pwlf==2.2.1
 #test that import: test_sac_estimator.py

 # To build PyTorch itself
-pyyaml==6.0.2
+pyyaml
 pyzstd
-setuptools==78.1.1
-packaging==23.1
+setuptools>=70.1.0
 six

 scons==4.5.2 ; platform_machine == "aarch64"
@ -385,10 +384,7 @@ cmake==3.31.6
 tlparse==0.4.0
 #Description: required for log parsing

-filelock==3.18.0
-#Description: required for inductor testing
-
-cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x" and platform_system != "Darwin"
+cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"
 #Description: required for testing CUDAGraph::raw_cuda_graph(). See https://nvidia.github.io/cuda-python/cuda-bindings/latest/support.html for how this version was chosen. Note "Any fix in the latest bindings would be backported to the prior major version" means that only the newest version of cuda-bindings will get fixes. Depending on the latest version of 12.x is okay because all 12.y versions will be supported via "CUDA minor version compatibility". Pytorch builds against 13.z versions of cuda toolkit work with 12.x versions of cuda-bindings as well because newer drivers work with old toolkits.
 #test that import: test_cuda.py

--- a/.ci/pytorch/build.sh
+++ b/.ci/pytorch/build.sh
@ -89,7 +89,7 @@ fi
 if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
  export USE_MKLDNN=1
  export USE_MKLDNN_ACL=1
-  export ACL_ROOT_DIR=/acl
+  export ACL_ROOT_DIR=/ComputeLibrary
 fi

 if [[ "$BUILD_ENVIRONMENT" == *riscv64* ]]; then
--- a/.ci/pytorch/check_binary.sh
+++ b/.ci/pytorch/check_binary.sh
@ -67,7 +67,7 @@ fi
 #       wheels with cxx11-abi

 echo "Checking that the gcc ABI is what we expect"
-if [[ "$(uname)" != 'Darwin' &&  "$(uname -m)" != "s390x" ]]; then
+if [[ "$(uname)" != 'Darwin' ]]; then
  # We also check that there are cxx11 symbols in libtorch
  #
  echo "Checking that symbols in libtorch.so have the right gcc abi"
--- a/.ci/pytorch/smoke_test/check_binary_symbols.py
+++ b/.ci/pytorch/smoke_test/check_binary_symbols.py
@ -32,9 +32,6 @@ LIBTORCH_NAMESPACE_LIST = (
    "torch::",
 )

-# Patterns for detecting statically linked libstdc++ symbols
-STATICALLY_LINKED_CXX11_ABI = [re.compile(r".*recursive_directory_iterator.*")]
-

 def _apply_libtorch_symbols(symbols):
    return [
@ -56,17 +53,12 @@ def get_symbols(lib: str) -> list[tuple[str, str, str]]:
    return [x.split(" ", 2) for x in lines.decode("latin1").split("\n")[:-1]]


-def grep_symbols(
-    lib: str, patterns: list[Any], symbol_type: str | None = None
-) -> list[str]:
+def grep_symbols(lib: str, patterns: list[Any]) -> list[str]:
    def _grep_symbols(
        symbols: list[tuple[str, str, str]], patterns: list[Any]
    ) -> list[str]:
        rc = []
        for _s_addr, _s_type, s_name in symbols:
-            # Filter by symbol type if specified
-            if symbol_type and _s_type != symbol_type:
-                continue
            for pattern in patterns:
                if pattern.match(s_name):
                    rc.append(s_name)
@ -88,18 +80,6 @@ def grep_symbols(
        return functools.reduce(list.__add__, (x.result() for x in tasks), [])


-def check_lib_statically_linked_libstdc_cxx_abi_symbols(lib: str) -> None:
-    cxx11_statically_linked_symbols = grep_symbols(
-        lib, STATICALLY_LINKED_CXX11_ABI, symbol_type="T"
-    )
-    num_statically_linked_symbols = len(cxx11_statically_linked_symbols)
-    print(f"num_statically_linked_symbols (T): {num_statically_linked_symbols}")
-    if num_statically_linked_symbols > 0:
-        raise RuntimeError(
-            f"Found statically linked libstdc++ symbols (recursive_directory_iterator): {cxx11_statically_linked_symbols[:100]}"
-        )
-
-
 def check_lib_symbols_for_abi_correctness(lib: str) -> None:
    print(f"lib: {lib}")
    cxx11_symbols = grep_symbols(lib, LIBTORCH_CXX11_PATTERNS)
@ -127,7 +107,6 @@ def main() -> None:

    libtorch_cpu_path = str(install_root / "lib" / "libtorch_cpu.so")
    check_lib_symbols_for_abi_correctness(libtorch_cpu_path)
-    check_lib_statically_linked_libstdc_cxx_abi_symbols(libtorch_cpu_path)


 if __name__ == "__main__":
--- a/.ci/pytorch/test.sh
+++ b/.ci/pytorch/test.sh
@ -34,14 +34,12 @@ fi


 # Patch numba to avoid CUDA-13 crash, see https://github.com/pytorch/pytorch/issues/162878
-if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
-  NUMBA_CUDA_DIR=$(python -c "import os;import numba.cuda; print(os.path.dirname(numba.cuda.__file__))" 2>/dev/null || true)
-  if [ -n "$NUMBA_CUDA_DIR" ]; then
-    NUMBA_PATCH="$(dirname "$(realpath "${BASH_SOURCE[0]}")")/numba-cuda-13.patch"
-    pushd "$NUMBA_CUDA_DIR"
-    patch -p4 <"$NUMBA_PATCH"
-    popd
-  fi
+NUMBA_CUDA_DIR=$(python -c "import os;import numba.cuda; print(os.path.dirname(numba.cuda.__file__))" 2>/dev/null || true)
+if [ -n "$NUMBA_CUDA_DIR" ]; then
+  NUMBA_PATCH="$(dirname "$(realpath "${BASH_SOURCE[0]}")")/numba-cuda-13.patch"
+  pushd "$NUMBA_CUDA_DIR"
+  patch -p4 <"$NUMBA_PATCH"
+  popd
 fi

 echo "Environment variables:"
--- a/.ci/pytorch/test_fa3_abi_stable.sh
+++ b/.ci/pytorch/test_fa3_abi_stable.sh
@ -1,32 +0,0 @@
-#!/bin/bash
-
-set -ex -o pipefail
-
-# Suppress ANSI color escape sequences
-export TERM=vt100
-
-# shellcheck source=./common.sh
-source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
-# shellcheck source=./common-build.sh
-source "$(dirname "${BASH_SOURCE[0]}")/common-build.sh"
-
-echo "Environment variables"
-env
-
-echo "Testing FA3 stable wheel still works with currently built torch"
-
-echo "Installing ABI Stable FA3 wheel"
-# The wheel was built on https://github.com/Dao-AILab/flash-attention/commit/b3846b059bf6b143d1cd56879933be30a9f78c81
-# on torch nightly torch==2.9.0.dev20250830+cu129
-$MAYBE_SUDO pip -q install https://s3.amazonaws.com/ossci-linux/wheels/flash_attn_3-3.0.0b1-cp39-abi3-linux_x86_64.whl
-
-pushd flash-attention/hopper
-export PYTHONPATH=$PWD
-pytest -v -s \
-  "test_flash_attn.py::test_flash_attn_output[1-1-192-False-False-False-0.0-False-False-mha-dtype0]" \
-  "test_flash_attn.py::test_flash_attn_varlen_output[511-1-64-True-False-False-0.0-False-False-gqa-dtype2]" \
-  "test_flash_attn.py::test_flash_attn_kvcache[1-128-128-False-False-True-None-0.0-False-False-True-False-True-False-gqa-dtype0]" \
-  "test_flash_attn.py::test_flash_attn_race_condition[97-97-192-True-dtype0]" \
-  "test_flash_attn.py::test_flash_attn_combine[2-3-64-dtype1]" \
-  "test_flash_attn.py::test_flash3_bw_compatibility"
-popd
--- a/.ci/pytorch/win-test-helpers/build_pytorch.bat
+++ b/.ci/pytorch/win-test-helpers/build_pytorch.bat
@ -38,12 +38,10 @@ if errorlevel 1 goto fail
 if not errorlevel 0 goto fail

 :: Update CMake
-:: TODO: Investigate why this helps MKL detection, even when CMake from choco is not used
 call choco upgrade -y cmake --no-progress --installargs 'ADD_CMAKE_TO_PATH=System' --apply-install-arguments-to-dependencies --version=3.27.9
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail

-:: TODO: Move to .ci/docker/requirements-ci.txt
 call pip install mkl==2024.2.0 mkl-static==2024.2.0 mkl-include==2024.2.0
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail
--- a/.ci/pytorch/win-test.sh
+++ b/.ci/pytorch/win-test.sh
@ -37,8 +37,27 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
  export PYTORCH_TESTING_DEVICE_ONLY_FOR="cuda"
 fi

-# TODO: Move this to .ci/docker/requirements-ci.txt
-python -m pip install "psutil==5.9.1" "pynvml==11.4.1" "pytest-shard==0.1.2"
+# TODO: Move both of them to Windows AMI
+python -m pip install tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
+
+# Copied from https://github.com/pytorch/test-infra/blob/be01a40157c36cd5a48391fdf44a7bc3ebd4c7e3/aws/ami/windows/scripts/Installers/Install-Pip-Dependencies.ps1#L16 with some adjustments
+# pytest-rerunfailures==10.3 as 10.2 fails with INTERNALERROR> pluggy._manager.PluginValidationError: unknown hook 'pytest_configure_node'
+# scipy from 1.6.3 to 1.10
+# expecttest from 0.1.3 to 0.3.0
+# xdoctest from 1.0.2 to 1.3.0
+python -m pip install "future==0.18.2" "hypothesis==5.35.1" "expecttest==0.3.0" "librosa>=0.6.2" "scipy==1.10.1" "psutil==5.9.1" "pynvml==11.4.1" "pillow==9.2.0" "unittest-xml-reporting<=3.2.0,>=2.0.0" "pytest==7.1.3" "pytest-xdist==2.5.0" "pytest-flakefinder==1.1.0" "pytest-rerunfailures==10.3" "pytest-shard==0.1.2" "sympy==1.11.1" "xdoctest==1.3.0" "pygments==2.12.0" "opt-einsum>=3.3" "networkx==2.8.8" "mpmath==1.2.1" "pytest-cpp==2.3.0" "boto3==1.35.42"
+
+# Install Z3 optional dependency for Windows builds.
+python -m pip install z3-solver==4.15.1.0
+
+# Install tlparse for test\dynamo\test_structured_trace.py UTs.
+python -m pip install tlparse==0.4.0
+
+# Install parameterized
+python -m pip install parameterized==0.8.1
+
+# Install pulp for testing ilps under torch\distributed\_tools
+python -m pip install pulp==2.9.0

 run_tests() {
    # Run nvidia-smi if available
--- a/.ci/pytorch/windows/cuda128.bat
+++ b/.ci/pytorch/windows/cuda128.bat
@ -37,10 +37,10 @@ IF "%CUDA_PATH_V128%"=="" (
 )

 IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=7.0;7.5;8.0;8.6;9.0;10.0;12.0
+    set TORCH_CUDA_ARCH_LIST=6.1;7.0;7.5;8.0;8.6;9.0;10.0;12.0
    set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
 ) ELSE (
-    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90 -gencode=arch=compute_100,code=compute_100 -gencode=arch=compute_120,code=compute_120
+    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90 -gencode=arch=compute_100,code=compute_100 -gencode=arch=compute_120,code=compute_120
 )

 set "CUDA_PATH=%CUDA_PATH_V128%"
--- a/.clang-tidy
+++ b/.clang-tidy
@ -59,14 +59,13 @@ performance-*,
 -performance-enum-size,
 readability-container-size-empty,
 readability-delete-null-pointer,
-readability-duplicate-include,
+readability-duplicate-include
 readability-misplaced-array-index,
-readability-redundant*,
+readability-redundant*
 readability-simplify-subscript-expr,
 readability-string-compare,
 -readability-redundant-access-specifiers,
 -readability-redundant-control-flow,
-readability-redundant-inline-specifier,
 '
 HeaderFilterRegex: '^(aten/|c10/|torch/).*$'
 WarningsAsErrors: '*'
--- a/.github/actions/teardown-win/action.yml
+++ b/.github/actions/teardown-win/action.yml
@ -23,6 +23,9 @@ runs:
      run: |
        .github\scripts\kill_active_ssh_sessions.ps1

+    - name: Clean up leftover processes on non-ephemeral Windows runner
+      uses: pytorch/test-infra/.github/actions/cleanup-runner@main
+
    # Cleaning up Windows workspace sometimes fails flakily with device or resource busy
    # error, meaning one or more processes haven't stopped completely yet. So trying to
    # retry this step several time similar to how checkout-pytorch GHA does
--- a/.github/ci_commit_pins/vllm.txt
+++ b/.github/ci_commit_pins/vllm.txt
@ -1 +1 @@
-0ad9951c416d33c5da4f7a504fb162cbe62386f5
+78a47f87ce259a48f0391fa9ae15add05ea7432b
--- a/.github/ci_commit_pins/xla.txt
+++ b/.github/ci_commit_pins/xla.txt
@ -1 +1 @@
-2a9138a26ee257fef05310ad3fecf7c55fe80d73
+0fc62aa26a30ed7ca419d285f285cb5ba02c4394
--- a/.github/ci_configs/vllm/Dockerfile.tmp_vllm
+++ b/.github/ci_configs/vllm/Dockerfile.tmp_vllm
@ -202,7 +202,7 @@ ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
 ARG nvcc_threads=4
 ENV NVCC_THREADS=$nvcc_threads
-ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
+ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}

 ARG USE_SCCACHE
@ -297,28 +297,16 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
    echo "[INFO] Showing torch_build_versions.txt content:" && \
    cat torch_build_versions.txt

-# Install build and runtime dependencies, this is needed for flashinfer install
-COPY requirements/build.txt requirements/build.txt
-COPY use_existing_torch.py use_existing_torch.py
-RUN python3 use_existing_torch.py
-RUN cat requirements/build.txt
-
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
    if ! python3 -m uv --version > /dev/null 2>&1; then \
        python3 -m pip install uv==0.8.4; \
    fi
-
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy

-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/build.txt
-
-
 # Default mount file as placeholder, this just avoid the mount error
 ARG TORCH_WHEELS_PATH="./requirements"
 # Install torch, torchaudio and torchvision
@ -344,11 +332,13 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Install xformers wheel from previous stage
 RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --system /wheels/xformers/*.whl --verbose
+
 # Build flashinfer from source.
 ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
 # install package for build flashinfer
 # see issue: https://github.com/flashinfer-ai/flashinfer/issues/738

+RUN pip install build==1.3.0
 RUN pip freeze | grep -E 'setuptools|packaging|build'

 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
--- a/.github/ci_configs/vllm/use_existing_torch.py
+++ b/.github/ci_configs/vllm/use_existing_torch.py
@ -1,14 +1,9 @@
 import glob
-import os


 requires_files = glob.glob("requirements/*.txt")
 requires_files += ["pyproject.toml"]
-
 for file in requires_files:
-    if not os.path.exists(file):
-        print(f"!!! skipping missing {file}")
-        continue
    print(f">>> cleaning {file}")
    with open(file) as f:
        lines = f.readlines()
--- a/.github/requirements/pip-requirements-macOS.txt
+++ b/.github/requirements/pip-requirements-macOS.txt
@ -0,0 +1,37 @@
+boto3==1.35.42
+build==1.2.2.post1
+cmake==3.27.*
+expecttest==0.3.0
+fbscribelogger==0.1.7
+filelock==3.18.0
+hypothesis==6.56.4
+librosa>=0.6.2
+mpmath==1.3.0
+networkx==2.8.7
+ninja==1.10.2.4
+numba==0.59.0
+numpy==1.26.4
+opt-einsum>=3.3
+optree==0.13.0
+packaging==23.1
+parameterized==0.8.1
+pillow==10.3.0
+protobuf==5.29.5
+psutil==5.9.8
+pygments==2.15.0
+pytest-cpp==2.3.0
+pytest-flakefinder==1.1.0
+pytest-rerunfailures==10.3
+pytest-subtests==0.13.1
+pytest-xdist==3.3.1
+pytest==7.3.2
+pyyaml==6.0.2
+scipy==1.12.0
+setuptools==78.1.1
+sympy==1.13.3
+tlparse==0.4.0
+tensorboard==2.13.0
+typing-extensions==4.12.2
+unittest-xml-reporting<=3.2.0,>=2.0.0
+xdoctest==1.1.0
+z3-solver==4.15.1.0
--- a/.github/scripts/filter_test_configs.py
+++ b/.github/scripts/filter_test_configs.py
@ -502,7 +502,6 @@ def perform_misc_tasks(
    job_name: str,
    pr_body: str,
    branch: Optional[str] = None,
-    tag: Optional[str] = None,
 ) -> None:
    """
    In addition to apply the filter logic, the script also does the following
@ -510,9 +509,7 @@ def perform_misc_tasks(
    """
    set_output(
        "keep-going",
-        branch == MAIN_BRANCH
-        or bool(tag and re.match(r"^trunk/[a-f0-9]{40}$", tag))
-        or check_for_setting(labels, pr_body, "keep-going"),
+        branch == MAIN_BRANCH or check_for_setting(labels, pr_body, "keep-going"),
    )
    set_output(
        "ci-verbose-test-logs",
@ -637,7 +634,6 @@ def main() -> None:
        job_name=args.job_name,
        pr_body=pr_body if pr_body else "",
        branch=args.branch,
-        tag=tag,
    )

    # Set the filtered test matrix as the output
--- a/.github/scripts/generate_binary_build_matrix.py
+++ b/.github/scripts/generate_binary_build_matrix.py
@ -53,7 +53,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
+        "nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | "
        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
@ -70,7 +70,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
+        "nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | "
        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
@ -87,7 +87,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
-        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
+        "nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | "
        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
        "nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
--- a/.github/workflows/_get-changed-files.yml
+++ b/.github/workflows/_get-changed-files.yml
@ -40,15 +40,6 @@ jobs:
              # Use gh CLI to get changed files in the PR with explicit repo
              CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')

