Leave ROCm alone for now

Signed-off-by: Huy Do <huydhn@gmail.com>
Install the correct torchao version
2025-10-23 14:59:34 +08:00 · 2025-09-11 21:20:56 -07:00 · 2025-09-11 19:45:43 -07:00 · 2025-09-10 23:01:07 -07:00 · 2025-09-11 05:57:13 +00:00 · 2025-09-11 05:52:46 +00:00
1216 changed files with 54464 additions and 32004 deletions
--- a/.ci/aarch64_linux/aarch64_ci_build.sh
+++ b/.ci/aarch64_linux/aarch64_ci_build.sh
@ -3,12 +3,13 @@ set -eux -o pipefail

 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}

-if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
+# Set CUDA architecture lists to match x86 build_cuda.sh
+if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
-fi
-
-if [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0"
+elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
 fi

 # Compress the fatbin with -compress-mode=size for CUDA 13
@ -27,14 +28,22 @@ cd /
 # on the mounted pytorch repo
 git config --global --add safe.directory /pytorch
 pip install -r /pytorch/requirements.txt
-pip install auditwheel==6.2.0
+pip install auditwheel==6.2.0 wheel
 if [ "$DESIRED_CUDA" = "cpu" ]; then
    echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
-    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
-    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
+    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
    echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
    export USE_SYSTEM_NCCL=1
-    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
-    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
+
+    # Check if we should use NVIDIA libs from PyPI (similar to x86 build_cuda.sh logic)
+    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
+        echo "Bundling CUDA libraries with wheel for aarch64."
+    else
+        echo "Using nvidia libs from pypi for aarch64."
+        echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
+        export USE_NVIDIA_PYPI_LIBS=1
+    fi
+
+    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi
--- a/.ci/aarch64_linux/aarch64_wheel_ci_build.py
+++ b/.ci/aarch64_linux/aarch64_wheel_ci_build.py
@ -13,49 +13,6 @@ 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()
@ -69,83 +26,186 @@ def replace_tag(filename) -> None:
        f.writelines(lines)