-              # See https://github.com/pytorch/pytorch/pull/134215#issuecomment-2332128790
-              PYI_FILES_TO_ADD=""
-              for file in ${CHANGED_FILES}; do
-                if [[ "${file}" == *".pyi.in" ]]; then
-                  PYI_FILES_TO_ADD="${PYI_FILES_TO_ADD} ${file//.in/}"
-                fi
-              done
-              CHANGED_FILES="${CHANGED_FILES}${PYI_FILES_TO_ADD}"
-
              if [ -z "$CHANGED_FILES" ]; then
                echo "No changed files found, setting to '*'"
                CHANGED_FILES="*"
--- a/.github/workflows/_linux-test-stable-fa3.yml
+++ b/.github/workflows/_linux-test-stable-fa3.yml
@ -1,255 +0,0 @@
-# The point of this workflow is to test that a FA3 wheel that was built based off the
-# stable ABI as of torch nightly 20250830 can still run on the newer torch.
-#
-# This workflow is very similar to the _linux-test.yml workflow, with the following
-# differences:
-#   1. It is simpler (there is no test matrix)
-#   2. It pulls flash-attention as a secondary repository in order to access the tests.
-#      Note that it does not BUILD anything from flash-attention, as we have a prebuilt
-#      wheel. We pull flash-attention only to run a few tests.
-#   3. It runs only FA3 tests. No PyTorch tests are run.
-name: linux-test-stable-fa3
-
-on:
-  workflow_call:
-    inputs:
-      build-environment:
-        required: true
-        type: string
-        description: Top-level label for what's being built/tested.
-      docker-image:
-        required: true
-        type: string
-        description: Docker image to run in.
-      timeout-minutes:
-        required: false
-        type: number
-        default: 30
-        description: |
-          Set the maximum (in minutes) how long the workflow should take to finish
-      s3-bucket:
-        description: S3 bucket to download artifact
-        required: false
-        type: string
-        default: "gha-artifacts"
-    secrets:
-      HUGGING_FACE_HUB_TOKEN:
-        required: false
-        description: |
-          HF Auth token to avoid rate limits when downloading models or datasets from hub
-      VLLM_TEST_HUGGING_FACE_TOKEN:
-        required: false
-        description: |
-          HF Auth token to test vllm
-      SCRIBE_GRAPHQL_ACCESS_TOKEN:
-        required: false
-        description: |
-          FB app token to write to scribe endpoint
-
-env:
-  GIT_DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
-
-jobs:
-  test:
-    # Don't run on forked repos
-    if: github.repository_owner == 'pytorch'
-    runs-on: linux.aws.h100
-    timeout-minutes: ${{ inputs.timeout-minutes || 30 }}
-    permissions:
-      id-token: write
-      contents: read
-    steps:
-      - name: Checkout PyTorch
-        uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
-        with:
-          no-sudo: true
-
-      - name: Checkout flash-attention as a secondary repository
-        uses: actions/checkout@v4
-        with:
-          repository: Dao-AILab/flash-attention
-          path: flash-attention
-
-      - name: Setup Linux
-        uses: ./.github/actions/setup-linux
-
-      - name: Calculate docker image
-        id: calculate-docker-image
-        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
-        with:
-          docker-image-name: ${{ inputs.docker-image }}
-
-      - name: Use following to pull public copy of the image
-        id: print-ghcr-mirror
-        env:
-          ECR_DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
-        shell: bash
-        run: |
-          tag=${ECR_DOCKER_IMAGE##*:}
-          echo "docker pull ghcr.io/pytorch/ci-image:${tag/:/-}"
-
-      - name: Pull docker image
-        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
-        with:
-          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
-
-      - name: Check if in a container runner
-        shell: bash
-        id: check_container_runner
-        run: echo "IN_CONTAINER_RUNNER=$(if [ -f /.inarc ] || [ -f /.incontainer ]; then echo true ; else echo false; fi)" >> "$GITHUB_OUTPUT"
-
-      - name: Setup GPU_FLAG for docker run
-        id: setup-gpu-flag
-        run: echo "GPU_FLAG=--gpus all -e NVIDIA_DRIVER_CAPABILITIES=all" >> "${GITHUB_ENV}"
-
-      - name: Setup SCCACHE_SERVER_PORT environment for docker run when on container
-        id: setup-sscache-port-flag
-        run: echo "SCCACHE_SERVER_PORT_DOCKER_FLAG=-e SCCACHE_SERVER_PORT=$((RUNNER_UID + 4226))" >> "${GITHUB_ENV}"
-        if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' }}
-
-      - name: Get workflow job id
-        id: get-job-id
-        uses: ./.github/actions/get-workflow-job-id
-        if: always()
-        with:
-          github-token: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Download build artifacts
-        uses: ./.github/actions/download-build-artifacts
-        with:
-          name: ${{ inputs.build-environment }}
-          s3-bucket: ${{ inputs.s3-bucket }}
-
-      - name: Parse ref
-        id: parse-ref
-        run: .github/scripts/parse_ref.py
-
-      - name: Set Test step time
-        id: test-timeout
-        shell: bash
-        env:
-          JOB_TIMEOUT: ${{ inputs.timeout-minutes }}
-        run: |
-          echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
-
-      - name: Preserve github env variables for use in docker
-        shell: bash
-        run: |
-          env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-          env | grep '^CI' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-
-      - name: Test
-        id: test
-        timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
-        env:
-          BUILD_ENVIRONMENT: ${{ inputs.build-environment }}
-          PR_NUMBER: ${{ github.event.pull_request.number }}
-          GITHUB_REPOSITORY: ${{ github.repository }}
-          GITHUB_WORKFLOW: ${{ github.workflow }}
-          GITHUB_JOB: ${{ github.job }}
-          GITHUB_RUN_ID: ${{ github.run_id }}
-          GITHUB_RUN_NUMBER: ${{ github.run_number }}
-          GITHUB_RUN_ATTEMPT: ${{ github.run_attempt }}
-          JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
-          JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
-          BRANCH: ${{ steps.parse-ref.outputs.branch }}
-          SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha || github.sha }}
-          SHM_SIZE: '2g'
-          DOCKER_IMAGE: ${{ inputs.docker-image }}
-          VLLM_TEST_HUGGING_FACE_TOKEN: ${{ secrets.VLLM_TEST_HUGGING_FACE_TOKEN }}
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
-          SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.SCRIBE_GRAPHQL_ACCESS_TOKEN }}
-          ARTIFACTS_FILE_SUFFIX: ${{ github.job }}-${{ steps.get-job-id.outputs.job-id }}
-        run: |
-          set -x
-
-          TEST_COMMAND=.ci/pytorch/test_fa3_abi_stable.sh
-
-          # Leaving 1GB for the runner and other things
-          TOTAL_AVAILABLE_MEMORY_IN_GB=$(awk '/MemTotal/ { printf "%.3f \n", $2/1024/1024 - 1 }' /proc/meminfo)
-          # https://docs.docker.com/engine/containers/resource_constraints/#--memory-swap-details, the 3GB swap
-          # comes from https://github.com/pytorch/test-infra/pull/6058
-          TOTAL_MEMORY_WITH_SWAP=$(("${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}" + 3))
-
-
-          SHM_OPTS="--shm-size=${SHM_SIZE}"
-          JENKINS_USER="--user jenkins"
-          DOCKER_SHELL_CMD=
-
-          # detached container should get cleaned up by teardown_ec2_linux
-          # TODO: Stop building test binaries as part of the build phase
-          # Used for GPU_FLAG, SHM_OPTS, JENKINS_USER and DOCKER_SHELL_CMD since that doesn't play nice
-          # shellcheck disable=SC2086,SC2090
-          container_name=$(docker run \
-            ${GPU_FLAG:-} \
-            ${SCCACHE_SERVER_PORT_DOCKER_FLAG:-} \
-            -e BUILD_ENVIRONMENT \
-            -e PR_NUMBER \
-            -e GITHUB_ACTIONS \
-            -e GITHUB_REPOSITORY \
-            -e GITHUB_WORKFLOW \
-            -e GITHUB_JOB \
-            -e GITHUB_RUN_ID \
-            -e GITHUB_RUN_NUMBER \
-            -e GITHUB_RUN_ATTEMPT \
-            -e JOB_ID \
-            -e JOB_NAME \
-            -e BASE_SHA \
-            -e BRANCH \
-            -e SHA1 \
-            -e MAX_JOBS="$(nproc --ignore=2)" \
-            -e HUGGING_FACE_HUB_TOKEN \
-            -e VLLM_TEST_HUGGING_FACE_TOKEN \
-            -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
-            -e ARTIFACTS_FILE_SUFFIX \
-            --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
-            --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \
-            --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
-            --security-opt seccomp=unconfined \
-            --cap-add=SYS_PTRACE \
-            --ipc=host \
-            ${SHM_OPTS} \
-            --tty \
-            --detach \
-            --name="${container_name}" \
-            ${JENKINS_USER} \
-            -v "${GITHUB_WORKSPACE}:/var/lib/jenkins/workspace" \
-            -w /var/lib/jenkins/workspace \
-            "${DOCKER_IMAGE}" \
-            ${DOCKER_SHELL_CMD}
-          )
-
-          echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
-
-          docker exec -t "${container_name}" sh -c "python3 -m pip install $(echo dist/*.whl)[opt-einsum] && ${TEST_COMMAND}"
-
-      - name: Collect backtraces from coredumps (if any)
-        if: always()
-        run: |
-          # shellcheck disable=SC2156
-          find . -iname "core.[1-9]*" -exec docker exec "${DOCKER_CONTAINER_ID}" sh -c "gdb python {} -ex 'bt' -ex 'q'" \;
-
-      - name: Store Core dumps on S3
-        uses: seemethere/upload-artifact-s3@baba72d0712b404f646cebe0730933554ebce96a # v5.1.0
-        if: failure()
-        with:
-          name: coredumps-fa3-stable-abi-smoke-tests
-          retention-days: 14
-          if-no-files-found: ignore
-          path: ./**/core.[1-9]*
-
-      - name: Upload utilization stats
-        if: ${{ always() && steps.test.conclusion && steps.test.conclusion != 'skipped' }}
-        continue-on-error: true
-        uses: ./.github/actions/upload-utilization-stats
-        with:
-          job_id: ${{ steps.get-job-id.outputs.job-id }}
-          job_name: ${{ steps.get-job-id.outputs.job-name }}
-          workflow_name: ${{ github.workflow }}
-          workflow_run_id: ${{github.run_id}}
-          workflow_attempt: ${{github.run_attempt}}
-
-      - name: Teardown Linux
-        uses: pytorch/test-infra/.github/actions/teardown-linux@main
-        if: always() && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false'
--- a/.github/workflows/_mac-build.yml
+++ b/.github/workflows/_mac-build.yml
@ -85,7 +85,7 @@ jobs:
        uses: pytorch/test-infra/.github/actions/setup-python@main
        with:
          python-version: ${{ inputs.python-version }}
-          pip-requirements-file: .ci/docker/requirements-ci.txt
+          pip-requirements-file: .github/requirements/pip-requirements-macOS.txt

      - name: Install sccache (only for non-forked PRs, and pushes to trunk)
        uses: nick-fields/retry@7152eba30c6575329ac0576536151aca5a72780e # v3.0.0
--- a/.github/workflows/_mac-test.yml
+++ b/.github/workflows/_mac-test.yml
@ -122,7 +122,7 @@ jobs:
        uses: pytorch/test-infra/.github/actions/setup-python@main
        with:
          python-version: ${{ inputs.python-version }}
-          pip-requirements-file: .ci/docker/requirements-ci.txt
+          pip-requirements-file: .github/requirements/pip-requirements-macOS.txt

      - name: Start monitoring script
        id: monitor-script
--- a/.github/workflows/_win-build.yml
+++ b/.github/workflows/_win-build.yml
@ -84,6 +84,9 @@ jobs:
          # in https://github.com/actions/checkout/issues/1018
          git config --global core.fsmonitor false

+      - name: Clean up leftover processes on non-ephemeral Windows runner
+        uses: pytorch/test-infra/.github/actions/cleanup-runner@main
+
      - name: Setup SSH (Click me for login details)
        uses: pytorch/test-infra/.github/actions/setup-ssh@main
        with:
--- a/.github/workflows/_win-test.yml
+++ b/.github/workflows/_win-test.yml
@ -77,6 +77,9 @@ jobs:
          # in https://github.com/actions/checkout/issues/1018
          git config --global core.fsmonitor false

+      - name: Clean up leftover processes on non-ephemeral Windows runner
+        uses: pytorch/test-infra/.github/actions/cleanup-runner@main
+
      - name: Setup SSH (Click me for login details)
        uses: pytorch/test-infra/.github/actions/setup-ssh@main
        with:
@ -103,6 +106,18 @@ jobs:
        with:
          cuda-version: ${{ inputs.cuda-version }}

+      # TODO: Move to a requirements.txt file for windows
+      - name: Install pip dependencies
+        uses: nick-fields/retry@7152eba30c6575329ac0576536151aca5a72780e # v3.0.0
+        with:
+          shell: bash
+          timeout_minutes: 5
+          max_attempts: 5
+          retry_wait_seconds: 30
+          command: |
+            set -eu
+            python3 -m pip install 'xdoctest>=1.1.0'
+
      - name: Get workflow job id
        id: get-job-id
        uses: ./.github/actions/get-workflow-job-id
@ -257,6 +272,15 @@ jobs:
        shell: bash
        run: python3 .github/scripts/parse_ref.py

+      - name: Uninstall PyTorch
+        if: always()
+        continue-on-error: true
+        shell: bash
+        run: |
+          # This step removes PyTorch installed by the test to give a clean slate
+          # to the next job
+          python3 -mpip uninstall -y torch
+
      - name: Teardown Windows
        uses: ./.github/actions/teardown-win
        if: always()
--- a/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
@ -132,7 +132,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -178,7 +178,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -224,7 +224,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -335,7 +335,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_11-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -381,7 +381,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_11-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -427,7 +427,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_11-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -538,7 +538,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_12-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -584,7 +584,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_12-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -630,7 +630,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_12-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -741,7 +741,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -787,7 +787,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -833,7 +833,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -944,7 +944,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13t-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -990,7 +990,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13t-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1036,7 +1036,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13t-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1147,7 +1147,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1193,7 +1193,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1239,7 +1239,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1350,7 +1350,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14t-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1396,7 +1396,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14t-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1442,7 +1442,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14t-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-linux-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-nightly.yml
@ -127,7 +127,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_10-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda12_6-test:  # Testing
@ -193,7 +193,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_10-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda12_8-test:  # Testing
@ -259,7 +259,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_10-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda13_0-test:  # Testing
@ -721,7 +721,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_11-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda12_6-test:  # Testing
@ -787,7 +787,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_11-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda12_8-test:  # Testing
@ -853,7 +853,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_11-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda13_0-test:  # Testing
@ -1315,7 +1315,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_12-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_6-test:  # Testing
@ -1381,7 +1381,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_12-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_8-test:  # Testing
@ -1447,7 +1447,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_12-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda13_0-test:  # Testing
@ -1909,7 +1909,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda12_6-test:  # Testing
@ -1975,7 +1975,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda12_8-test:  # Testing
@ -2041,7 +2041,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda13_0-test:  # Testing
@ -2503,7 +2503,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13t-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda12_6-test:  # Testing
@ -2569,7 +2569,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13t-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda12_8-test:  # Testing
@ -2635,7 +2635,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13t-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda13_0-test:  # Testing
@ -3097,7 +3097,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda12_6-test:  # Testing
@ -3163,7 +3163,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda12_8-test:  # Testing
@ -3229,7 +3229,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda13_0-test:  # Testing
@ -3691,7 +3691,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14t-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda12_6-test:  # Testing
@ -3757,7 +3757,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14t-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda12_8-test:  # Testing
@ -3823,7 +3823,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14t-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda13_0-test:  # Testing
--- a/.github/workflows/inductor-periodic.yml
+++ b/.github/workflows/inductor-periodic.yml
@ -106,16 +106,6 @@ jobs:
          { config: "dynamic_aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.1" },
          { config: "dynamic_aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
          { config: "dynamic_aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_inductor_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_inductor_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_inductor_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_inductor_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
        ]}
    secrets: inherit

--- a/.github/workflows/operator_microbenchmark.yml
+++ b/.github/workflows/operator_microbenchmark.yml
@ -18,7 +18,6 @@ permissions:
  contents: read

 jobs:
-  # H100 A100 runners
  opmicrobenchmark-build:
    if: github.repository_owner == 'pytorch'
    name: opmicrobenchmark-build
@ -45,56 +44,3 @@ jobs:
      docker-image: ${{ needs.opmicrobenchmark-build.outputs.docker-image }}
      test-matrix: ${{ needs.opmicrobenchmark-build.outputs.test-matrix }}
    secrets: inherit
-
-  # B200 runner
-  opmicrobenchmark-build-b200:
-    if: github.repository_owner == 'pytorch'
-    name: opmicrobenchmark-build-b200
-    uses: ./.github/workflows/_linux-build.yml
-    with:
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '10.0'
-      test-matrix: |
-        { include: [
-          { config: "operator_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-        ]}
-    secrets: inherit
-
-  opmicrobenchmark-test-b200:
-    name: opmicrobenchmark-test-b200
-    uses: ./.github/workflows/_linux-test.yml
-    needs: opmicrobenchmark-build-b200
-    with:
-      timeout-minutes: 500
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      docker-image: ${{ needs.opmicrobenchmark-build-b200.outputs.docker-image }}
-      test-matrix: ${{ needs.opmicrobenchmark-build-b200.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-    secrets: inherit
-
-  # ROCM MI300 runner
-  opmicrobenchmark-build-rocm:
-    if: github.repository_owner == 'pytorch'
-    name: opmicrobenchmark-build-rocm
-    uses: ./.github/workflows/_linux-build.yml
-    with:
-      build-environment: linux-jammy-rocm-py3_10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
-      test-matrix: |
-        { include: [
-          { config: "operator_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.1" },
-        ]}
-    secrets: inherit
-
-  opmicrobenchmark-test-rocm:
-    name: opmicrobenchmark-test-rocm
-    uses: ./.github/workflows/_rocm-test.yml
-    needs: opmicrobenchmark-build-rocm
-    with:
-      timeout-minutes: 500
-      build-environment: linux-jammy-rocm-py3_10
-      docker-image: ${{ needs.opmicrobenchmark-build-rocm.outputs.docker-image }}
-      test-matrix: ${{ needs.opmicrobenchmark-build-rocm.outputs.test-matrix }}
-    secrets: inherit
--- a/.github/workflows/periodic.yml
+++ b/.github/workflows/periodic.yml
@ -213,9 +213,9 @@ jobs:
      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
      test-matrix: |
        { include: [
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
        ]}
    secrets: inherit

--- a/.github/workflows/pull.yml
+++ b/.github/workflows/pull.yml
@ -127,6 +127,8 @@ jobs:
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
+      # More memory is needed to build with asan
+      runner: linux.2xlarge.memory
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build-environment: linux-jammy-py3.10-clang18-asan
      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
--- a/.github/workflows/slow.yml
+++ b/.github/workflows/slow.yml
@ -140,6 +140,8 @@ jobs:
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
+      # More memory is needed to build with asan
+      runner: linux.2xlarge.memory
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build-environment: linux-jammy-py3.10-clang18-asan
      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
--- a/.github/workflows/test-h100.yml
+++ b/.github/workflows/test-h100.yml
@ -61,15 +61,3 @@ jobs:
      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm90-build.outputs.docker-image }}
      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm90-build.outputs.test-matrix }}
    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc11-sm90-FA3-ABI-stable-test:
-    name: linux-jammy-cuda12_8-py3_10-gcc11-sm90-FA3-ABI-stable-test
-    uses: ./.github/workflows/_linux-test-stable-fa3.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-sm90-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm90-build.outputs.docker-image }}
-      timeout-minutes: 30
-      s3-bucket: gha-artifacts
-    secrets: inherit
--- a/.github/workflows/trunk.yml
+++ b/.github/workflows/trunk.yml
@ -160,10 +160,9 @@ jobs:
      runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
      test-matrix: |
        { include: [
-          { config: "default", shard: 1, num_shards: 4, runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral" },
-          { config: "default", shard: 2, num_shards: 4, runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral" },
-          { config: "default", shard: 3, num_shards: 4, runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral" },
-          { config: "default", shard: 4, num_shards: 4, runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral" },
+          { config: "default", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral" },
+          { config: "default", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral" },
+          { config: "default", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral" },
        ]}
    secrets: inherit

@ -190,6 +189,41 @@ jobs:
      runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
    secrets: inherit

+  linux-jammy-rocm-py3_10-build:
+    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
+      sync-tag: rocm-build
+      test-matrix: |
+        { include: [
+          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "distributed", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.4" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-test:
+    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor distributed/test_c10d_common distributed/test_c10d_nccl"
+    secrets: inherit
+
  inductor-build:
    name: inductor-build
    uses: ./.github/workflows/_linux-build.yml
--- a/.github/workflows/update-viablestrict.yml
+++ b/.github/workflows/update-viablestrict.yml
@ -23,7 +23,7 @@ jobs:
        with:
          repository: pytorch/pytorch
          stable-branch: viable/strict
-          requires: '[\"pull\", \"trunk\", \"lint\",  \"linux-aarch64\"]'
+          requires: '[\"pull\", \"trunk\", \"lint\", \"^linux-binary-manywheel$\", \"^linux-binary-libtorch-release$\", \"linux-aarch64\"]'
          secret-bot-token: ${{ secrets.MERGEBOT_TOKEN }}
          clickhouse-url: ${{ secrets.CLICKHOUSE_URL }}
          clickhouse-username: ${{ secrets.CLICKHOUSE_VIABLESTRICT_USERNAME }}
@ -48,7 +48,4 @@ jobs:
            echo "{\"sha\": \"${LATEST_SHA}\", \"repository\":\"pytorch/pytorch\", \"timestamp\": ${TIME}}" > "/tmp/${LATEST_SHA}.json"
            pip install awscli==1.29.40
            aws s3 cp "/tmp/${LATEST_SHA}.json" "s3://ossci-raw-job-status/stable_pushes/pytorch/pytorch/${LATEST_SHA}.json"
-            # Push new viable/strict tag
-            cd pytorch/pytorch
-            git push origin "${LATEST_SHA}:refs/tags/viable/strict/${TIME}"
          fi
--- a/.github/workflows/vllm.yml
+++ b/.github/workflows/vllm.yml
@ -42,7 +42,7 @@ jobs:
      build-external-packages: "vllm"
      build-environment: linux-jammy-cuda12.8-py3.12-gcc11
      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm
-      cuda-arch-list: '8.0 8.9 9.0'
+      cuda-arch-list: '8.0;8.9;9.0'
      runner: linux.24xlarge.memory
      test-matrix: |
        { include: [
--- a/.lintrunner.toml
+++ b/.lintrunner.toml
@ -18,7 +18,6 @@ exclude_patterns = [
    'torch/_inductor/autoheuristic/artifacts/**',
    'scripts/**',
    'test/generated_type_hints_smoketest.py',
-    'test/test_torchfuzz_repros.py',
    # CPython tests
    'test/dynamo/cpython/**',
    # Tests from the NumPy test suite
@ -28,7 +27,6 @@ exclude_patterns = [
    'torch/lib/**',
    'venv/**',
    '**/*.pyi',
-    "tools/experimental/dynamic_shapes/torchfuzz/**",
    'tools/test/test_selective_build.py',
 ]
 command = [
@ -1262,7 +1260,6 @@ exclude_patterns = [
    'test/test_masked.py',
    'test/test_maskedtensor.py',
    'test/test_matmul_cuda.py',
-    'test/test_scaled_matmul_cuda.py',
    'test/test_meta.py',
    'test/test_metal.py',
    'test/test_mkl_verbose.py',
@ -1573,7 +1570,6 @@ exclude_patterns = [
    'torch/_inductor/fx_passes/serialized_patterns/**',
    'torch/_inductor/autoheuristic/artifacts/**',
    'test/dynamo/cpython/**',
-    'test/test_torchfuzz_repros.py',
    'scripts/**',
    'third_party/**',
    'fb/**',
--- a/18
+++ b/18
@ -181,15 +181,15 @@ caffe2/utils/hip @jeffdaily @jithunnair-amd
 /torch/csrc/jit/python/init.cpp @mikaylagawarecki

 # CUDA and CUDA math libraries
-aten/src/ATen/cuda/ @eqy @syed-ahmed @Aidyn-A
-aten/src/ATen/cudnn/ @eqy @syed-ahmed @Aidyn-A
-aten/src/ATen/native/cuda/ @eqy @syed-ahmed @Aidyn-A
-aten/src/ATen/native/cudnn/ @eqy @syed-ahmed @Aidyn-A
-c10/cuda @eqy @syed-ahmed @Aidyn-A
-torch/cuda/ @eqy @syed-ahmed @Aidyn-A
-torch/csrc/cuda/ @eqy @syed-ahmed @Aidyn-A
-torch/backends/cuda/ @eqy @syed-ahmed @Aidyn-A
-torch/backends/cudnn/ @eqy @syed-ahmed @Aidyn-A
+aten/src/ATen/cuda/ @eqy @syed-ahmed
+aten/src/ATen/cudnn/ @eqy @syed-ahmed
+aten/src/ATen/native/cuda/ @eqy @syed-ahmed
+aten/src/ATen/native/cudnn/ @eqy @syed-ahmed
+c10/cuda @eqy @syed-ahmed
+torch/cuda/ @eqy @syed-ahmed
+torch/csrc/cuda/ @eqy @syed-ahmed
+torch/backends/cuda/ @eqy @syed-ahmed
+torch/backends/cudnn/ @eqy @syed-ahmed

 # PyTree utilities
 /torch/utils/_pytree.py @XuehaiPan
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@ -81,7 +81,7 @@ git remote add upstream git@github.com:pytorch/pytorch.git
 make setup-env
 # Or run `make setup-env-cuda` for pre-built CUDA binaries
 # Or run `make setup-env-rocm` for pre-built ROCm binaries
-source venv/bin/activate  # or `. .\venv\Scripts\activate` on Windows
+source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
 ```

 ### Tips and Debugging
@ -182,36 +182,28 @@ You can use this script to check out a new nightly branch with the following:

 ```bash
 ./tools/nightly.py checkout -b my-nightly-branch
-source venv/bin/activate  # or `. .\venv\Scripts\activate` on Windows
+source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
 ```

 To install the nightly binaries built with CUDA, you can pass in the flag `--cuda`:

 ```bash
 ./tools/nightly.py checkout -b my-nightly-branch --cuda
-source venv/bin/activate  # or `. .\venv\Scripts\activate` on Windows
+source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
 ```

 To install the nightly binaries built with ROCm, you can pass in the flag `--rocm`:

 ```bash
 ./tools/nightly.py checkout -b my-nightly-branch --rocm
-source venv/bin/activate  # or `. .\venv\Scripts\activate` on Windows
+source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
 ```

 You can also use this tool to pull the nightly commits into the current branch:

 ```bash
-./tools/nightly.py pull
-source venv/bin/activate  # or `. .\venv\Scripts\activate` on Windows
-```
-
-To create the virtual environment with a specific Python interpreter, you can
-pass in the `--python` argument:
-
-```bash
-./tools/nightly.py --python /path/to/python3.12
-source venv/bin/activate  # or `. .\venv\Scripts\activate` on Windows
+./tools/nightly.py pull -p my-env
+source my-env/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
 ```

 Pulling will recreate a fresh virtual environment and reinstall the development
--- a/5
+++ b/5
@ -50,10 +50,11 @@ RUN git submodule update --init --recursive
 FROM conda as conda-installs
 ARG PYTHON_VERSION=3.11
 ARG CUDA_PATH=cu121
+ARG CUDA_CHANNEL=nvidia
 ARG INSTALL_CHANNEL=whl/nightly
 # Automatically set by buildx
-# pinning version of conda here see: https://github.com/pytorch/pytorch/issues/164574
-RUN /opt/conda/bin/conda install -c "${INSTALL_CHANNEL}" -y python=${PYTHON_VERSION} conda=25.7.0
+RUN /opt/conda/bin/conda update -y -n base -c defaults conda
+RUN /opt/conda/bin/conda install -y python=${PYTHON_VERSION}

 ARG TARGETPLATFORM

--- a/RELEASE.md
+++ b/RELEASE.md
@ -3,7 +3,6 @@
 <!-- toc -->

  - [Release Compatibility Matrix](#release-compatibility-matrix)
-    - [PyTorch CUDA Support Matrix](#pytorch-cuda-support-matrix)
  - [Release Cadence](#release-cadence)
  - [General Overview](#general-overview)
    - [Frequently Asked Questions](#frequently-asked-questions)
@ -64,22 +63,6 @@ Following is the Release Compatibility Matrix for PyTorch releases:
 | 1.13 | >=3.7, <=3.10 | C++14 | CUDA 11.6, CUDNN 8.3.2.44 | CUDA 11.7, CUDNN 8.5.0.96 | ROCm 5.2 |
 | 1.12 | >=3.7, <=3.10 | C++14 | CUDA 11.3, CUDNN 8.3.2.44 | CUDA 11.6, CUDNN 8.3.2.44 | ROCm 5.0 |

-### PyTorch CUDA Support Matrix
-
-For Release 2.9 PyTorch Supports following CUDA Architectures:
-
-| CUDA | architectures supported for Linux x86 and Windows builds | notes |
-| --- | --- | --- |
-| 12.6.3 | Maxwell(5.0), Pascal(6.0), Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0) | |
-| 12.8.1 | Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0), Blackwell(10.0, 12.0)  | |
-| 13.0.0 | Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0), Blackwell(10.0, 12.0+PTX) | +PTX available on linux builds only |
-
-| CUDA | architectures supported for Linux aarch64 builds |
-| --- | --- |
-| 12.6.3 | Ampere(8.0), Hopper(9.0) |
-| 12.8.1 | Ampere(8.0), Hopper(9.0), Blackwell(10.0, 12.0)  |
-| 13.0.0 | Ampere(8.0), Hopper(9.0), Blackwell(10.0, 11.0, 12.0+PTX) |
-
 ## Release Cadence

 Following is the release cadence. All future dates below are tentative. For latest updates on the release schedule, please follow [dev discuss](https://dev-discuss.pytorch.org/c/release-announcements/27). Please note: Patch Releases are optional.
--- a/aten/src/ATen/CMakeLists.txt
+++ b/aten/src/ATen/CMakeLists.txt
@ -605,11 +605,6 @@ if(UNIX)
  if(HAVE_MALLOC_USABLE_SIZE)
    add_definitions(-DHAVE_MALLOC_USABLE_SIZE=1)
  endif(HAVE_MALLOC_USABLE_SIZE)
-  set(CMAKE_EXTRA_INCLUDE_FILES "fcntl.h")
-  CHECK_FUNCTION_EXISTS(posix_fallocate HAVE_POSIX_FALLOCATE)
-  if(HAVE_POSIX_FALLOCATE)
-    add_definitions(-DHAVE_POSIX_FALLOCATE=1)
-  endif(HAVE_POSIX_FALLOCATE)
 endif(UNIX)

 ADD_DEFINITIONS(-DUSE_EXTERNAL_MZCRC)
--- a/aten/src/ATen/Context.cpp
+++ b/aten/src/ATen/Context.cpp
@ -40,6 +40,41 @@ namespace {
                ->conv
                ->rnn
 */
+const std::map<std::string, std::vector<std::string>> _fp32_precisions = {
+    {"generic", {{"ieee", "tf32", "bf16", "none"}}},
+    {"mkldnn", {{"ieee", "tf32", "bf16", "none"}}},
+    {"cuda", {{"ieee", "tf32", "none"}}}};
+
+// Check whether the backend and op are legal
+void check_fp32_prec_backend_and_op(
+    const std::string& backend,
+    const std::string& op) {
+  static std::vector<std::string> backends = {"generic", "mkldnn", "cuda"};
+  static std::vector<std::string> operators = {"conv", "matmul", "rnn", "all"};
+  TORCH_CHECK(
+      std::find(backends.begin(), backends.end(), backend) != backends.end(),
+      "Invalid backend: ",
+      backend);
+  TORCH_CHECK(
+      std::find(operators.begin(), operators.end(), op) != operators.end(),
+      "Invalid operator: ",
+      op);
+  if (backend == "generic") {
+    TORCH_CHECK(op == "all", "Invalid operation for generic backend: ", op);
+  }
+  }
+
+  // Return whether the precision is supported by backends
+  bool validate_fp32_prec(
+      const std::string& backend,
+      const std::string& precision) {
+    auto iterp = _fp32_precisions.find(backend);
+    TORCH_CHECK(iterp != _fp32_precisions.end());
+    auto precisions = iterp->second;
+    bool valid = std::find(precisions.begin(), precisions.end(), precision) !=
+        precisions.end();
+    return valid;
+  }

  C10_ALWAYS_INLINE void warn_deprecated_fp32_precision_api(){
    TORCH_WARN_ONCE(
@ -51,54 +86,6 @@ namespace {
  }
 } // namespace

-Float32Backend str2backend(const std::string& name) {
-  if (name == "generic")
-    return Float32Backend::GENERIC;
-  else if (name == "cuda")
-    return Float32Backend::CUDA;
-  else if (name == "mkldnn")
-    return Float32Backend::MKLDNN;
-  TORCH_CHECK(false, "Unknown backend: ", name);
-}
-
-Float32Op str2op(const std::string& name) {
-  if (name == "all")
-    return Float32Op::ALL;
-  else if (name == "conv")
-    return Float32Op::CONV;
-  else if (name == "rnn")
-    return Float32Op::RNN;
-  else if (name == "matmul")
-    return Float32Op::MATMUL;
-  TORCH_CHECK(false, "Unknown op: ", name);
-}
-
-Float32Precision str2precision(const std::string& name) {
-  if (name == "none")
-    return Float32Precision::NONE;
-  else if (name == "ieee")
-    return Float32Precision::IEEE;
-  else if (name == "tf32")
-    return Float32Precision::TF32;
-  else if (name == "bf16")
-    return Float32Precision::BF16;
-  TORCH_CHECK(false, "Unknown precision: ", name);
-}
-
-std::string precision2str(Float32Precision prec) {
-  switch (prec) {
-    case Float32Precision::NONE:
-      return "none";
-    case Float32Precision::IEEE:
-      return "ieee";
-    case Float32Precision::TF32:
-      return "tf32";
-    case Float32Precision::BF16:
-      return "bf16";
-  }
-  TORCH_CHECK(false, "Invalid enum Float32Precision(", static_cast<int>(prec), ")");
-}
-
 Context::Context() = default;

 // TODO: This could be bad juju if someone calls globalContext() in the
@ -192,10 +179,10 @@ void Context::setUserEnabledNNPACK(bool e) {
  enabled_nnpack = e;
 }

-bool Context::allowTF32CuDNN(std::optional<Float32Op> op) const {
-  if (!op.has_value()) {
-    bool allow_tf32_rnn = float32Precision(Float32Backend::CUDA, Float32Op::RNN) == Float32Precision::TF32;
-    bool allow_tf32_conv = float32Precision(Float32Backend::CUDA, Float32Op::CONV) == Float32Precision::TF32;
+bool Context::allowTF32CuDNN(const std::string& op) const {
+  if (op.empty()){
+    bool allow_tf32_rnn = float32Precision("cuda", "rnn") == "tf32";
+    bool allow_tf32_conv = float32Precision("cuda", "conv") == "tf32";
    TORCH_CHECK(
        allow_tf32_rnn == allow_tf32_conv && allow_tf32_rnn == allow_tf32_cudnn,
        "PyTorch is checking whether allow_tf32 is enabled for cuDNN without a specific operator name,",
@ -204,15 +191,15 @@ bool Context::allowTF32CuDNN(std::optional<Float32Op> op) const {
        "We suggest only using the new API to set the TF32 flag(s). See also: ",
        "https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices");
  } else {
-    return float32Precision(Float32Backend::CUDA, op.value()) == Float32Precision::TF32;
+    return float32Precision("cuda", op) == "tf32";
  }
  warn_deprecated_fp32_precision_api();
  return allow_tf32_cudnn;
 }

 void Context::setAllowTF32CuDNN(bool b) {
-  setFloat32Precision(Float32Backend::CUDA, Float32Op::RNN, b ? Float32Precision::TF32 : Float32Precision::NONE);
-  setFloat32Precision(Float32Backend::CUDA, Float32Op::CONV, b ? Float32Precision::TF32 : Float32Precision::NONE);
+  setFloat32Precision("cuda", "rnn", b ? "tf32" : "none");
+  setFloat32Precision("cuda", "conv", b ? "tf32" : "none");
  allow_tf32_cudnn = b;
  warn_deprecated_fp32_precision_api();
 }
@ -292,6 +279,42 @@ bool Context::userEnabledOverrideableSDP() const {
  return enabled_overrideable;
 }

+static constexpr const auto cublas_config_var_name = "CUBLAS_WORKSPACE_CONFIG";
+static constexpr const std::array<const char*, 2> cublas_deterministic_configs = {":4096:8", ":16:8"};
+
+bool Context::checkCuBLASConfigDeterministic() {
+  // If using CUDA 10.2 or greater, need to make sure CuBLAS workspace config
+  // is set to deterministic setting
+  if (hasCUDART()) {
+    const auto workspace_config = c10::utils::get_env(cublas_config_var_name);
+    return (workspace_config == cublas_deterministic_configs[0] || workspace_config == cublas_deterministic_configs[1]);
+  }
+  return true;
+}
+
+void Context::alertCuBLASConfigNotDeterministic() const {
+  static const bool cublas_config_deterministic = checkCuBLASConfigDeterministic();
+  if (C10_LIKELY(!deterministicAlgorithms() || cublas_config_deterministic)) {
+    return;
+  }
+
+  auto msg = c10::str(
+    "Deterministic behavior was enabled with either `torch.use_deterministic_algorithms(True)` or ",
+    "`at::Context::setDeterministicAlgorithms(true)`, but this operation is not deterministic because ",
+    "it uses CuBLAS and you have CUDA >= 10.2. To enable deterministic behavior in this ",
+    "case, you must set an environment variable before running your PyTorch application: ",
+    cublas_config_var_name, "=", cublas_deterministic_configs[0], " or ",
+    cublas_config_var_name, "=", cublas_deterministic_configs[1], ". For more information, go to ",
+    "https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility"
+  );
+
+  if (deterministicAlgorithmsWarnOnly()) {
+    TORCH_WARN(msg);
+  } else {
+    TORCH_CHECK(false, msg);
+  }
+}
+
 bool Context::benchmarkCuDNN() const {
  return benchmark_cudnn;
 }
@ -318,7 +341,7 @@ void Context::setImmediateMiopen(bool b) {

 bool Context::allowTF32CuBLAS() const {
  bool legacy_allow_tf32 = float32_matmul_precision != at::Float32MatmulPrecision::HIGHEST;
-  bool allow_tf32_new = float32Precision(Float32Backend::CUDA, Float32Op::MATMUL) == Float32Precision::TF32;
+  bool allow_tf32_new = float32Precision("cuda", "matmul") == "tf32";
  TORCH_CHECK(
      legacy_allow_tf32 == allow_tf32_new,
      "PyTorch is checking whether allow_tf32_new is enabled for cuBlas matmul,",
@ -331,17 +354,17 @@ bool Context::allowTF32CuBLAS() const {

 void Context::setAllowTF32CuBLAS(bool b) {
  float32_matmul_precision = b ? at::Float32MatmulPrecision::HIGH : at::Float32MatmulPrecision::HIGHEST;
-  setFloat32Precision(Float32Backend::CUDA, Float32Op::MATMUL, b ? Float32Precision::TF32 : Float32Precision::IEEE);
+  setFloat32Precision("cuda", "matmul", b ? "tf32" : "ieee");
 }

 Float32MatmulPrecision Context::float32MatmulPrecision() const {
-  bool invalid = float32Precision(Float32Backend::CUDA, Float32Op::MATMUL) == Float32Precision::TF32 &&
+  bool invalid = float32Precision("cuda", "matmul") == "tf32" &&
      float32_matmul_precision == at::Float32MatmulPrecision::HIGHEST;
  invalid = invalid ||
-      (float32Precision(Float32Backend::MKLDNN, Float32Op::MATMUL) == Float32Precision::BF16 &&
+      (float32Precision("mkldnn", "matmul") == "bf16" &&
       float32_matmul_precision != at::Float32MatmulPrecision::MEDIUM);
  invalid = invalid ||
-      (float32Precision(Float32Backend::MKLDNN, Float32Op::MATMUL) == Float32Precision::TF32 &&
+      (float32Precision("mkldnn", "matmul") == "tf32" &&
       float32_matmul_precision != at::Float32MatmulPrecision::HIGH);
  TORCH_CHECK(
      !invalid,
@ -353,26 +376,15 @@ Float32MatmulPrecision Context::float32MatmulPrecision() const {
  return float32_matmul_precision;
 }

-Float32Precision Context::float32Precision(Float32Backend backend, Float32Op op) const {
-  std::pair<Float32Backend, Float32Op> key{backend, op};
-  auto it = fp32_precision.find(key);
-  TORCH_CHECK(it != fp32_precision.end(), "Invalid (backend, op) pair: (", backend, ", ", op, ")");
-
-  Float32Precision precision = it->second;
-  if (precision == Float32Precision::NONE) {
-    key.second = Float32Op::ALL;
-    precision = fp32_precision.find(key)->second;
-  }
-  if (precision == Float32Precision::NONE) {
-    key.first = Float32Backend::GENERIC;
-    precision = fp32_precision.find(key)->second;
-  }
-
-  // "cuda" does not support "bf16"
-  if (backend == Float32Backend::CUDA && precision == Float32Precision::BF16) {
-    return Float32Precision::NONE;
-  }
-  return precision;
+std::string Context::float32Precision(const std::string& backend, const std::string& op) const {
+  check_fp32_prec_backend_and_op(backend, op);
+  auto precision = fp32_precision.find(backend)->second.find(op)->second;
+  if (precision == "none")
+    precision = fp32_precision.find(backend)->second.find("all")->second;
+  if (precision == "none")
+    precision = fp32_precision.find("generic")->second.find("all")->second;
+  bool valid_prec = validate_fp32_prec(backend, precision);
+  return valid_prec ? precision : "none";
 }

 void Context::setFloat32MatmulPrecision(const std::string &s) {
@ -381,18 +393,18 @@ void Context::setFloat32MatmulPrecision(const std::string &s) {
    // TODO: consider if CuDNN field needs to also be set for potential future CuDNN ops like multi-headed attention
    if (s_ == "highest") {
      float32_matmul_precision = at::Float32MatmulPrecision::HIGHEST;
-      setFloat32Precision(Float32Backend::CUDA, Float32Op::MATMUL, Float32Precision::IEEE);
-      setFloat32Precision(Float32Backend::MKLDNN, Float32Op::MATMUL, Float32Precision::IEEE);
+      setFloat32Precision("cuda", "matmul", "ieee");
+      setFloat32Precision("mkldnn", "matmul", "ieee");
      return true;
    } else if (s_ == "high") {
      float32_matmul_precision = at::Float32MatmulPrecision::HIGH;
-      setFloat32Precision(Float32Backend::CUDA, Float32Op::MATMUL, Float32Precision::TF32);
-      setFloat32Precision(Float32Backend::MKLDNN, Float32Op::MATMUL, Float32Precision::TF32);
+      setFloat32Precision("cuda", "matmul", "tf32");
+      setFloat32Precision("mkldnn", "matmul", "tf32");
      return true;
    } else if (s_ == "medium") {
      float32_matmul_precision = at::Float32MatmulPrecision::MEDIUM;
-      setFloat32Precision(Float32Backend::CUDA, Float32Op::MATMUL, Float32Precision::TF32);
-      setFloat32Precision(Float32Backend::MKLDNN, Float32Op::MATMUL, Float32Precision::BF16);
+      setFloat32Precision("cuda", "matmul", "tf32");
+      setFloat32Precision("mkldnn", "matmul", "bf16");
      return true;
    }
    return false;
@ -406,16 +418,25 @@ void Context::setFloat32MatmulPrecision(const std::string &s) {
    "setFloat32MatmulPrecision call has no effect.");
 }

-void Context::setFloat32Precision(Float32Backend backend, Float32Op op, Float32Precision p) {
-  auto it = fp32_precision.find(std::make_pair(backend, op));
-  TORCH_CHECK(
-      it != fp32_precision.end(),
-      "Invalid (backend, op) pair: (", backend, ", ", op, ")");
-  TORCH_CHECK(
-      !(backend == Float32Backend::CUDA && p == Float32Precision::BF16),
-      "backend 'cuda' does not support precision 'bf16'");
-
-  it->second = p;
+void Context::setFloat32Precision(const std::string& backend, const std::string& op, const std::string& p) {
+  check_fp32_prec_backend_and_op(backend, op);
+  if (validate_fp32_prec(backend, p)) {
+    fp32_precision[backend][op] = p;
+  } else {
+    std::string msg;
+    auto iterp = _fp32_precisions.find(backend);
+    TORCH_CHECK(iterp != _fp32_precisions.end());
+    for (const auto& p : iterp->second) {
+      msg += p;
+      msg += " ";
+    }
+    TORCH_WARN(
+        "you have set wrong precision for backend:",
+        backend,
+        " setFloat32Precision call has no effect.",
+        "Please choose precision from: ",
+        msg);
+  }
 }

 at::LinalgBackend Context::linalgPreferredBackend() const {
@ -587,33 +608,20 @@ void Context::setROCmFAPreferredBackend(at::ROCmFABackend b) {
  rocm_fa_preferred_backend = b;
 }

-CuBLASReductionOption Context::allowFP16ReductionCuBLAS() const {
+bool Context::allowFP16ReductionCuBLAS() const {
  return allow_fp16_reduction_cublas;
 }

-CuBLASReductionOption inline get_reduction_option(bool allow_reduced_precision, bool allow_splitk) {
-  TORCH_CHECK(
-      !(allow_reduced_precision && !allow_splitk),
-      "allow_splitk=False is not supported when reduced precision reductions are enabled");
-  if (allow_reduced_precision) {
-    return CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
-  } else if (allow_splitk) {
-    return CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK;
-  } else {
-    return CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK;
-  }
+void Context::setAllowFP16ReductionCuBLAS(bool b) {
+  allow_fp16_reduction_cublas = b;
 }

-void Context::setAllowFP16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
-  allow_fp16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
-}
-
-CuBLASReductionOption Context::allowBF16ReductionCuBLAS() const {
+bool Context::allowBF16ReductionCuBLAS() const {
  return allow_bf16_reduction_cublas;
 }

-void Context::setAllowBF16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
-  allow_bf16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
+void Context::setAllowBF16ReductionCuBLAS(bool b) {
+  allow_bf16_reduction_cublas = b;
 }

 bool Context::allowFP16AccumulationCuBLAS() const {
--- a/aten/src/ATen/Context.h
+++ b/aten/src/ATen/Context.h
@ -25,13 +25,11 @@
 #include <c10/util/CallOnce.h>
 #include <c10/util/Exception.h>