+def patch_library_rpath(
+    folder: str,
+    lib_name: str,
+    use_nvidia_pypi_libs: bool = False,
+    desired_cuda: str = "",
+) -> None:
+    """Apply patchelf to set RPATH for a library in torch/lib"""
+    lib_path = f"{folder}/tmp/torch/lib/{lib_name}"
+
+    if use_nvidia_pypi_libs:
+        # For PyPI NVIDIA libraries, construct CUDA RPATH
+        cuda_rpaths = [
+            "$ORIGIN/../../nvidia/cudnn/lib",
+            "$ORIGIN/../../nvidia/nvshmem/lib",
+            "$ORIGIN/../../nvidia/nccl/lib",
+            "$ORIGIN/../../nvidia/cusparselt/lib",
+        ]
+
+        if "130" in desired_cuda:
+            cuda_rpaths.append("$ORIGIN/../../nvidia/cu13/lib")
+        else:
+            cuda_rpaths.extend(
+                [
+                    "$ORIGIN/../../nvidia/cublas/lib",
+                    "$ORIGIN/../../nvidia/cuda_cupti/lib",
+                    "$ORIGIN/../../nvidia/cuda_nvrtc/lib",
+                    "$ORIGIN/../../nvidia/cuda_runtime/lib",
+                    "$ORIGIN/../../nvidia/cufft/lib",
+                    "$ORIGIN/../../nvidia/curand/lib",
+                    "$ORIGIN/../../nvidia/cusolver/lib",
+                    "$ORIGIN/../../nvidia/cusparse/lib",
+                    "$ORIGIN/../../nvidia/nvtx/lib",
+                    "$ORIGIN/../../nvidia/cufile/lib",
+                ]
+            )
+
+        # Add $ORIGIN for local torch libs
+        rpath = ":".join(cuda_rpaths) + ":$ORIGIN"
+    else:
+        # For bundled libraries, just use $ORIGIN
+        rpath = "$ORIGIN"
+
+    if os.path.exists(lib_path):
+        os.system(
+            f"cd {folder}/tmp/torch/lib/; "
+            f"patchelf --set-rpath '{rpath}' --force-rpath {lib_name}"
+        )
+
+
+def copy_and_patch_library(
+    src_path: str,
+    folder: str,
+    use_nvidia_pypi_libs: bool = False,
+    desired_cuda: str = "",
+) -> None:
+    """Copy a library to torch/lib and patch its RPATH"""
+    if os.path.exists(src_path):
+        lib_name = os.path.basename(src_path)
+        shutil.copy2(src_path, f"{folder}/tmp/torch/lib/{lib_name}")
+        patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
+
+
 def package_cuda_wheel(wheel_path, desired_cuda) -> None:
    """
    Package the cuda wheel libraries
    """
    folder = os.path.dirname(wheel_path)
-    wheelname = os.path.basename(wheel_path)
    os.mkdir(f"{folder}/tmp")
    os.system(f"unzip {wheel_path} -d {folder}/tmp")
-    # Common libraries for all CUDA versions
-    common_libs = [
-        # Non-NVIDIA system libraries
-        "/lib64/libgomp.so.1",
-        "/usr/lib64/libgfortran.so.5",
-        "/acl/build/libarm_compute.so",
-        "/acl/build/libarm_compute_graph.so",
-        # Common CUDA libraries (same for all versions)
-        "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_lapack_core.so.0",
-        "/usr/local/lib/libnvpl_blas_core.so.0",
-        "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
-        "/usr/local/cuda/lib64/libcudnn.so.9",
-        "/usr/local/cuda/lib64/libcusparseLt.so.0",
-        "/usr/local/cuda/lib64/libcurand.so.10",
-        "/usr/local/cuda/lib64/libnccl.so.2",
-        "/usr/local/cuda/lib64/libnvshmem_host.so.3",
-        "/usr/local/cuda/lib64/libcudnn_adv.so.9",
-        "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
-        "/usr/local/cuda/lib64/libcudnn_graph.so.9",
-        "/usr/local/cuda/lib64/libcudnn_ops.so.9",
-        "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
-        "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
-        "/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
-        "/usr/local/cuda/lib64/libcufile.so.0",
-        "/usr/local/cuda/lib64/libcufile_rdma.so.1",
-        "/usr/local/cuda/lib64/libcusparse.so.12",
-    ]
+    # Delete original wheel since it will be repackaged
+    os.system(f"rm {wheel_path}")

-    # CUDA version-specific libraries
-    if "130" in desired_cuda:
-        version_specific_libs = [
-            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
-            "/usr/local/cuda/lib64/libcublas.so.13",
-            "/usr/local/cuda/lib64/libcublasLt.so.13",
-            "/usr/local/cuda/lib64/libcudart.so.13",
-            "/usr/local/cuda/lib64/libcufft.so.12",
-            "/usr/local/cuda/lib64/libcusolver.so.12",
-            "/usr/local/cuda/lib64/libnvJitLink.so.13",
-            "/usr/local/cuda/lib64/libnvrtc.so.13",
-            "/usr/local/cuda/lib64/libnvrtc-builtins.so.13.0",
-        ]
-    elif "12" in desired_cuda:
-        # Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
-        minor_version = desired_cuda[-1]
-        version_specific_libs = [
-            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
-            "/usr/local/cuda/lib64/libcublas.so.12",
-            "/usr/local/cuda/lib64/libcublasLt.so.12",
-            "/usr/local/cuda/lib64/libcudart.so.12",
-            "/usr/local/cuda/lib64/libcufft.so.11",
-            "/usr/local/cuda/lib64/libcusolver.so.11",
-            "/usr/local/cuda/lib64/libnvJitLink.so.12",
-            "/usr/local/cuda/lib64/libnvrtc.so.12",
-            f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
+    # Check if we should use PyPI NVIDIA libraries or bundle system libraries
+    use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
+
+    if use_nvidia_pypi_libs:
+        print("Using nvidia libs from pypi - skipping CUDA library bundling")
+        # For PyPI approach, we don't bundle CUDA libraries - they come from PyPI packages
+        # We only need to bundle non-NVIDIA libraries
+        minimal_libs_to_copy = [
+            "/lib64/libgomp.so.1",
+            "/usr/lib64/libgfortran.so.5",
+            "/acl/build/libarm_compute.so",
+            "/acl/build/libarm_compute_graph.so",
+            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_lapack_core.so.0",
+            "/usr/local/lib/libnvpl_blas_core.so.0",
        ]

-    # Combine all libraries
-    libs_to_copy = common_libs + version_specific_libs
+        # Copy minimal libraries to unzipped_folder/torch/lib
+        for lib_path in minimal_libs_to_copy:
+            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)

-    # Copy libraries to unzipped_folder/a/lib
-    for lib_path in libs_to_copy:
-        lib_name = os.path.basename(lib_path)
-        shutil.copy2(lib_path, f"{folder}/tmp/torch/lib/{lib_name}")
-        os.system(
-            f"cd {folder}/tmp/torch/lib/; "
-            f"patchelf --set-rpath '$ORIGIN' --force-rpath {folder}/tmp/torch/lib/{lib_name}"
-        )
+        # Patch torch libraries used for searching libraries
+        torch_libs_to_patch = [
+            "libtorch.so",
+            "libtorch_cpu.so",
+            "libtorch_cuda.so",
+            "libtorch_cuda_linalg.so",
+            "libtorch_global_deps.so",
+            "libtorch_python.so",
+            "libtorch_nvshmem.so",
+            "libc10.so",
+            "libc10_cuda.so",
+            "libcaffe2_nvrtc.so",
+            "libshm.so",
+        ]
+        for lib_name in torch_libs_to_patch:
+            patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
+    else:
+        print("Bundling CUDA libraries with wheel")
+        # Original logic for bundling system CUDA libraries
+        # Common libraries for all CUDA versions
+        common_libs = [
+            # Non-NVIDIA system libraries
+            "/lib64/libgomp.so.1",
+            "/usr/lib64/libgfortran.so.5",
+            "/acl/build/libarm_compute.so",
+            "/acl/build/libarm_compute_graph.so",
+            # Common CUDA libraries (same for all versions)
+            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_lapack_core.so.0",
+            "/usr/local/lib/libnvpl_blas_core.so.0",
+            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
+            "/usr/local/cuda/lib64/libcudnn.so.9",
+            "/usr/local/cuda/lib64/libcusparseLt.so.0",
+            "/usr/local/cuda/lib64/libcurand.so.10",
+            "/usr/local/cuda/lib64/libnccl.so.2",
+            "/usr/local/cuda/lib64/libnvshmem_host.so.3",
+            "/usr/local/cuda/lib64/libcudnn_adv.so.9",
+            "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
+            "/usr/local/cuda/lib64/libcudnn_graph.so.9",
+            "/usr/local/cuda/lib64/libcudnn_ops.so.9",
+            "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
+            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
+            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
+            "/usr/local/cuda/lib64/libcufile.so.0",
+            "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            "/usr/local/cuda/lib64/libcusparse.so.12",
+        ]
+
+        # CUDA version-specific libraries
+        if "13" in desired_cuda:
+            minor_version = desired_cuda[-1]
+            version_specific_libs = [
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
+                "/usr/local/cuda/lib64/libcublas.so.13",
+                "/usr/local/cuda/lib64/libcublasLt.so.13",
+                "/usr/local/cuda/lib64/libcudart.so.13",
+                "/usr/local/cuda/lib64/libcufft.so.12",
+                "/usr/local/cuda/lib64/libcusolver.so.12",
+                "/usr/local/cuda/lib64/libnvJitLink.so.13",
+                "/usr/local/cuda/lib64/libnvrtc.so.13",
+                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.{minor_version}",
+            ]
+        elif "12" in desired_cuda:
+            # Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
+            minor_version = desired_cuda[-1]
+            version_specific_libs = [
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
+                "/usr/local/cuda/lib64/libcublas.so.12",
+                "/usr/local/cuda/lib64/libcublasLt.so.12",
+                "/usr/local/cuda/lib64/libcudart.so.12",
+                "/usr/local/cuda/lib64/libcufft.so.11",
+                "/usr/local/cuda/lib64/libcusolver.so.11",
+                "/usr/local/cuda/lib64/libnvJitLink.so.12",
+                "/usr/local/cuda/lib64/libnvrtc.so.12",
+                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
+            ]
+        else:
+            raise ValueError(f"Unsupported CUDA version: {desired_cuda}.")
+
+        # Combine all libraries
+        libs_to_copy = common_libs + version_specific_libs
+
+        # Copy libraries to unzipped_folder/torch/lib
+        for lib_path in libs_to_copy:
+            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)