 #include <c10/util/env.h>
-#include <c10/util/hash.h>
 #include <c10/util/irange.h>

 #include <cstdint>
 #include <map>
 #include <mutex>
-#include <unordered_map>

 namespace at {

@ -39,20 +37,6 @@ class Tensor;

 enum class TORCH_API Float32MatmulPrecision { HIGHEST, HIGH, MEDIUM };

-enum class CuBLASReductionOption : uint8_t {
-  AllowReducedPrecisionWithSplitK = 0,
-  DisallowReducedPrecisionAllowSplitK = 1,
-  DisallowReducedPrecisionDisallowSplitK = 2,
-};
-enum class TORCH_API Float32Backend { GENERIC, CUDA, MKLDNN };
-enum class TORCH_API Float32Op { ALL, CONV, RNN, MATMUL };
-enum class TORCH_API Float32Precision { NONE, IEEE, TF32, BF16 };
-
-TORCH_API Float32Backend str2backend(const std::string& name);
-TORCH_API Float32Op str2op(const std::string& name);
-TORCH_API Float32Precision str2precision(const std::string& name);
-TORCH_API std::string precision2str(Float32Precision prec);
-
 class TORCH_API Context {
 public:
  Context();
@ -326,7 +310,13 @@ class TORCH_API Context {
  //
  // * Throw an error when `Context::deterministicAlgorithms()` is true. Most
  //   of the time, this should be accomplished by calling
-  //   `at::globalContext().alertNotDeterminstic().
+  //   `at::globalContext().alertNotDeterminstic()`.  However, if the
+  //   nondeterministic behavior is caused by the CuBLAS workspace
+  //   configuration in CUDA >= 10.2,
+  //   `at::globalContext().alertCuBLASConfigNotDeterministic()` should be
+  //   called instead (in this case, a comment explaining why the operation is
+  //   nondeterministic is not necessary). See below for details on these
+  //   methods.
  //
  // * Have an entry in the list of nondeterministic PyTorch operations in the
  //   docstring of `use_deterministic_algorithms()` in torch/__init__.py
@ -350,27 +340,31 @@ class TORCH_API Context {
  // Throws an error if `Context::deterministicAlgorithms()` is true
  static void alertNotDeterministic(std::string_view const& caller);

+  // Throws an error if `Context::deterministicAlgorithms()` is true, CUDA
+  // >= 10.2, and CUBLAS_WORKSPACE_CONFIG is not set to either ":16:8" or
+  // ":4096:8". For more details:
+  // https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility
+  void alertCuBLASConfigNotDeterministic() const;
+
  void setFloat32MatmulPrecision(const std::string& s);
  void setFloat32Precision(
-      Float32Backend backend,
-      Float32Op op,
-      Float32Precision p);
-  bool allowTF32CuDNN(std::optional<Float32Op> op = std::nullopt) const;
+      const std::string& backend,
+      const std::string& op,
+      const std::string& s);
+  bool allowTF32CuDNN(const std::string& op = std::string()) const;
  void setAllowTF32CuDNN(bool);
  bool allowTF32OneDNN() const;
  void setAllowTF32OneDNN(bool);
  bool allowTF32CuBLAS() const;
  void setAllowTF32CuBLAS(bool);
  Float32MatmulPrecision float32MatmulPrecision() const;
-  Float32Precision float32Precision(Float32Backend backend, Float32Op op) const;
-  CuBLASReductionOption allowFP16ReductionCuBLAS() const;
-  void setAllowFP16ReductionCuBLAS(
-      bool allow_reduced_precision,
-      bool allow_splitk = true);
-  CuBLASReductionOption allowBF16ReductionCuBLAS() const;
-  void setAllowBF16ReductionCuBLAS(
-      bool allow_reduced_precision,
-      bool allow_splitk = true);
+  std::string float32Precision(
+      const std::string& backend,
+      const std::string& op) const;
+  bool allowFP16ReductionCuBLAS() const;
+  void setAllowFP16ReductionCuBLAS(bool);
+  bool allowBF16ReductionCuBLAS() const;
+  void setAllowBF16ReductionCuBLAS(bool);
  bool allowFP16AccumulationCuBLAS() const;
  void setAllowFP16AccumulationCuBLAS(bool);

@ -435,6 +429,7 @@ class TORCH_API Context {
  }

 private:
+  static bool checkCuBLASConfigDeterministic();
  std::array<c10::once_flag, at::COMPILE_TIME_MAX_DEVICE_TYPES> init_;
  bool enabled_cudnn = true;
  bool deterministic_cudnn = false;
@ -462,10 +457,8 @@ class TORCH_API Context {
      : at::Float32MatmulPrecision::HIGHEST;
  int benchmark_limit_cudnn = 10;
  bool allow_tf32_cudnn = true;
-  CuBLASReductionOption allow_fp16_reduction_cublas =
-      CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
-  CuBLASReductionOption allow_bf16_reduction_cublas =
-      CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
+  bool allow_fp16_reduction_cublas = true;
+  bool allow_bf16_reduction_cublas = true;
  bool allow_fp16_accumulation_cublas = false;
  std::optional<int32_t> sm_carveout = std::nullopt;
  bool enabled_mkldnn = true;
@ -495,20 +488,21 @@ class TORCH_API Context {
  bool enable_sparse_tensor_invariant_checks = false;
  bool allow_fp16_reduction_cpu = false;

-  using Key = std::pair<Float32Backend, Float32Op>;
-  std::unordered_map<Key, Float32Precision, c10::hash<Key>> fp32_precision = {
-      {{Float32Backend::GENERIC, Float32Op::ALL}, Float32Precision::NONE},
-      {{Float32Backend::MKLDNN, Float32Op::ALL}, Float32Precision::NONE},
-      {{Float32Backend::MKLDNN, Float32Op::CONV}, Float32Precision::NONE},
-      {{Float32Backend::MKLDNN, Float32Op::RNN}, Float32Precision::NONE},
-      {{Float32Backend::MKLDNN, Float32Op::MATMUL}, Float32Precision::NONE},
-      {{Float32Backend::CUDA, Float32Op::ALL}, Float32Precision::NONE},
-      {{Float32Backend::CUDA, Float32Op::CONV}, Float32Precision::TF32},
-      {{Float32Backend::CUDA, Float32Op::RNN}, Float32Precision::TF32},
-      {{Float32Backend::CUDA, Float32Op::MATMUL},
-       float32_matmul_precision == at::Float32MatmulPrecision::HIGHEST
-           ? Float32Precision::NONE
-           : Float32Precision::TF32},
+  std::map<std::string, std::map<std::string, std::string>> fp32_precision = {
+      {"generic", {{"all", "none"}}},
+      {"mkldnn",
+       {{"matmul", "none"},
+        {"conv", "none"},
+        {"rnn", "none"},
+        {"all", "none"}}},
+      {"cuda",
+       {{"matmul",
+         float32_matmul_precision == at::Float32MatmulPrecision::HIGHEST
+             ? "none"
+             : "tf32"},
+        {"conv", "tf32"},
+        {"rnn", "tf32"},
+        {"all", "none"}}},
  };

  Allocator* prev_allocator_ptr_{nullptr};
@ -690,4 +684,5 @@ struct TORCH_API ROCmBackwardPassGuard {
  ~ROCmBackwardPassGuard();
  static bool is_backward_pass();
 };
+
 } // namespace at
--- a/aten/src/ATen/MapAllocator.cpp
+++ b/aten/src/ATen/MapAllocator.cpp
@ -292,28 +292,6 @@ MapAllocator::MapAllocator(WithFd, std::string_view filename, int fd, int flags,
          if (ftruncate(fd, static_cast<off_t>(size)) == -1) {
            TORCH_CHECK(false, "unable to resize file <", filename_, "> to the right size: ", c10::utils::str_error(errno), " (", errno, ")");
          }
-
-#ifdef HAVE_POSIX_FALLOCATE
-          if (flags_ & ALLOCATOR_MAPPED_SHAREDMEM) {
-            for (;;) {
-              if (posix_fallocate(fd, 0, static_cast<off_t>(size)) == 0) {
-                break;
-              }
-
-              if (errno == EINTR) {
-                continue;
-              }
-
-              if (errno == EINVAL || errno == EOPNOTSUPP) {
-                // the underlying filesystem does not support the operation
-                break;
-              }
-
-              TORCH_CHECK(false, "unable to allocate shared memory(shm) for file <", filename_, ">: ", c10::utils::str_error(errno), " (", errno, ")");
-            }
-          }
-#endif
-
          if (fstat(fd, &file_stat) == -1 || file_stat.st_size < static_cast<int64_t>(size)) {
 #ifndef STRIP_ERROR_MESSAGES
            int last_err = errno;
--- a/aten/src/ATen/NamedTensorUtils.cpp
+++ b/aten/src/ATen/NamedTensorUtils.cpp
@ -179,7 +179,7 @@ void propagate_names_except(const Tensor& result, const Tensor& src, IntArrayRef
    return;
  }
  const auto src_names = src.names();
-  const auto result_dim = result.dim();
+  const auto result_dim = static_cast<int64_t>(result.dim());
  const auto src_dim = static_cast<int64_t>(src_names.size());
  const auto excluded_dim = static_cast<int64_t>(excluded_idxs.size());
  TORCH_INTERNAL_ASSERT(src_dim - excluded_dim == result_dim);
--- a/aten/src/ATen/TensorIndexing.h
+++ b/aten/src/ATen/TensorIndexing.h
@ -214,7 +214,7 @@ inline Tensor applySlice(
      "step must be greater than zero");

  // See NOTE [nested tensor size for indexing]
-  if (self_sizes.has_value() && !self_sizes.value().empty()) {
+  if (self_sizes.has_value() && self_sizes.value().size() > 0) {
    // Skip this optimization if we are tracing, as the trace may be polymorphic
    // over the shape of the `self` tensor, and we still want to record
    // the slice.
--- a/aten/src/ATen/TensorUtils.cpp
+++ b/aten/src/ATen/TensorUtils.cpp
@ -273,11 +273,11 @@ void checkLayout(CheckedFrom c, at::ArrayRef<Tensor> tensors, at::Layout layout)
 }

 void * maybe_data_ptr(const Tensor& tensor) {
-  return tensor.defined() ? tensor.data_ptr() : nullptr;
+  return tensor.defined() ? (void *)tensor.data_ptr() : nullptr;
 }

 void * maybe_data_ptr(const TensorArg& tensor) {
-  return tensor->defined() ? tensor->data_ptr() : nullptr;
+  return tensor->defined() ? (void *)tensor->data_ptr() : nullptr;
 }

 void check_dim_size(
--- a/aten/src/ATen/core/CachingHostAllocator.h
+++ b/aten/src/ATen/core/CachingHostAllocator.h
@ -6,7 +6,6 @@
 #include <c10/core/thread_pool.h>
 #include <c10/util/flat_hash_map.h>
 #include <c10/util/llvmMathExtras.h>
-#include <iostream>
 #include <optional>

 #include <deque>
@ -50,57 +49,19 @@ namespace {
  constexpr size_t MAX_SIZE_INDEX = 64;
 }

-// A large reserved pinned memory segment that is created in advance which is used
-// to allocate small pinned memory requests to avoid calling into expensive APIs.
-// We never free this memory and move up the pointer as we allocate new blocks
-// and when blocks are freed, they are cached in the free lists.
-struct PinnedReserveSegment {
-  PinnedReserveSegment(void *start, size_t size) : start_(start), size_(size),
-    current_ptr_(start_), initialized_(true) {}
-
-  PinnedReserveSegment() : start_(nullptr), size_(0), current_ptr_(nullptr), initialized_(false) {}
-
-  bool initialized() {
-    return initialized_;
-  }
-
-  void* allocate(size_t bytes) {
-    std::lock_guard<std::mutex> guard(mutex_);
-
-    // Round up the requested size to 4KB boundary for all including the small ones.
-    size_t rounded_bytes = (bytes + 4096 - 1) & ~(4096 - 1);
-
-    if (((uint8_t*)current_ptr_ + rounded_bytes) > ((uint8_t*)start_ + size_)) {
-      return nullptr;
-    }
-
-    void* ptr = current_ptr_;
-    current_ptr_ = (uint8_t*)current_ptr_ + rounded_bytes;
-    return ptr;
-  }
-
-  bool owns(void* ptr) {
-    return ptr >= start_ && ptr < (uint8_t*)start_ + size_;
-  }
-
-  std::mutex mutex_;
-  void* start_;
-  size_t size_;
-  void* current_ptr_;
-  bool initialized_;
-};
-
 // Struct containing memory allocator summary statistics for host.
 struct TORCH_API HostStats {
-  // COUNT: total allocations (active)
-  Stat active_requests;
-  // SUM: bytes allocated/reserved by this memory alocator. (active)
-  Stat active_bytes;
-  // COUNT: total allocations (active + free)
-  Stat allocations;
-  // SUM: bytes allocated/reserved by this memory alocator. This accounts
-  // for both free and in-use blocks.
+  // COUNT: allocations requested by client code. Note that active
+  // count can be extracted by looking at current allocations
+  Stat allocation;
+  // COUNT: number of allocated segments from host memory allocation.
+  Stat segment;
+
+  // SUM: bytes allocated by this memory alocator. Note that active bytes
+  // can be extracted by looking at current bytes allocated
  Stat allocated_bytes;
+  // SUM: bytes reserved by this memory allocator (both free and used)
+  Stat reserved_bytes;

  // SUM: time spent in cudaHostAlloc/cudaHostRegister in microseconds
  DurationStat host_alloc_time;
@ -114,9 +75,6 @@ struct TORCH_API HostStats {

  // COUNT: number of times cudaHostFree/cudaHostUnregister was called.
  int64_t num_host_free = 0; // This is derived from segment or timing
-
-  // Count of cudaHostAlloc/cudaHostRegister per bucket
-  std::vector<int64_t> bucket_allocation = std::vector<int64_t>(MAX_SIZE_INDEX);
 };

 // Struct containing memory allocator summary statistics for host, as they
@ -124,22 +82,17 @@ struct TORCH_API HostStats {
 // avoid locking the allocator while collecting stats.
 struct alignas(64) HostStatsStaged {
  std::mutex timing_mutex_;
-  // COUNT: total allocations (active + free)
+  // COUNT: allocations requested by client code resulting in a new segment/block allocation
+  // LOCK: access to this stat is protected by the allocator's blocks_mutex_
+  Stat allocation;
+  // SUM: bytes within active memory blocks, including blocks that are
+  // currently in the free list.
  // LOCK: access to this stat is protected by the allocator's blocks_mutex_
-  Stat allocations;
-  // SUM: bytes allocated/reserved by this memory alocator. This accounts
-  // for both free and in-use blocks.
  Stat allocated_bytes;
-  // COUNT: number of allocations per bucket (active)
-  // LOCK: access to this stat is protected by the per bucket free_list_[index].mutex_
-  std::vector<Stat> active_bucket_stats = std::vector<Stat>(MAX_SIZE_INDEX);
-  // SUM: bytes of allocation per bucket (active)
-  // LOCK: access to this stat is protected by the per bucket free_list_[index].mutex_
-  std::vector<Stat> active_bytes_bucket_stats = std::vector<Stat>(MAX_SIZE_INDEX);
-  // COUNT: number of allocations per bucket (active + free)
+  // COUNT: number of allocations per bucket
  // LOCK: access to this stat is protected by the per bucket free_list_[index].mutex_
  std::vector<Stat> allocation_bucket_stats = std::vector<Stat>(MAX_SIZE_INDEX);
-  // SUM: bytes of allocation per bucket (active + free)
+  // SUM: bytes of allocation per bucket
  // LOCK: access to this stat is protected by the per bucket free_list_[index].mutex_
  std::vector<Stat> allocated_bytes_bucket_stats = std::vector<Stat>(MAX_SIZE_INDEX);
  // SUM: time spent in cudaHostAlloc/cudaHostRegister
@ -258,6 +211,12 @@ struct CachingHostAllocatorImpl {
    // Check in the recently freed blocks with pending events to see if we
    // can reuse them. Call get_free_block again after processing events
    if (pinned_use_background_threads()) {
+      process_events_for_specific_size(roundSize);
+      block = get_free_block(roundSize);
+      if (block) {
+        return {block->ptr_, reinterpret_cast<void*>(block)};
+      }
+
      // Launch the background thread and process events in a loop.
      static bool background_thread_flag [[maybe_unused]] = [this] {
        getBackgroundThreadPool()->run([&]() {
@ -319,6 +278,8 @@ struct CachingHostAllocatorImpl {
      auto index = size_index(block->size_);
      std::lock_guard<std::mutex> g(free_list_[index].mutex_);
      free_list_[index].list_.push_back(block);
+      stats_.allocation_bucket_stats[index].decrease(1);
+      stats_.allocated_bytes_bucket_stats[index].decrease(block->size_);
    } else {
      // restore these events that record by used streams.
      std::lock_guard<std::mutex> g(events_mutex_);
@ -378,12 +339,9 @@ struct CachingHostAllocatorImpl {
      for (auto* block : blocks_to_remove) {
        blocks_.erase(block);
        ptr_to_block_.erase(block->ptr_);
-        auto index = size_index(block->size_);
-        free_block(block);
-        stats_.allocations.decrease(1);
+        stats_.allocation.decrease(1);
        stats_.allocated_bytes.decrease(block->size_);
-        stats_.allocation_bucket_stats[index].decrease(1);
-        stats_.allocated_bytes_bucket_stats[index].decrease(block->size_);
+        free_block(block);
        delete block;
      }
    }
@ -430,17 +388,16 @@ struct CachingHostAllocatorImpl {
      // per bucket (we pick index 0 arbitrarily). These are also all the host
      // allocations, not taking into account caching and free lists.
      if (i == 0) {
-        stats.allocations = stats_.allocations;
-        stats.allocated_bytes = stats_.allocated_bytes;
-        stats.num_host_alloc = stats.allocations.allocated;
-        stats.num_host_free = stats.allocations.freed;
+        stats.segment = stats_.allocation;
+        stats.reserved_bytes = stats_.allocated_bytes;
+        stats.num_host_alloc = stats.segment.allocated;
+        stats.num_host_free = stats.segment.freed;
      }

      // Bucket stats need to be merged with the slow-path stats. We do this in
      // a best effort manner, since we can't really replay the cached events per bucket.
-      add_bucket_stats(stats.active_requests, stats_.active_bucket_stats[i]);
-      add_bucket_stats(stats.active_bytes, stats_.active_bytes_bucket_stats[i]);
-      stats.bucket_allocation[i] = stats_.allocation_bucket_stats[i].allocated;
+      add_bucket_stats(stats.allocation, stats_.allocation_bucket_stats[i]);
+      add_bucket_stats(stats.allocated_bytes, stats_.allocated_bytes_bucket_stats[i]);
    }

    // Get the timing stats
@ -464,11 +421,9 @@ struct CachingHostAllocatorImpl {
      std::lock_guard<std::mutex> gb(blocks_mutex_, std::adopt_lock);

      if (i == 0) {
-        stats_.allocations.reset_accumulated();
+        stats_.allocation.reset_accumulated();
        stats_.allocated_bytes.reset_accumulated();
      }
-      stats_.active_bucket_stats[i].reset_accumulated();
-      stats_.active_bytes_bucket_stats[i].reset_accumulated();
      stats_.allocation_bucket_stats[i].reset_accumulated();
      stats_.allocated_bytes_bucket_stats[i].reset_accumulated();
    }
@ -491,11 +446,9 @@ struct CachingHostAllocatorImpl {
      std::lock_guard<std::mutex> gb(blocks_mutex_, std::adopt_lock);

      if (i == 0) {
-        stats_.allocations.reset_peak();