    # Make sure the wheel is tagged with manylinux_2_28
    for f in os.scandir(f"{folder}/tmp/"):
@ -153,14 +213,8 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
            replace_tag(f"{f.path}/WHEEL")
            break

-    os.mkdir(f"{folder}/cuda_wheel")
-    os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
-    shutil.move(
-        f"{folder}/cuda_wheel/{wheelname}",
-        f"{folder}/{wheelname}",
-        copy_function=shutil.copy2,
-    )
-    os.system(f"rm -rf {folder}/tmp/ {folder}/cuda_wheel/")
+    os.system(f"wheel pack {folder}/tmp/ -d {folder}")
+    os.system(f"rm -rf {folder}/tmp/")


 def complete_wheel(folder: str) -> str:
@ -183,14 +237,7 @@ def complete_wheel(folder: str) -> str:
            f"/{folder}/dist/{repaired_wheel_name}",
        )
    else:
-        repaired_wheel_name = wheel_name.replace(
-            "linux_aarch64", "manylinux_2_28_aarch64"
-        )
-        print(f"Renaming {wheel_name} wheel to {repaired_wheel_name}")
-        os.rename(
-            f"/{folder}/dist/{wheel_name}",
-            f"/{folder}/dist/{repaired_wheel_name}",
-        )
+        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]

    print(f"Copying {repaired_wheel_name} to artifacts")
    shutil.copy2(
@ -227,11 +274,21 @@ if __name__ == "__main__":
    ).decode()

    print("Building PyTorch wheel")
-    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
+    build_vars = ""
    # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
    if enable_cuda:
        build_vars += "MAX_JOBS=5 "

+        # Handle PyPI NVIDIA libraries vs bundled libraries
+        use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
+        if use_nvidia_pypi_libs:
+            print("Configuring build for PyPI NVIDIA libraries")
+            # Configure for dynamic linking (matching x86 logic)
+            build_vars += "ATEN_STATIC_CUDA=0 USE_CUDA_STATIC_LINK=0 USE_CUPTI_SO=1 "
+        else:
+            print("Configuring build for bundled NVIDIA libraries")
+            # Keep existing static linking approach - already configured above
+
    override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
    desired_cuda = os.getenv("DESIRED_CUDA")
    if override_package_version is not None:
@ -256,19 +313,13 @@ 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 "
-            "ACL_ROOT_DIR=/acl "
-            "LD_LIBRARY_PATH=/pytorch/build/lib:/acl/build:$LD_LIBRARY_PATH "
-            "ACL_INCLUDE_DIR=/acl/build "
-            "ACL_LIBRARY=/acl/build "
-        )
+        build_vars += "USE_MKLDNN=ON USE_MKLDNN_ACL=ON "
+        build_vars += "ACL_ROOT_DIR=/acl "
        if enable_cuda:
            build_vars += "BLAS=NVPL "
        else:
-            build_vars += "BLAS=OpenBLAS OpenBLAS_HOME=/OpenBLAS "
+            build_vars += "BLAS=OpenBLAS OpenBLAS_HOME=/opt/OpenBLAS "
    else:
        print("build pytorch without mkldnn backend")