+        stats_.allocation.reset_peak();
        stats_.allocated_bytes.reset_peak();
      }
-      stats_.active_bucket_stats[i].reset_peak();
-      stats_.active_bytes_bucket_stats[i].reset_peak();
      stats_.allocation_bucket_stats[i].reset_peak();
      stats_.allocated_bytes_bucket_stats[i].reset_peak();
    }
@ -512,7 +465,7 @@ struct CachingHostAllocatorImpl {
  virtual void add_allocated_block(B* block) {
    std::lock_guard<std::mutex> g(blocks_mutex_);
    blocks_.insert(block);
-    stats_.allocations.increase(1);
+    stats_.allocation.increase(1);
    stats_.allocated_bytes.increase(block->size_);
    ptr_to_block_.insert({block->ptr_, block});

@ -525,8 +478,6 @@ struct CachingHostAllocatorImpl {
      std::lock_guard<std::mutex> g(free_list_[index].mutex_);
      stats_.allocation_bucket_stats[index].increase(1);
      stats_.allocated_bytes_bucket_stats[index].increase(size);
-      stats_.active_bucket_stats[index].increase(1);
-      stats_.active_bytes_bucket_stats[index].increase(size);
    }
  }

@ -537,8 +488,8 @@ struct CachingHostAllocatorImpl {
      B* block = free_list_[index].list_.back();
      free_list_[index].list_.pop_back();
      block->allocated_ = true;
-      stats_.active_bucket_stats[index].increase(1);
-      stats_.active_bytes_bucket_stats[index].increase(size);
+      stats_.allocation_bucket_stats[index].increase(1);
+      stats_.allocated_bytes_bucket_stats[index].increase(size);
      return block;
    }
    return nullptr;
@ -632,8 +583,8 @@ struct CachingHostAllocatorImpl {
        auto index = size_index(block->size_);
        std::lock_guard<std::mutex> g(free_list_[index].mutex_);
        free_list_[index].list_.push_back(block);
-        stats_.active_bucket_stats[index].decrease(1);
-        stats_.active_bytes_bucket_stats[index].decrease(size);
+        stats_.allocation_bucket_stats[index].decrease(1);
+        stats_.allocated_bytes_bucket_stats[index].decrease(size);
        if (size != -1) {
          return;
        }
--- a/aten/src/ATen/core/PythonFallbackKernel.cpp
+++ b/aten/src/ATen/core/PythonFallbackKernel.cpp
@ -2,7 +2,6 @@
 #include <c10/core/impl/PythonDispatcherTLS.h>
 #include <ATen/core/PythonFallbackKernel.h>
 #include <c10/core/SafePyObject.h>
-#include <ATen/record_function.h>

 namespace {

@ -54,24 +53,20 @@ void pythonFallback(const c10::OperatorHandle& op, c10::DispatchKeySet dispatch_
  TORCH_INTERNAL_ASSERT(tls_on_entry.has_value());
  // c10::impl::ForceDispatchKeyGuard dispatcher_guard(tls_on_entry.value());
  // StashTLSOnEntryGuard stash_guard;
-  c10::impl::ExcludeDispatchKeyGuard exclude_guard(after_Python_keyset);
+  c10::impl::ExcludeDispatchKeyGuard guard(after_Python_keyset);

-  const auto& schema = op.schema();
-  const auto num_arguments = schema.arguments().size();

  // If Torch Dispatch Mode is active, use its PyInterpreter for dispatch
  const auto mode_stack_len = c10::impl::TorchDispatchModeTLS::stack_len();
  if (mode_stack_len > 0) {
-    RECORD_FUNCTION("PythonDispatchMode", torch::jit::last(*stack, num_arguments));
    const auto& cur_torch_dispatch_mode_state = c10::impl::TorchDispatchModeTLS::get_stack_at(mode_stack_len - 1);
    cur_torch_dispatch_mode_state->pyinterpreter()->dispatch(op, stack);
    return;
  }

-  RECORD_FUNCTION("PythonSubclass", torch::jit::last(*stack, num_arguments));
-
  // Otherwise, find a PyInterpreter on a Tensor
-
+  const auto& schema = op.schema();
+  const auto num_arguments = schema.arguments().size();
  // It is safe to dispatch on the very first Tensor with a pyobj_interpreter
  // without checking the interpreters of any of the arguments, because when
  // we actually run dispatch(), we will take out PyObjects in the context
--- a/aten/src/ATen/core/TransformationHelper.h
+++ b/aten/src/ATen/core/TransformationHelper.h
@ -117,7 +117,7 @@ C10_HOST_DEVICE inline T cauchy(T val, T median, T sigma) {
 template <>
 C10_HOST_DEVICE inline double cauchy(double val, double median, double sigma) {
  // https://en.wikipedia.org/wiki/Cauchy_distribution#Cumulative_distribution_function
-  return median + sigma * at::tan(c10::pi<double> * (val - 0.5));
+  return median + sigma * at::tan(c10::pi<double> * (val - static_cast<double>(0.5)));
 }

 /**
--- a/aten/src/ATen/core/boxing/BoxedKernel_impl.h
+++ b/aten/src/ATen/core/boxing/BoxedKernel_impl.h
@ -2,7 +2,7 @@

 namespace c10 {

-inline BoxedKernel::BoxedKernel() : boxed_kernel_func_(nullptr) {}
+inline BoxedKernel::BoxedKernel() : functor_(), boxed_kernel_func_(nullptr) {}

 inline BoxedKernel::BoxedKernel(
    std::unique_ptr<OperatorKernel> functor,
--- a/aten/src/ATen/core/boxing/KernelFunction_impl.h
+++ b/aten/src/ATen/core/boxing/KernelFunction_impl.h
@ -20,7 +20,9 @@ make_unique_base(Args&&... args) {
 } // namespace detail

 inline KernelFunction::KernelFunction()
-    : unboxed_kernel_func_(nullptr), sym_unboxed_kernel_func_(nullptr) {}
+    : boxed_kernel_func_(),
+      unboxed_kernel_func_(nullptr),
+      sym_unboxed_kernel_func_(nullptr) {}

 inline KernelFunction::~KernelFunction() {
  if (tokens_) {
--- a/aten/src/ATen/core/dispatch/Dispatcher.cpp
+++ b/aten/src/ATen/core/dispatch/Dispatcher.cpp
@ -76,7 +76,13 @@ void _print_dispatch_trace(const std::string& label, const std::string& op_name,

 OpRegistrationListener::~OpRegistrationListener()= default;

-Dispatcher::Dispatcher(): backendFallbackKernels_(), listeners_(std::make_unique<detail::RegistrationListenerList>()), guard_(std::make_shared<Guard>())
+Dispatcher::Dispatcher()
+: operators_()
+, operatorLookupTable_()
+, backendFallbackKernels_()
+, listeners_(std::make_unique<detail::RegistrationListenerList>())
+, cond_var_()
+, guard_(std::make_shared<Guard>())
 {}

 Dispatcher::~Dispatcher() {
--- a/aten/src/ATen/core/dispatch/Dispatcher.h
+++ b/aten/src/ATen/core/dispatch/Dispatcher.h
@ -96,7 +96,7 @@ class TORCH_API Dispatcher final {
  friend class TypedOperatorHandle;

  struct Guard final {
-    Guard() : alive(true) {}
+    Guard() : alive(true), mutex() {}
    std::atomic<bool> alive;
    std::mutex mutex;
  };
--- a/aten/src/ATen/core/dispatch/OperatorEntry.cpp
+++ b/aten/src/ATen/core/dispatch/OperatorEntry.cpp
@ -62,7 +62,17 @@ static const auto& getDispatchTableIndexToKey() {
 }

 OperatorEntry::OperatorEntry(OperatorName&& operator_name)
-: name_(std::move(operator_name)), dispatchTable_(), dispatchKeyExtractor_(DispatchKeyExtractor::makeUninitialized()), is_observed_(ObservedOperators::isObserved(name_))
+: name_(std::move(operator_name))
+, schema_()
+#ifndef C10_MOBILE
+, tags_()
+#endif
+, dispatchTable_()
+, dispatchKeyExtractor_(DispatchKeyExtractor::makeUninitialized())
+, kernels_()
+, cpp_signature_()
+, sym_cpp_signature_()
+, is_observed_(ObservedOperators::isObserved(name_))
 {
  // Pick up any backend fallbacks that were registered prior to this
  // OperatorEntry being created.
--- a/aten/src/ATen/core/ivalue.cpp
+++ b/aten/src/ATen/core/ivalue.cpp
@ -357,7 +357,7 @@ IValue IValue::equals(const IValue& rhs) const {
    case Tag::Enum:
      return lhs.toEnumHolder()->is(*rhs.toEnumHolder());
    case Tag::Uninitialized:
-      // Uninitialized ivalues show up in no-ops when the compiler can prove a
+      // Unitialized ivalues show up in no-ops when the compiler can prove a
      // value will never be used. Just return false on any equality comparison.
      return false;
  }
--- a/aten/src/ATen/core/op_registration/op_allowlist.h
+++ b/aten/src/ATen/core/op_registration/op_allowlist.h
@ -114,7 +114,7 @@ constexpr bool allowlist_contains(std::string_view allowlist, std::string_view i
        }
        next++;
      } else {
-        if (allowlist.substr(cur) == item) {
+        if (allowlist.substr(cur).compare(item) == 0) {
          return true;
        }
        break;
--- a/aten/src/ATen/core/op_registration/op_registration.cpp
+++ b/aten/src/ATen/core/op_registration/op_registration.cpp
@ -73,7 +73,7 @@ c10::FunctionSchema RegisterOperators::inferSchemaFromKernels_(

  std::optional<FunctionSchema> inferred_schema = std::nullopt;
  for (const auto& kernel : options.kernels) {
-    if (nullptr != kernel.inferred_function_schema) {
+    if (nullptr != kernel.inferred_function_schema.get()) {
      if (!inferred_schema.has_value()) {
        inferred_schema = *kernel.inferred_function_schema;
        break;
--- a/aten/src/ATen/core/op_registration/op_registration.h
+++ b/aten/src/ATen/core/op_registration/op_registration.h
@ -411,6 +411,7 @@ public:

    Options()
    : schemaOrName_(std::nullopt)
+    , kernels()
    , aliasAnalysisKind_(std::nullopt)
    {}

@ -419,6 +420,7 @@ public:
    struct KernelRegistrationConfig final {
      KernelRegistrationConfig()
        : dispatch_key(std::nullopt)
+        , func()
        , cpp_signature(std::nullopt)
        , inferred_function_schema(nullptr)
      {}
--- a/aten/src/ATen/cpu/vec/vec256/vec256_int.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_int.h
@ -905,7 +905,7 @@ class Vectorized8 : public Vectorizedi {
    // Because loadu(const void* ptr, T count) requires zero initialization for
    // upper 128 bits. However, by using _mm256_castsi128_si256, the upper 128
    // bits of the result are undefined.
-    // TODO<leslie> We can use _mm256_zextsi128_si256 in the future,
+    // TODO<leslie> We can use _mm256_zextsi128_si256 in the furture,
    // since gcc 9.3 doesn't support it now.
    __m128i input_128 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(ptr));
    return _mm256_castsi128_si256(input_128);
@ -1844,7 +1844,7 @@ Vectorized<int16_t> inline shift_256_16(
    c0 = _mm256_srav_epi32(a0, b0);
  c0 = _mm256_shuffle_epi8(c0, ctl_1_0);

-  // Perform shifting the same way for input array elements with
+  // Peform shifting the same way for input array elements with
  // idx%2==1.
  __m256i a1 = _mm256_and_si256(a, keep_1);
  __m256i b1 = _mm256_shuffle_epi8(b, ctl_1_0);
@ -2180,7 +2180,7 @@ Vectorized<T> inline shift_256_8(
    c0 = _mm256_srlv_epi32(a0, b0);
  c0 = _mm256_shuffle_epi8(c0, ctl_3_0);

-  // Perform shifting the same way for input array elements with
+  // Peform shifting the same way for input array elements with
  // idx%4==1.
  __m256i a1 = _mm256_shuffle_epi8(a, ctl_1_3);
  __m256i b1 = _mm256_shuffle_epi8(b, ctl_1_0);
@ -2193,7 +2193,7 @@ Vectorized<T> inline shift_256_8(
    c1 = _mm256_srlv_epi32(a1, b1);
  c1 = _mm256_shuffle_epi8(c1, ctl_3_1);

-  // Perform shifting the same way for input array elements with
+  // Peform shifting the same way for input array elements with
  // idx%4==2.
  __m256i a2 = _mm256_shuffle_epi8(a, ctl_2_3);
  __m256i b2 = _mm256_shuffle_epi8(b, ctl_2_0);
@ -2206,7 +2206,7 @@ Vectorized<T> inline shift_256_8(
    c2 = _mm256_srlv_epi32(a2, b2);
  c2 = _mm256_shuffle_epi8(c2, ctl_3_2);

-  // Perform shifting the same way for input array elements with
+  // Peform shifting the same way for input array elements with
  // idx%4==3.
  __m256i a3 = _mm256_and_si256(a, keep_3);
  __m256i b3 = _mm256_shuffle_epi8(b, ctl_3_0);
--- a/aten/src/ATen/cpu/vec/vec512/vec512_int.h
+++ b/aten/src/ATen/cpu/vec/vec512/vec512_int.h
@ -1088,7 +1088,7 @@ class Vectorized8 : public Vectorizedi {
    // Because loadu(const void* ptr, T count) requires zero initialization for
    // upper 384 bits. However, by using _mm512_castsi128_si512, the upper 384
    // bits of the result are undefined.
-    // TODO<leslie> We can use _mm512_zextsi128_si512 in the future,
+    // TODO<leslie> We can use _mm512_zextsi128_si512 in the furture,
    // since gcc 9.3 doesn't support it now.
    __m128i input_128 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(ptr));
    return _mm512_castsi128_si512(input_128);
@ -2022,7 +2022,7 @@ Vectorized<T> inline shift_512_8(
    c0 = _mm512_srlv_epi16(a0, b0);
  c0 = _mm512_shuffle_epi8(c0, ctl_1_0);

-  // Perform shifting the same way for input array elements with
+  // Peform shifting the same way for input array elements with
  // idx%2==1.
  __m512i a1 = _mm512_and_si512(a, keep_1);
  __m512i b1 = _mm512_shuffle_epi8(b, ctl_1_0);
--- a/aten/src/ATen/cuda/CUDABlas.cpp
+++ b/aten/src/ATen/cuda/CUDABlas.cpp
@ -191,10 +191,6 @@ uint32_t _getAlignment(uintptr_t address) {

 #ifdef USE_ROCM
 static c10::cuda::CUDAStream _getCarveoutStream(int32_t value) {
-  // 0 is default value, meaning full CUs i.e. no mask
-  if (value == 0) {
-    return at::cuda::getCurrentCUDAStream();
-  }
  static int32_t last_value = 0;
  static hipStream_t stream;
  if (last_value == 0) {
@ -213,15 +209,15 @@ static c10::cuda::CUDAStream _getCarveoutStream(int32_t value) {
  int32_t CUs = at::cuda::getCurrentDeviceProperties()->multiProcessorCount;
  // how many uint32_t do we need to cover all CUs, fill bitmask with 1
  uint32_t mask_size = static_cast<uint32_t>((CUs + 32 - 1) / 32);
-  std::vector<uint32_t> mask(mask_size, uint32_t{0x00000000});
+  std::vector<uint32_t> mask(mask_size, uint32_t{0xffffffff});
  // starting from lowest order bits, in 32-bit chunks
  // set bits to 0 based on how many CUs to carve out
  int32_t full_shifts = value / 32;
  int32_t remainder = value % 32;
  for (int32_t i = 0; i < full_shifts; i++) {
-    mask[i] = uint32_t{0xffffffff};
+    mask[i] = uint32_t{0x00000000};
  }
-  mask[full_shifts] = uint32_t{0xffffffff} << (32 - remainder);
+  mask[full_shifts] = uint32_t{0xffffffff} << remainder;

  // finally, create masked stream
  AT_CUDA_CHECK(hipExtStreamCreateWithCUMask(&stream, mask_size, &mask[0]));
@ -323,7 +319,7 @@ class CuBlasLtMatmulDescriptor : public CuBlasLtDescriptor<
    descriptor_.reset(raw_descriptor);
  }
  template <typename T>
-  void setAttribute(cublasLtMatmulDescAttributes_t attr, const T value) {
+  inline void setAttribute(cublasLtMatmulDescAttributes_t attr, const T value) {
    // NOLINTNEXTLINE(bugprone-sizeof-expression)
    TORCH_CUDABLAS_CHECK(::cublasLtMatmulDescSetAttribute(descriptor(), attr, &value, sizeof(value)));
  }
@ -345,7 +341,7 @@ class CuBlasLtMatrixLayout : public CuBlasLtDescriptor<
    descriptor_.reset(raw_descriptor);
  }
  template <typename T>
-  void setAttribute(cublasLtMatrixLayoutAttribute_t attr, const T value) {
+  inline void setAttribute(cublasLtMatrixLayoutAttribute_t attr, const T value) {
    TORCH_CUDABLAS_CHECK(::cublasLtMatrixLayoutSetAttribute(descriptor(), attr, &value, sizeof(T)));
  }
 };
@ -360,7 +356,7 @@ class CuBlasLtMatmulPreference : public CuBlasLtDescriptor<
    descriptor_.reset(raw_descriptor);
  }
  template <typename T>
-  void setAttribute(cublasLtMatmulPreferenceAttributes_t attr, const T value) {
+  inline void setAttribute(cublasLtMatmulPreferenceAttributes_t attr, const T value) {
    TORCH_CUDABLAS_CHECK(::cublasLtMatmulPreferenceSetAttribute(descriptor(), attr, &value, sizeof(T)));
  }
 };
@ -395,7 +391,7 @@ static inline bool bgemm_internal_cublaslt(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(D
    computeType = CUBLAS_COMPUTE_64F;
    scaleType = CUDA_R_64F;
  } else if constexpr (std::is_same_v<Dtype, float>) {
-    if (at::globalContext().float32Precision(at::Float32Backend::CUDA, at::Float32Op::MATMUL) == at::Float32Precision::TF32) {
+    if (at::globalContext().float32Precision("cuda", "matmul") == "tf32") {
      computeType = CUBLAS_COMPUTE_32F_FAST_TF32;
    }
  } else if constexpr (std::is_same_v<Dtype, c10::complex<double>>) {
@ -422,40 +418,25 @@ static inline bool bgemm_internal_cublaslt(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(D
    abType = CUDA_R_16F;
    cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16F;
 #ifndef USE_ROCM
-    auto fp16_reduction = at::globalContext().allowFP16ReductionCuBLAS();
-    if (fp16_reduction !=
-        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
-      uint32_t mask =
-          fp16_reduction ==
-                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
-              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
-                 CUBLASLT_REDUCTION_SCHEME_NONE)
-              : CUBLASLT_REDUCTION_SCHEME_NONE;
-      preference.setAttribute(
-          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
+    if (!at::globalContext().allowFP16ReductionCuBLAS()) {
+      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
+        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
    }
 #endif
  } else if constexpr (std::is_same_v<Dtype, at::BFloat16>) {
    abType = CUDA_R_16BF;
    cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16BF;
 #ifndef USE_ROCM
-    auto bf16_reduction = at::globalContext().allowBF16ReductionCuBLAS();
-    if (bf16_reduction !=
-        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
-      uint32_t mask =
-          bf16_reduction ==
-                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
-              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
-                 CUBLASLT_REDUCTION_SCHEME_NONE)
-              : CUBLASLT_REDUCTION_SCHEME_NONE;
-      preference.setAttribute(
-          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
+    if (!at::globalContext().allowBF16ReductionCuBLAS()) {
+      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
+        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
    }
 #endif
  } else {
    static_assert(false && sizeof(Dtype), "at::cuda::blas::bgemm_internal_cublaslt: not implemented");
  }

+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasLtHandle_t ltHandle = at::cuda::getCurrentCUDABlasLtHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -589,6 +570,8 @@ inline void bgemm_internal_cublas(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(Dtype, C_D

 template <>