--- a/.ci/aarch64_linux/build_aarch64_wheel.py
+++ b/.ci/aarch64_linux/build_aarch64_wheel.py
@ -299,40 +299,6 @@ 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(
@ -700,7 +666,6 @@ 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")
@ -723,6 +688,8 @@ 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:
@ -743,15 +710,18 @@ def start_build(
    if host.using_docker():
        build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
    if enable_mkldnn:
-        build_ArmComputeLibrary(host, git_clone_flags)
+        host.run_cmd("pytorch/.ci/docker/common/install_acl.sh")
        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 $HOME/pytorch && export ACL_ROOT_DIR=$HOME/ComputeLibrary && {build_vars} python3 setup.py bdist_wheel{build_opts}"
+            f"cd $HOME/pytorch && {build_vars} python3 setup.py bdist_wheel{build_opts}"
        )
        print("Repair the wheel")
        pytorch_wheel_name = host.list_dir("pytorch/dist")[0]
-        ld_library_path = "$HOME/acl/build:$HOME/pytorch/build/lib"
+        ld_library_path = "/acl/build:$HOME/pytorch/build/lib"
        host.run_cmd(
            f"export LD_LIBRARY_PATH={ld_library_path} && auditwheel repair $HOME/pytorch/dist/{pytorch_wheel_name}"
        )
@ -907,7 +877,7 @@ def terminate_instances(instance_type: str) -> None:
 def parse_arguments():
    from argparse import ArgumentParser

-    parser = ArgumentParser("Builid and test AARCH64 wheels using EC2")
+    parser = ArgumentParser("Build 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/build.sh
+++ b/.ci/docker/build.sh
@ -81,8 +81,8 @@ elif [[ "$image" == *riscv* ]]; then
  DOCKERFILE="ubuntu-cross-riscv/Dockerfile"
 fi

-_UCX_COMMIT=7bb2722ff2187a0cad557ae4a6afa090569f83fb
-_UCC_COMMIT=20eae37090a4ce1b32bcce6144ccad0b49943e0b
+_UCX_COMMIT=7836b165abdbe468a2f607e7254011c07d788152
+_UCC_COMMIT=430e241bf5d38cbc73fc7a6b89155397232e3f96
 if [[ "$image" == *rocm* ]]; then
  _UCX_COMMIT=cc312eaa4655c0cc5c2bcd796db938f90563bcf6
  _UCC_COMMIT=0c0fc21559835044ab107199e334f7157d6a0d3d
@ -114,6 +114,16 @@ case "$tag" in
    UCC_COMMIT=${_UCC_COMMIT}
    TRITON=yes
    ;;
+  pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
+    CUDA_VERSION=13.0.0
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=11
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    ;;
  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
    CUDA_VERSION=12.8.1
    ANACONDA_PYTHON_VERSION=3.10
@ -125,28 +135,6 @@ case "$tag" in
    TRITON=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=9
-    VISION=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    TRITON=yes
-    INDUCTOR_BENCHMARKS=yes
-    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.13
-    GCC_VERSION=9
-    VISION=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    TRITON=yes
-    INDUCTOR_BENCHMARKS=yes
-    ;;
  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm)
    CUDA_VERSION=12.8.1
    ANACONDA_PYTHON_VERSION=3.12
@ -225,7 +213,8 @@ case "$tag" in
    NINJA_VERSION=1.9.0
    TRITON=yes
    ;;
-  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
+  pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
+    # TODO (huydhn): Upgrade this to Python >= 3.10
    ANACONDA_PYTHON_VERSION=3.9
    GCC_VERSION=11
    VISION=yes
--- a/.ci/docker/centos-rocm/Dockerfile
+++ b/.ci/docker/centos-rocm/Dockerfile
@ -56,9 +56,13 @@ ENV INSTALLED_VISION ${VISION}