 void bgemm_internal_cublas<double>(CUDABLAS_BGEMM_ARGTYPES(double)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -600,6 +583,8 @@ void bgemm_internal_cublas<double>(CUDABLAS_BGEMM_ARGTYPES(double)) {

 template <>
 void bgemm_internal_cublas<float>(CUDABLAS_BGEMM_ARGTYPES(float)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -611,6 +596,8 @@ void bgemm_internal_cublas<float>(CUDABLAS_BGEMM_ARGTYPES(float)) {

 template <>
 void bgemm_internal_cublas<c10::complex<double>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<double>)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -624,6 +611,8 @@ void bgemm_internal_cublas<c10::complex<double>>(CUDABLAS_BGEMM_ARGTYPES(c10::co

 template <>
 void bgemm_internal_cublas<c10::complex<float>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<float>)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -637,6 +626,8 @@ void bgemm_internal_cublas<c10::complex<float>>(CUDABLAS_BGEMM_ARGTYPES(c10::com

 template <typename C_Dtype>
 inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(at::Half, C_Dtype)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -708,6 +699,8 @@ inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYP

 template <typename C_Dtype>
 inline void bgemm_internal_cublas_bfloat16_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(at::BFloat16, C_Dtype)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  BGEMM_CHECK_ARGVALUES(at::BFloat16);
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
@ -1031,6 +1024,8 @@ inline void gemm_internal_cublas(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(Dtype, C_Dty

 template <>
 void gemm_internal_cublas<double>(CUDABLAS_GEMM_ARGTYPES(double)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -1042,6 +1037,8 @@ void gemm_internal_cublas<double>(CUDABLAS_GEMM_ARGTYPES(double)) {

 template <>
 void gemm_internal_cublas<float>(CUDABLAS_GEMM_ARGTYPES(float)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -1053,6 +1050,8 @@ void gemm_internal_cublas<float>(CUDABLAS_GEMM_ARGTYPES(float)) {

 template <>
 void gemm_internal_cublas<c10::complex<double>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<double>)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -1066,6 +1065,8 @@ void gemm_internal_cublas<c10::complex<double>>(CUDABLAS_GEMM_ARGTYPES(c10::comp

 template <>
 void gemm_internal_cublas<c10::complex<float>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<float>)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -1079,6 +1080,8 @@ void gemm_internal_cublas<c10::complex<float>>(CUDABLAS_GEMM_ARGTYPES(c10::compl

 template <typename C_Dtype>
 inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(at::Half, C_Dtype)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -1136,15 +1139,8 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
  }
  if (prop->major >= 5) {
    cublasMath_t cublas_flags = CUBLAS_DEFAULT_MATH;
-    auto fp16_reduction = at::globalContext().allowFP16ReductionCuBLAS();
-    TORCH_CHECK(fp16_reduction !=
-        at::CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK,
-          "torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction("
-          "..., allow_splitk=False) requires the cuBLASLt backend");
-    if (fp16_reduction !=
-        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
-      cublas_flags = static_cast<cublasMath_t>(
-          cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
+    if (!at::globalContext().allowFP16ReductionCuBLAS()) {
+      cublas_flags = static_cast<cublasMath_t>(cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
    }
    // Disallow fp16 reductions that could lead to unexpected overflow issues.
    TORCH_CUDABLAS_CHECK(cublasSetMathMode(handle, cublas_flags));
@ -1194,6 +1190,7 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(

 template <typename C_Dtype>
 inline void gemm_internal_cublas_bfloat16_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(at::BFloat16, C_Dtype)) {
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t opa = _cublasOpFromChar(transa);
  cublasOperation_t opb = _cublasOpFromChar(transb);
@ -1203,15 +1200,8 @@ inline void gemm_internal_cublas_bfloat16_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DT
  GEMM_CHECK_ARGVALUES(at::BFloat16);
 #ifndef USE_ROCM
  cublasMath_t cublas_flags = CUBLAS_DEFAULT_MATH;
-  auto bf16_reduction = at::globalContext().allowBF16ReductionCuBLAS();
-  TORCH_CHECK(bf16_reduction !=
-      at::CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK,
-        "torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction("
-        "..., allow_splitk=False) requires the cuBLASLt backend");
-  if (bf16_reduction !=
-      at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
-    cublas_flags = static_cast<cublasMath_t>(
-        cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
+  if (!at::globalContext().allowBF16ReductionCuBLAS()) {
+    cublas_flags = static_cast<cublasMath_t>(cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
  }
 #endif
 #if defined(USE_ROCM)
@ -1589,7 +1579,7 @@ bool gemm_and_bias(
    computeType = CUBLAS_COMPUTE_64F;
    scaleType = CUDA_R_64F;
  } else if constexpr (std::is_same_v<Dtype, float>) {
-    if (at::globalContext().float32Precision(at::Float32Backend::CUDA, at::Float32Op::MATMUL) == at::Float32Precision::TF32) {
+    if (at::globalContext().float32Precision("cuda", "matmul") == "tf32") {
      computeType = CUBLAS_COMPUTE_32F_FAST_TF32;
    }
  } else if constexpr (std::is_same_v<Dtype, at::Half>) {
@ -1607,34 +1597,18 @@ bool gemm_and_bias(
    abType = CUDA_R_16F;
    cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16F;
 #ifndef USE_ROCM
-    auto fp16_reduction = at::globalContext().allowFP16ReductionCuBLAS();
-    if (fp16_reduction !=
-        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
-      uint32_t mask =
-          fp16_reduction ==
-                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
-              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
-                 CUBLASLT_REDUCTION_SCHEME_NONE)
-              : CUBLASLT_REDUCTION_SCHEME_NONE;
-      preference.setAttribute(
-          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
+    if (!at::globalContext().allowFP16ReductionCuBLAS()) {
+      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
+        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
    }
 #endif
  } else if constexpr (std::is_same_v<Dtype, at::BFloat16>) {
    abType = CUDA_R_16BF;
    cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16BF;
 #ifndef USE_ROCM
-    auto bf16_reduction = at::globalContext().allowBF16ReductionCuBLAS();
-    if (bf16_reduction !=
-        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
-      uint32_t mask =
-          bf16_reduction ==
-                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
-              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
-                 CUBLASLT_REDUCTION_SCHEME_NONE)
-              : CUBLASLT_REDUCTION_SCHEME_NONE;
-      preference.setAttribute(
-          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
+    if (!at::globalContext().allowBF16ReductionCuBLAS()) {
+      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
+        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
    }
 #endif
  }
@ -2430,6 +2404,8 @@ void trsmBatched<c10::complex<double>>(

 template <>
 void gemv<c10::complex<double>>(CUDABLAS_GEMV_ARGTYPES(c10::complex<double>)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t op = _cublasOpFromChar(trans);
  _cublasAdjustLdLevel2(m, n, &lda);
@ -2445,6 +2421,8 @@ void gemv<c10::complex<float>>(CUDABLAS_GEMV_ARGTYPES(c10::complex<float>)) {
  // gemv is bw bound, and does not benefit from TF32. But the precision
  // loss still happens on TF32. So we disable it here.
  NoTF32Guard disable_tf32;
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t op = _cublasOpFromChar(trans);
  _cublasAdjustLdLevel2(m, n, &lda);
@ -2457,6 +2435,8 @@ void gemv<c10::complex<float>>(CUDABLAS_GEMV_ARGTYPES(c10::complex<float>)) {

 template <>
 void gemv<double>(CUDABLAS_GEMV_ARGTYPES(double)) {
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t op = _cublasOpFromChar(trans);
  _cublasAdjustLdLevel2(m, n, &lda);
@ -2470,6 +2450,8 @@ void gemv<float>(CUDABLAS_GEMV_ARGTYPES(float)) {
  // gemv is bw bound, and does not benefit from TF32. But the precision
  // loss still happens on TF32. So we disable it here.
  NoTF32Guard disable_tf32;
+  // See Note [Writing Nondeterministic Operations]
+  globalContext().alertCuBLASConfigNotDeterministic();
  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
  cublasOperation_t op = _cublasOpFromChar(trans);
  _cublasAdjustLdLevel2(m, n, &lda);
--- a/aten/src/ATen/cuda/CUDAGeneratorImpl.cpp
+++ b/aten/src/ATen/cuda/CUDAGeneratorImpl.cpp
@ -109,7 +109,7 @@ void CUDAGeneratorState::increase(uint64_t increment) {
        offset_intragraph_ % 4 == 0, "RNG offset must be a multiple of 4.");
    // Ensures the increment does not cause overflow.
    TORCH_INTERNAL_ASSERT(
-        offset_intragraph_ <= std::numeric_limits<uint64_t>::max() - increment,
+        offset_intragraph_ <= std::numeric_limits<uint32_t>::max() - increment,
        "Increment causes overflow in the offset value.");
    offset_intragraph_ += increment;
  } else {
@ -461,7 +461,7 @@ void CUDAGeneratorImpl::unregister_graph(cuda::CUDAGraph* graph) {
 */
 PhiloxCudaState CUDAGeneratorImpl::philox_cuda_state(uint64_t increment) {
  if (at::cuda::currentStreamCaptureStatus() != at::cuda::CaptureStatus::None) {
-    uint64_t offset = state_->offset_intragraph_;
+    uint32_t offset = state_->offset_intragraph_;
    state_->increase(increment);
    return PhiloxCudaState(
        state_->seed_extragraph_.data_ptr<int64_t>(),
--- a/aten/src/ATen/cuda/CUDAGeneratorImpl.h
+++ b/aten/src/ATen/cuda/CUDAGeneratorImpl.h
@ -96,16 +96,16 @@ struct CUDAGraph;
 struct CUDAGeneratorState : public c10::intrusive_ptr_target {
  uint64_t seed_;
  uint64_t philox_offset_per_thread_;
-  uint64_t offset_intragraph_;
+  uint32_t offset_intragraph_;
  bool capturing_{};
  std::unordered_set<cuda::CUDAGraph*> registered_graphs_;
-  at::TensorBase seed_extragraph_;
-  at::TensorBase offset_extragraph_;
+  at::TensorBase seed_extragraph_{};
+  at::TensorBase offset_extragraph_{};

  CUDAGeneratorState(
      uint64_t seed = default_rng_seed_val,
      uint64_t philox_offset_per_thread = 0,
-      uint64_t offset_intragraph = 0)
+      uint32_t offset_intragraph = 0)
      : seed_(seed),
        philox_offset_per_thread_(philox_offset_per_thread),
        offset_intragraph_(offset_intragraph) {}
@ -167,7 +167,7 @@ struct TORCH_CUDA_CPP_API CUDAGeneratorImpl : public c10::GeneratorImpl {
  CUDAGeneratorImpl* clone_impl() const override;

  c10::intrusive_ptr<CUDAGeneratorState> state_;
-  std::atomic_flag no_reset_rnn_state_;
+  std::atomic_flag no_reset_rnn_state_{};
 };

 namespace cuda::detail {
--- a/aten/src/ATen/cuda/CUDAGraph.h
+++ b/aten/src/ATen/cuda/CUDAGraph.h
@ -56,7 +56,7 @@ struct TORCH_CUDA_CPP_API CUDAGraph {

  // the ID assigned by cuda during graph capture,
  // used to identify when a stream is participating in capture
-  CaptureId_t capture_id_ = 0;
+  CaptureId_t capture_id_ = -1;

  // uuid used to request a particular private mempool from CUDACachingAllocator.
  // By default, this will be set to {id_, 0}.
--- a/aten/src/ATen/cuda/CUDASparse.h
+++ b/aten/src/ATen/cuda/CUDASparse.h
@ -6,15 +6,43 @@
 #define HIPSPARSE_VERSION ((hipsparseVersionMajor*100000) + (hipsparseVersionMinor*100) + hipsparseVersionPatch)
 #endif

+// cuSparse Generic API added in CUDA 10.1
+// Windows support added in CUDA 11.0
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && ((CUSPARSE_VERSION >= 10300) || (CUSPARSE_VERSION >= 11000 && defined(_WIN32)))
+#define AT_USE_CUSPARSE_GENERIC_API() 1
+#else
+#define AT_USE_CUSPARSE_GENERIC_API() 0
+#endif
+
+// cuSparse Generic API descriptor pointers were changed to const in CUDA 12.0
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && \
+    (CUSPARSE_VERSION < 12000)
+#define AT_USE_CUSPARSE_NON_CONST_DESCRIPTORS() 1
+#else
+#define AT_USE_CUSPARSE_NON_CONST_DESCRIPTORS() 0
+#endif
+
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && \
+    (CUSPARSE_VERSION >= 12000)
+#define AT_USE_CUSPARSE_CONST_DESCRIPTORS() 1
+#else
+#define AT_USE_CUSPARSE_CONST_DESCRIPTORS() 0
+#endif

 #if defined(USE_ROCM)
 // hipSparse const API added in v2.4.0
 #if HIPSPARSE_VERSION >= 200400
+#define AT_USE_HIPSPARSE_CONST_DESCRIPTORS() 1
+#define AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS() 0
 #define AT_USE_HIPSPARSE_GENERIC_API() 1
 #else
+#define AT_USE_HIPSPARSE_CONST_DESCRIPTORS() 0
+#define AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS() 1
 #define AT_USE_HIPSPARSE_GENERIC_API() 1
 #endif
 #else // USE_ROCM
+#define AT_USE_HIPSPARSE_CONST_DESCRIPTORS() 0
+#define AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS() 0
 #define AT_USE_HIPSPARSE_GENERIC_API() 0
 #endif // USE_ROCM

--- a/aten/src/ATen/cuda/CUDASparseDescriptors.cpp
+++ b/aten/src/ATen/cuda/CUDASparseDescriptors.cpp
@ -12,6 +12,8 @@ cusparseStatus_t destroyConstDnMat(const cusparseDnMatDescr* dnMatDescr) {
  return cusparseDestroyDnMat(const_cast<cusparseDnMatDescr*>(dnMatDescr));
 }

+#if AT_USE_CUSPARSE_GENERIC_API() || AT_USE_HIPSPARSE_GENERIC_API()
+
 namespace {

 // If a specific GPU model does not provide native support for a given data
@ -208,4 +210,6 @@ CuSparseSpMatCsrDescriptor::CuSparseSpMatCsrDescriptor(const Tensor& input, int6
  descriptor_.reset(raw_descriptor);
 }

+#endif // AT_USE_CUSPARSE_GENERIC_API() || AT_USE_HIPSPARSE_GENERIC_API()
+
 } // namespace at::cuda::sparse
--- a/aten/src/ATen/cuda/CUDASparseDescriptors.h
+++ b/aten/src/ATen/cuda/CUDASparseDescriptors.h
@ -35,6 +35,7 @@ class CuSparseDescriptor {
  std::unique_ptr<T, CuSparseDescriptorDeleter<T, destructor>> descriptor_;
 };

+#if AT_USE_CUSPARSE_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_CONST_DESCRIPTORS()
 template <typename T, cusparseStatus_t (*destructor)(const T*)>
 struct ConstCuSparseDescriptorDeleter {
  void operator()(T* x) {
@ -57,6 +58,7 @@ class ConstCuSparseDescriptor {
 protected:
  std::unique_ptr<T, ConstCuSparseDescriptorDeleter<T, destructor>> descriptor_;
 };
+#endif // AT_USE_CUSPARSE_CONST_DESCRIPTORS || AT_USE_HIPSPARSE_CONST_DESCRIPTORS

 #if defined(USE_ROCM)
 using cusparseMatDescr = std::remove_pointer_t<hipsparseMatDescr_t>;
@ -121,8 +123,39 @@ class TORCH_CUDA_CPP_API CuSparseBsrsm2Info

 #endif // AT_USE_HIPSPARSE_TRIANGULAR_SOLVE

+#if AT_USE_CUSPARSE_GENERIC_API() || AT_USE_HIPSPARSE_GENERIC_API()
+
 cusparseIndexType_t getCuSparseIndexType(const c10::ScalarType& scalar_type);

+#if AT_USE_CUSPARSE_NON_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS()
+class TORCH_CUDA_CPP_API CuSparseDnMatDescriptor
+    : public CuSparseDescriptor<cusparseDnMatDescr, &cusparseDestroyDnMat> {
+ public:
+  explicit CuSparseDnMatDescriptor(const Tensor& input, int64_t batch_offset = -1);
+};
+
+class TORCH_CUDA_CPP_API CuSparseConstDnMatDescriptor
+    : public CuSparseDescriptor<const cusparseDnMatDescr, &destroyConstDnMat> {
+ public:
+  explicit CuSparseConstDnMatDescriptor(const Tensor& input, int64_t batch_offset = -1);
+  cusparseDnMatDescr* unsafe_mutable_descriptor() const {
+    return const_cast<cusparseDnMatDescr*>(descriptor());
+  }
+  cusparseDnMatDescr* unsafe_mutable_descriptor() {
+    return const_cast<cusparseDnMatDescr*>(descriptor());
+  }
+};
+
+class TORCH_CUDA_CPP_API CuSparseDnVecDescriptor
+    : public CuSparseDescriptor<cusparseDnVecDescr, &cusparseDestroyDnVec> {
+ public:
+  explicit CuSparseDnVecDescriptor(const Tensor& input);
+};
+
+class TORCH_CUDA_CPP_API CuSparseSpMatDescriptor
+    : public CuSparseDescriptor<cusparseSpMatDescr, &cusparseDestroySpMat> {};
+
+#elif AT_USE_CUSPARSE_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_CONST_DESCRIPTORS()
  class TORCH_CUDA_CPP_API CuSparseDnMatDescriptor
      : public ConstCuSparseDescriptor<
            cusparseDnMatDescr,
@ -161,6 +194,7 @@ cusparseIndexType_t getCuSparseIndexType(const c10::ScalarType& scalar_type);
      : public ConstCuSparseDescriptor<
            cusparseSpMatDescr,
            &cusparseDestroySpMat> {};
+#endif // AT_USE_CUSPARSE_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_CONST_DESCRIPTORS()

 class TORCH_CUDA_CPP_API CuSparseSpMatCsrDescriptor
    : public CuSparseSpMatDescriptor {
@ -249,4 +283,6 @@ class TORCH_CUDA_CPP_API CuSparseSpGEMMDescriptor
  }
 };

+#endif // AT_USE_CUSPARSE_GENERIC_API() || AT_USE_HIPSPARSE_GENERIC_API()
+
 } // namespace at::cuda::sparse
--- a/aten/src/ATen/cuda/CachingHostAllocator.cpp
+++ b/aten/src/ATen/cuda/CachingHostAllocator.cpp
@ -9,6 +9,7 @@

 #include <cuda_runtime_api.h>
 #include <future>
+#include <unordered_map>

 namespace at::cuda {
 namespace {
@ -71,20 +72,9 @@ using Block = HostBlock<CUDAStream>;
 struct CUDACachingHostAllocatorImpl
    : public CachingHostAllocatorImpl<CUDAStream, EventPool::Event> {
 private:
-  ska::flat_hash_map<void*, bool> use_host_register;
+  std::unordered_map<void*, bool> use_host_register;

  void allocate_host_memory(size_t size, void** ptr) override {
-    // try allocating from reserve segment first before calling into expensive APIs
-    if (get_reserve_segment().initialized()) {
-      *ptr = get_reserve_segment().allocate(size);
-      if (*ptr != nullptr) {
-        return;
-      }
-    }
-    allocate_host_memory_slowpath(size, ptr);
-  }
-
-  void allocate_host_memory_slowpath(size_t size, void** ptr) {
    // Pinned memory pointers allocated by any device can be directly used by
    // any other device, regardless of the current device at the time of
    // allocation, since we assume unified addressing. So we grab any existing
@ -123,18 +113,6 @@ struct CUDACachingHostAllocatorImpl
  }

  void free_block(Block* block) override {
-    // We never free blocks from the reserve segment
-    if (get_reserve_segment().initialized()) {
-      // Check if the block is from the reserve segment
-      if (get_reserve_segment().owns(block->ptr_)) {
-        return;
-      }
-    }
-
-    free_block_slowpath(block);
-  }
-
-  void free_block_slowpath(Block* block) {
    auto start = std::chrono::steady_clock::now();
    // Users may change the allocator config at will. torch unit tests do this.
    // However, allocations using cudaHostRegister should use corresonding
@ -173,6 +151,11 @@ struct CUDACachingHostAllocatorImpl
  }

  bool query_event(EventPool::Event& event) override {
+    // Do not call cudaEventQuery if capturing is underway
+    if (at::cuda::currentStreamCaptureStatusMayInitCtx() !=
+        at::cuda::CaptureStatus::None) {
+      return false;
+    }
    cudaError_t err = cudaEventQuery(*event);
    if (err == cudaErrorNotReady) {
      (void)cudaGetLastError(); // clear CUDA error
@ -194,20 +177,6 @@ struct CUDACachingHostAllocatorImpl
    return event_pool->get(idx);
  }

-  PinnedReserveSegment& get_reserve_segment() {
-    static auto reserve_segment = [&]() {
-      if (c10::cuda::CUDACachingAllocator::CUDAAllocatorConfig::pinned_reserve_segment_size_mb() > 0) {
-        void *ptr;
-        size_t sz = c10::cuda::CUDACachingAllocator::CUDAAllocatorConfig::pinned_reserve_segment_size_mb() * 1024 * 1024;
-        allocate_host_memory_slowpath(sz, &ptr);
-        return PinnedReserveSegment(ptr, sz);
-      } else {
-        return PinnedReserveSegment();
-      }
-    } ();
-    return reserve_segment;
-  }
-
  TaskThreadPool* getThreadPool() {
    static TaskThreadPool* pool = new TaskThreadPool(
        static_cast<int>(c10::cuda::CUDACachingAllocator::CUDAAllocatorConfig::
@ -222,15 +191,15 @@ struct CUDACachingHostAllocatorImpl
      size_t numThreads,
      size_t pageSize) {
    uintptr_t start = (uintptr_t)ptr + (size * i / numThreads);
-    uintptr_t end = start + (size / numThreads);
+    uintptr_t end = (uintptr_t)start + (size / numThreads);
    if (i == (numThreads - 1)) {
      end = (uintptr_t)ptr + size;
    }

    // pre-fault/map the pages by setting the first byte of the page
    uintptr_t alignedStart =
-        ((start + pageSize - 1) & ~(pageSize - 1));
-    for (uintptr_t p = alignedStart; p < (end); p += pageSize) {
+        (((uintptr_t)start + pageSize - 1) & ~(pageSize - 1));
+    for (uintptr_t p = alignedStart; p < ((uintptr_t)end); p += pageSize) {
      // NOLINTNEXTLINE(performance-no-int-to-ptr)
      memset((void*)p, 0, 1);
    }
--- a/aten/src/ATen/cuda/CublasHandlePool.cpp
+++ b/aten/src/ATen/cuda/CublasHandlePool.cpp
@ -310,7 +310,7 @@ cublasHandle_t getCurrentCUDABlasHandle() {
  // FP32 data type calculations based on the value of the allow_tf32 flag.
  // To enable TF32, set the math mode of the handle to CUBLAS_TF32_TENSOR_OP_MATH.
  if (!NoTF32Guard::should_disable_tf32() &&
-      at::globalContext().float32Precision(at::Float32Backend::CUDA, at::Float32Op::MATMUL) == at::Float32Precision::TF32) {
+      at::globalContext().float32Precision("cuda", "matmul") == "tf32") {
    TORCH_CUDABLAS_CHECK(cublasSetMathMode(handle, CUBLAS_TF32_TENSOR_OP_MATH));
  } else {
    TORCH_CUDABLAS_CHECK(cublasSetMathMode(handle, CUBLAS_DEFAULT_MATH));
--- a/aten/src/ATen/cuda/detail/DeviceThreadHandles.h
+++ b/aten/src/ATen/cuda/detail/DeviceThreadHandles.h
@ -122,7 +122,7 @@ struct DeviceThreadHandlePool : public std::enable_shared_from_this<DeviceThread

    // Called by the destructor.  Releases this thread's handles back into the pool.
    void release() {
-        if(!my_handles.empty()) {
+        if(my_handles.size() > 0) {
            auto parent = weak_parent.lock();
            if (!parent) {
                // If this thread exits after atexit handlers have completed, the
--- a/aten/src/ATen/cuda/detail/PhiloxCudaStateRaw.cuh
+++ b/aten/src/ATen/cuda/detail/PhiloxCudaStateRaw.cuh
@ -19,7 +19,7 @@ struct PhiloxCudaState {
  // Called if graph capture is underway
  PhiloxCudaState(int64_t* seed,
                  int64_t* offset_extragraph,
-                  uint64_t offset_intragraph) {
+                  uint32_t offset_intragraph) {
    seed_.ptr = seed;
    offset_.ptr = offset_extragraph;
    offset_intragraph_ = offset_intragraph;
@ -36,7 +36,7 @@ struct PhiloxCudaState {

  Payload seed_{};
  Payload offset_{};
-  uint64_t offset_intragraph_ = 0;
+  uint32_t offset_intragraph_ = 0;
  bool captured_ = false;
 };

--- a/aten/src/ATen/cuda/tunable/GemmCommon.h
+++ b/aten/src/ATen/cuda/tunable/GemmCommon.h
@ -162,7 +162,7 @@ inline std::string ComputeTypeFor() {
 // ROCBLAS and hipBLASLt.
 template <>
 inline std::string ComputeTypeFor<float>() {
-  if (at::globalContext().float32Precision(at::Float32Backend::CUDA, at::Float32Op::MATMUL) != at::Float32Precision::TF32) {
+  if (at::globalContext().float32Precision("cuda", "matmul") != "tf32") {
    return "f32_r";
  } else {
    return "xf32_r";
--- a/aten/src/ATen/cuda/tunable/GemmHipblaslt.h
+++ b/aten/src/ATen/cuda/tunable/GemmHipblaslt.h
@ -506,7 +506,7 @@ class HipblasltGemmOp : public Callable<ParamsT> {
      }

      hipblasComputeType_t computeType = HIPBLAS_COMPUTE_32F;
-      if (at::globalContext().float32Precision(at::Float32Backend::CUDA, at::Float32Op::MATMUL) == at::Float32Precision::TF32) {
+      if (at::globalContext().float32Precision("cuda", "matmul") == "tf32") {
        computeType = HIPBLAS_COMPUTE_32F_FAST_TF32;
      }
      HipBlasLtMatmulDescriptor matmul(computeType, HIP_R_32F);
--- a/aten/src/ATen/cuda/tunable/GemmRocblas.h
+++ b/aten/src/ATen/cuda/tunable/GemmRocblas.h
@ -141,7 +141,7 @@ class RocblasGemmOp : public Callable<GemmParams<T>> {

    TuningStatus Call(const GemmParams<T>* params) override {
      auto input_output_type = RocBlasDataTypeFor<T>();
-      if (at::globalContext().float32Precision(at::Float32Backend::CUDA, at::Float32Op::MATMUL) == at::Float32Precision::TF32 && input_output_type == rocblas_datatype_f32_r)
+      if (at::globalContext().float32Precision("cuda", "matmul") == "tf32" && input_output_type == rocblas_datatype_f32_r)
        return FAIL;  // no support for TF32 in rocBLAS
      auto compute_type = RocBlasComputeTypeFor<T>();
      auto h_a = DoCastForHalfOrBfloat16(params->alpha);
@ -209,7 +209,7 @@ class RocblasGemmStridedBatchedOp : public Callable<GemmStridedBatchedParams<T>>

    TuningStatus Call(const GemmStridedBatchedParams<T>* params) override {
      auto input_output_type = RocBlasDataTypeFor<T>();
-      if (at::globalContext().float32Precision(at::Float32Backend::CUDA, at::Float32Op::MATMUL) == at::Float32Precision::TF32 && input_output_type == rocblas_datatype_f32_r)
+      if (at::globalContext().float32Precision("cuda", "matmul") == "tf32" && input_output_type == rocblas_datatype_f32_r)
        return FAIL;  // no support for TF32 in rocBLAS
      auto compute_type = RocBlasComputeTypeFor<T>();
      auto h_a = DoCastForHalfOrBfloat16(params->alpha);
--- a/aten/src/ATen/cuda/tunable/Tunable.cpp
+++ b/aten/src/ATen/cuda/tunable/Tunable.cpp
@ -404,6 +404,8 @@ TuningContext::TuningContext() :
    max_warmup_iterations_{0},
    icache_flush_{true},
    rotating_buffer_size_{-1},
+    filename_{},
+    untuned_file_{},
    results_count_from_input_file_{0},
    is_shutting_down_{false}
 {
--- a/aten/src/ATen/cudnn/Descriptors.cpp
+++ b/aten/src/ATen/cudnn/Descriptors.cpp
@ -141,7 +141,7 @@ void FilterDescriptor::set(const at::Tensor &t, const at::MemoryFormat memory_fo
    size[i] = (int) t.size(i);
  }
  for (const auto i : c10::irange(dim, pad)) {
-    size[i] = 1;
+    size[i] = (int) 1;
  }
  dim = std::max(dim, pad);
  cudnnTensorFormat_t filter_format{};
--- a/aten/src/ATen/functorch/BatchRulesLinearAlgebra.cpp
+++ b/aten/src/ATen/functorch/BatchRulesLinearAlgebra.cpp
@ -176,7 +176,7 @@ struct LinalgCheckMatrixUnaryRuleHelper;

 template <char const *op_name, typename F, F Func, typename A, typename... T>
 struct LinalgCheckMatrixUnaryRuleHelper<op_name, F, Func, typelist<A, T...>> {
-  static Tensor check_and_reshape_input(const Tensor& tensor, std::optional<int64_t> batch_dim) {
+  static inline Tensor check_and_reshape_input(const Tensor& tensor, std::optional<int64_t> batch_dim) {
    TORCH_CHECK(rankWithoutBatchDim(tensor, batch_dim) >= 2, op_name, ": The input tensor A must have at least 2 dimensions.");
    return moveBatchDimToFront(tensor, batch_dim);
  }
@ -222,7 +222,7 @@ struct LinalgCheckMatrixBinaryRuleHelper;

 template <char const *op_name, typename F, F Func, typename A, typename B, typename... T>
 struct LinalgCheckMatrixBinaryRuleHelper<op_name, F, Func, typelist<A, B, T...>> {
-  static std::tuple<Tensor, Tensor> check_inputs_and_reshape_inputs(
+  static inline std::tuple<Tensor, Tensor> check_inputs_and_reshape_inputs(
      const Tensor& first, std::optional<int64_t> first_bdim,
      const Tensor& second, std::optional<int64_t> second_bdim) {
    TORCH_CHECK(rankWithoutBatchDim(first, first_bdim) >= 2,
--- a/aten/src/ATen/hip/impl/HIPAllocatorMasqueradingAsCUDA.h
+++ b/aten/src/ATen/hip/impl/HIPAllocatorMasqueradingAsCUDA.h
@ -90,10 +90,6 @@ public:
    allocator_->setMemoryFraction(fraction, device);
  }

-  std::vector<HIPCachingAllocator::StreamSegmentSize> getExpandableSegmentSizes(c10::DeviceIndex device) override {
-    return allocator_->getExpandableSegmentSizes(device);
-  }
-
  void enable(bool value) override {
    allocator_->enable(value);
  }
--- a/aten/src/ATen/native/Blas.cpp
+++ b/aten/src/ATen/native/Blas.cpp
@ -58,7 +58,7 @@ scalar_t dot_impl(int64_t n, const scalar_t *x, int64_t incx, const scalar_t *y,
 template<typename scalar_t>
 scalar_t vdot_impl(int64_t n, const scalar_t *x, int64_t incx, const scalar_t *y, int64_t incy);

-static constexpr bool lda_cond(int64_t m, int64_t n, int64_t lda) {
+static constexpr inline bool lda_cond(int64_t m, int64_t n, int64_t lda) {
  return n == 1 || lda >= std::max<int64_t>(1L, m);
 }

--- a/aten/src/ATen/native/CPUBlas.cpp
+++ b/aten/src/ATen/native/CPUBlas.cpp
@ -991,7 +991,7 @@ std::size_t UnsafeUkernelKeyHasher<PackKey>::operator()(const PackKey& key) cons
 template <typename key_t, typename value_t>
 struct KernelCache  {
  using kstore_t = std::unordered_map<key_t, std::shared_ptr<value_t>, UnsafeUkernelKeyHasher<key_t>>;
-  static std::shared_ptr<value_t>&& fetch_or_create(
+  static inline std::shared_ptr<value_t>&& fetch_or_create(
      const key_t& key,
      const std::function<std::shared_ptr<value_t>()>& callback) {
    auto&& search = get_store().find(key);
@ -1003,7 +1003,7 @@ struct KernelCache  {
    }
  }

-  static kstore_t& get_store() {
+  static inline kstore_t& get_store() {
    static thread_local kstore_t cache_kernels;
    return cache_kernels;
  }
@ -1067,7 +1067,7 @@ struct GemmHelper {
 struct Brgemm : public KernelCache <BrgemmKey, GemmHelper> {
  // Fetch/create GemmHelper object and execute brgemm with batch size = 1
  template <typename scalar_t_a, typename scalar_t_b, typename scalar_t_c>
-  static void call(
+  static inline void call(
      int64_t M,
      int64_t N,
      int64_t K,
@ -1118,12 +1118,12 @@ struct Brgemm : public KernelCache <BrgemmKey, GemmHelper> {
        .execute(A, B, (*value).A_B_offsets, C, (*value).scratchpad.data());
  }

-  static std::shared_ptr<GemmHelper>& get_current() {
+  static inline std::shared_ptr<GemmHelper>& get_current() {
    static thread_local std::shared_ptr<GemmHelper> current;
    return current;
  }

-  static bool device_check(ScalarType dtype) {
+  static inline bool device_check(ScalarType dtype) {
    if (!at::globalContext().userEnabledMkldnn()) {
      return false;
    }
@ -1153,7 +1153,7 @@ using pack_t = dnnl::ukernel::brgemm_pack_B;
 using pack_t = dnnl::ukernel::transform;
 #endif
 struct Pack : public KernelCache <PackKey, pack_t> {
-  static void call(
+  static inline void call(
      int64_t K,
      int64_t N,
      int64_t ld_in,
@ -1182,7 +1182,7 @@ struct Pack : public KernelCache <PackKey, pack_t> {
    }
  }

-  static bool could_pack(ScalarType dtype) {
+  static inline bool could_pack(ScalarType dtype) {
    if (!at::globalContext().userEnabledMkldnn()) {
      return false;
    }
--- a/aten/src/ATen/native/Convolution.cpp
+++ b/aten/src/ATen/native/Convolution.cpp
@ -702,7 +702,7 @@ static void check_shape_forward(const at::Tensor& input,
      // If kernel size is incorrect
      std::ostringstream input_ss;
      std::ostringstream kernel_ss;
-      std::string separator;
+      std::string separator = "";

      for (int i = 0, len = input_shape.size(); i < len; ++i) {
        input_ss << separator << input_shape[i];
@ -1019,7 +1019,7 @@ static Tensor convolution_same(

  if (symmetric_padding) {
    // All backends handle symmetric padding natively
-    SymDimVector output_padding(dim);
+    SymDimVector output_padding(static_cast<size_t>(dim));
    return at::convolution_symint(input, weight, bias, stride, padding_l, dilation,
                               false, output_padding, groups);
  }
@ -1039,7 +1039,7 @@ static Tensor convolution_same(
    }
  }
  auto padded_input = at::constant_pad_nd_symint(input, pad_nd, 0);
-  SymDimVector output_padding(dim);
+  SymDimVector output_padding(static_cast<size_t>(dim));
  return at::convolution_symint(padded_input, weight, bias, stride, padding_l,
                                dilation, false, output_padding, groups);
 }
@ -1174,7 +1174,7 @@ at::Tensor convolution(
  bool deterministic = ctx.deterministicCuDNN() || ctx.deterministicAlgorithms();
  return at::_convolution(input, weight, bias, stride, padding, dilation,
                          transposed, output_padding, groups,
-                          ctx.benchmarkCuDNN(), deterministic, ctx.userEnabledCuDNN(), ctx.allowTF32CuDNN(at::Float32Op::CONV));
+                          ctx.benchmarkCuDNN(), deterministic, ctx.userEnabledCuDNN(), ctx.allowTF32CuDNN("conv"));
 }

 at::Tensor convolution_overrideable(
@ -1319,7 +1319,7 @@ ConvBackend select_conv_backend(
  params.benchmark = ctx.benchmarkCuDNN();
  params.deterministic = ctx.deterministicCuDNN() || ctx.deterministicAlgorithms();
  params.cudnn_enabled = ctx.userEnabledCuDNN();
-  params.allow_tf32 = ctx.allowTF32CuDNN(at::Float32Op::CONV);
+  params.allow_tf32 = ctx.allowTF32CuDNN("conv");

  auto input = input_r;
  auto weight = weight_r;
@ -1699,7 +1699,7 @@ at::Tensor _convolution(
  c10::MaybeOwned<Tensor> bias_r_maybe_owned = at::borrow_from_optional_tensor(bias_r_opt);
  const Tensor& bias_r = *bias_r_maybe_owned;

-  return at::_convolution(input_r, weight_r, bias_r, stride_, padding_, dilation_, transposed_, output_padding_, groups_, benchmark, deterministic, cudnn_enabled, at::globalContext().allowTF32CuDNN(at::Float32Op::CONV));
+  return at::_convolution(input_r, weight_r, bias_r, stride_, padding_, dilation_, transposed_, output_padding_, groups_, benchmark, deterministic, cudnn_enabled, at::globalContext().allowTF32CuDNN("conv"));
 }

 std::tuple<Tensor, Tensor, Tensor> convolution_backward_overrideable(
@ -1997,7 +1997,7 @@ std::tuple<Tensor, Tensor, Tensor> convolution_backward(
  params.benchmark = ctx.benchmarkCuDNN();
  params.deterministic = ctx.deterministicCuDNN() || ctx.deterministicAlgorithms();
  params.cudnn_enabled = ctx.userEnabledCuDNN();
-  params.allow_tf32 = ctx.allowTF32CuDNN(at::Float32Op::CONV);
+  params.allow_tf32 = ctx.allowTF32CuDNN("conv");

  // Validate inputs.
  check_shape_backward(input, weight.sizes(), params);
--- a/aten/src/ATen/native/Copy.cpp
+++ b/aten/src/ATen/native/Copy.cpp
@ -1,5 +1,6 @@
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
 #include <ATen/native/Copy.h>
+#include <ATen/native/Copy.h>

 #include <ATen/core/Tensor.h>
 #include <ATen/Dispatch.h>
--- a/aten/src/ATen/native/LinearAlgebra.cpp
+++ b/aten/src/ATen/native/LinearAlgebra.cpp
@ -2801,7 +2801,6 @@ Tensor matrix_exp(const Tensor& a) {
 // TODO This should be deprecated in favor of linalg_matrix_exp_differential
 //      in FunctionsManual.cpp
 Tensor matrix_exp_backward(const Tensor& self, const Tensor& grad) {
-  squareCheckInputs(self, "matrix_exp_backward");
  NoTF32Guard disable_tf32;
  return backward_analytic_function_of_a_matrix(
    self, grad,
--- a/aten/src/ATen/native/PadNd.cpp
+++ b/aten/src/ATen/native/PadNd.cpp
@ -70,7 +70,7 @@ Tensor constant_pad_nd(const Tensor& self, IntArrayRef pad, const Scalar& value)
        new_shape.emplace_back(input_sizes[i]);
    }

-    for (const auto i : c10::irange(l_pad)) {
+    for (const auto i : c10::irange((size_t)l_pad)) {
        auto pad_idx = pad.size() - ((i + 1) * 2);
        auto new_dim = input_sizes[l_diff + i] + pad[pad_idx] + pad[pad_idx + 1];
        TORCH_CHECK(new_dim >= 0, "The input size ", input_sizes[l_diff + i], ", plus negative padding ",
--- a/aten/src/ATen/native/RangeUtils.h
+++ b/aten/src/ATen/native/RangeUtils.h
@ -47,7 +47,7 @@ int64_t compute_arange_size(const Scalar& start, const Scalar& end, const Scalar
    int64_t sgn = (xstep > 0) - (xstep < 0);
    size_d = std::ceil((xend - xstart + xstep - sgn) / xstep);
  } else {
-    size_d = std::ceil((end.to<double>() - start.to<double>())
+    size_d = std::ceil(static_cast<double>(end.to<double>() - start.to<double>())
                        / step.to<double>());
  }

--- a/Show More
+++ b/Show More