 # Install rocm
 ARG ROCM_VERSION
+RUN mkdir ci_commit_pins
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh
+RUN rm install_rocm.sh common_utils.sh
+RUN rm -r ci_commit_pins
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh
--- a/.ci/docker/ci_commit_pins/rocm-composable-kernel.txt
+++ b/.ci/docker/ci_commit_pins/rocm-composable-kernel.txt
@ -0,0 +1 @@
+7fe50dc3da2069d6645d9deb8c017a876472a977
--- a/.ci/docker/ci_commit_pins/torchbench.txt
+++ b/.ci/docker/ci_commit_pins/torchbench.txt
@ -1 +1 @@
-22bc29b4d503fc895ff73bc720ff396e9723465f
+74a23feff57432129df84d8099e622773cf77925
--- a/.ci/docker/ci_commit_pins/triton-xpu.txt
+++ b/.ci/docker/ci_commit_pins/triton-xpu.txt
@ -1 +1 @@
-d0e80f39c562c70986fc548fa6e5852ad86e16e7
+1b0418a9a454b2b93ab8d71f40e59d2297157fae
--- a/.ci/docker/ci_commit_pins/triton.txt
+++ b/.ci/docker/ci_commit_pins/triton.txt
@ -1 +1 @@
-f7888497a1eb9e98d4c07537f0d0bcfe180d1363
+fccfc522864cf8bc172abe0cd58ae5581e2d44b9
--- a/.ci/docker/common/install_acl.sh
+++ b/.ci/docker/common/install_acl.sh
@ -1,16 +1,27 @@
-set -euo pipefail
+#!/bin/bash
+# Script used only in CD pipeline

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

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

+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 $ACL_CHECKOUT_DIR
--- a/.ci/docker/common/install_cpython.sh
+++ b/.ci/docker/common/install_cpython.sh
@ -83,9 +83,9 @@ function build_cpython {
        py_suffix=${py_ver::-1}
        py_folder=$py_suffix
    fi
-    # Only b3 is available now
+    # Update to rc2 due to https://github.com/python/cpython/commit/c72699086fe4
    if [ "$py_suffix" == "3.14.0" ]; then
-        py_suffix="3.14.0b3"
+        py_suffix="3.14.0rc2"
    fi
    wget -q $PYTHON_DOWNLOAD_URL/$py_folder/Python-$py_suffix.tgz -O Python-$py_ver.tgz
    do_cpython_build $py_ver Python-$py_suffix
--- a/.ci/docker/common/install_cuda.sh
+++ b/.ci/docker/common/install_cuda.sh
@ -147,7 +147,7 @@ function install_128 {
 }

 function install_130 {
-  CUDNN_VERSION=9.12.0.46
+  CUDNN_VERSION=9.13.0.50
  echo "Installing CUDA 13.0 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
  # install CUDA 13.0 in the same container
  install_cuda 13.0.0 cuda_13.0.0_580.65.06_linux
--- a/.ci/docker/common/install_openblas.sh
+++ b/.ci/docker/common/install_openblas.sh
@ -3,8 +3,10 @@

 set -ex

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

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

-make -j8 ${OPENBLAS_BUILD_FLAGS} -C ${OPENBLAS_CHECKOUT_DIR}
-make -j8 ${OPENBLAS_BUILD_FLAGS} install -C ${OPENBLAS_CHECKOUT_DIR}
+make -j8 ${OPENBLAS_BUILD_FLAGS} -C $OPENBLAS_CHECKOUT_DIR
+sudo make install -C $OPENBLAS_CHECKOUT_DIR
+
+rm -rf $OPENBLAS_CHECKOUT_DIR
--- a/.ci/docker/common/install_rocm.sh
+++ b/.ci/docker/common/install_rocm.sh
@ -2,6 +2,11 @@

 set -ex

+# for pip_install function
+source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
+
+ROCM_COMPOSABLE_KERNEL_VERSION="$(cat $(dirname $0)/../ci_commit_pins/rocm-composable-kernel.txt)"
+
 ver() {
    printf "%3d%03d%03d%03d" $(echo "$1" | tr '.' ' ');
 }
@ -113,6 +118,8 @@ EOF
        rm -rf HIP clr
    fi

+    pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
+
    # Cleanup
    apt-get autoclean && apt-get clean
    rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
@ -176,6 +183,8 @@ install_centos() {
      sqlite3 $kdb "PRAGMA journal_mode=off; PRAGMA VACUUM;"
  done

+  pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
+
  # Cleanup
  yum clean all
  rm -rf /var/cache/yum
--- a/.ci/docker/common/install_ucc.sh
+++ b/.ci/docker/common/install_ucc.sh
@ -44,8 +44,12 @@ function install_ucc() {

  ./autogen.sh

-  # We only run distributed tests on Tesla M60 and A10G
-  NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_86,code=compute_86"
+  if [[ -n "$CUDA_VERSION"  && $CUDA_VERSION == 13* ]]; then
+    NVCC_GENCODE="-gencode=arch=compute_86,code=compute_86"
+  else
+    # We only run distributed tests on Tesla M60 and A10G
+    NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_86,code=compute_86"
+  fi

  if [[ -n "$ROCM_VERSION" ]]; then
    if [[ -n "$PYTORCH_ROCM_ARCH" ]]; then
--- a/.ci/docker/common/install_xpu.sh
+++ b/.ci/docker/common/install_xpu.sh
@ -65,10 +65,14 @@ function install_ubuntu() {

 function install_rhel() {
    . /etc/os-release
-
-    if [[ ! " 8.8 8.10 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
-        echo "RHEL version ${VERSION_ID} not supported"
-        exit
+    if [[ "${ID}" == "rhel" ]]; then
+        if [[ ! " 8.8 8.9 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
+            echo "RHEL version ${VERSION_ID} not supported"
+            exit
+        fi
+    elif [[ "${ID}" == "almalinux" ]]; then
+        # Workaround for almalinux8 which used by quay.io/pypa/manylinux_2_28_x86_64
+        VERSION_ID="8.8"
    fi

    dnf install -y 'dnf-command(config-manager)'
--- a/.ci/docker/libtorch/Dockerfile
+++ b/.ci/docker/libtorch/Dockerfile
@ -74,6 +74,14 @@ RUN bash ./install_cuda.sh 13.0
 RUN bash ./install_magma.sh 13.0
 RUN ln -sf /usr/local/cuda-13.0 /usr/local/cuda

+# Install libibverbs for libtorch and copy to CUDA directory
+RUN apt-get update -y && \
+    apt-get install -y libibverbs-dev librdmacm-dev && \
+    cp /usr/lib/x86_64-linux-gnu/libmlx5.so* /usr/local/cuda/lib64/ && \
+    cp /usr/lib/x86_64-linux-gnu/librdmacm.so* /usr/local/cuda/lib64/ && \
+    cp /usr/lib/x86_64-linux-gnu/libibverbs.so* /usr/local/cuda/lib64/ && \
+    cp /usr/lib/x86_64-linux-gnu/libnl* /usr/local/cuda/lib64/
+
 FROM cpu as rocm
 ARG ROCM_VERSION
 ARG PYTORCH_ROCM_ARCH
--- a/.ci/docker/manywheel/Dockerfile_2_28_aarch64
+++ b/.ci/docker/manywheel/Dockerfile_2_28_aarch64
@ -62,6 +62,13 @@ 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
@ -70,4 +77,5 @@ 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/
-ENV LD_LIBRARY_PATH=/opt/OpenBLAS/lib:$LD_LIBRARY_PATH
+COPY --from=arm_compute /acl /acl
+ENV LD_LIBRARY_PATH=/opt/OpenBLAS/lib:/acl/build/:$LD_LIBRARY_PATH
--- a/.ci/docker/manywheel/build.sh
+++ b/.ci/docker/manywheel/build.sh
@ -28,6 +28,7 @@ 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)
@ -41,7 +42,6 @@ case ${image} in
        GPU_IMAGE=arm64v8/almalinux:8
        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13 --build-arg NINJA_VERSION=1.12.1"
        MANY_LINUX_VERSION="2_28_aarch64"
-        OPENBLAS_VERSION="v0.3.30"
        ;;
    manylinuxcxx11-abi-builder:cpu-cxx11-abi)
        TARGET=final
@ -121,7 +121,8 @@ tmp_tag=$(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=${OPENBLAS_VERSION}" \
+    --build-arg "OPENBLAS_VERSION=${OPENBLAS_VERSION:-}" \
+    --build-arg "ACL_VERSION=${ACL_VERSION:-}" \
    --target "${TARGET}" \
    -t "${tmp_tag}" \
    $@ \
--- a/.ci/docker/requirements-docs.txt
+++ b/.ci/docker/requirements-docs.txt
@ -1,7 +1,7 @@
 sphinx==5.3.0
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 5.3.0
-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@722b7e6f9ca512fcc526ad07d62b3d28c50bb6cd#egg=pytorch_sphinx_theme2
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@1657ad2fc1acdc98aa719eebecbb0128a7c13ce4#egg=pytorch_sphinx_theme2

 # TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
 # but it doesn't seem to work and hangs around idly. The initial thought that it is probably
--- a/.ci/docker/triton_version.txt
+++ b/.ci/docker/triton_version.txt
@ -1 +1 @@
-3.4.0
+3.5.0
--- a/.ci/docker/triton_xpu_version.txt
+++ b/.ci/docker/triton_xpu_version.txt
@ -1 +1 @@
-3.4.0
+3.5.0
--- a/.ci/docker/ubuntu-rocm/Dockerfile
+++ b/.ci/docker/ubuntu-rocm/Dockerfile
@ -52,9 +52,13 @@ ENV INSTALLED_VISION ${VISION}

 # Install rocm
 ARG ROCM_VERSION
+RUN mkdir ci_commit_pins
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh
+RUN rm install_rocm.sh common_utils.sh
+RUN rm -r ci_commit_pins
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh
--- a/.ci/docker/ubuntu/Dockerfile
+++ b/.ci/docker/ubuntu/Dockerfile
@ -66,6 +66,7 @@ ENV NCCL_LIB_DIR="/usr/local/cuda/lib64/"
 # (optional) Install UCC
 ARG UCX_COMMIT
 ARG UCC_COMMIT
+ARG CUDA_VERSION
 ENV UCX_COMMIT $UCX_COMMIT
 ENV UCC_COMMIT $UCC_COMMIT
 ENV UCX_HOME /usr
--- a/.ci/lumen_cli/cli/lib/core/vllm/lib.py
+++ b/.ci/lumen_cli/cli/lib/core/vllm/lib.py
@ -76,7 +76,6 @@ def sample_vllm_test_library():
                ),
                "pytest -v -s entrypoints/llm/test_lazy_outlines.py",
                "pytest -v -s entrypoints/llm/test_generate.py ",
-                "pytest -v -s entrypoints/llm/test_generate_multiple_loras.py",
                "VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode",
            ],
        },
@ -97,14 +96,24 @@ def sample_vllm_test_library():
            "num_gpus": 4,
            "steps": [
                "pytest -v -s -x lora/test_chatglm3_tp.py",
-                "echo $VLLM_WORKER_MULTIPROC_METHOD",
                "pytest -v -s -x lora/test_llama_tp.py",
-                "pytest -v -s -x lora/test_multi_loras_with_tp.py",
+                "pytest -v -s -x lora/test_llm_with_multi_loras.py",
            ],
        },
-        "vllm_lora_280_failure_test": {
-            "title": "LoRA 280 failure test",
-            "id": "vllm_lora_280_failure_test",
+        "vllm_distributed_test_28_failure_test": {
+            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
+            "id": "vllm_distributed_test_28_failure_test",
+            "env_vars": {
+                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
+            },
+            "num_gpus": 4,
+            "steps": [
+                "pytest -v -s distributed/test_sequence_parallel.py",
+            ],
+        },
+        "vllm_lora_28_failure_test": {
+            "title": "LoRA pytorch 2.8 failure test",
+            "id": "vllm_lora_28_failure_test",
            "steps": ["pytest -v lora/test_quant_model.py"],
        },
        "vllm_multi_model_processor_test": {
@ -115,6 +124,15 @@ def sample_vllm_test_library():
                "pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py",
            ],
        },
+        "vllm_multi_model_test_28_failure_test": {
+            "title": "Multi-Model Test (Failed 2.8 release)",
+            "id": "vllm_multi_model_test_28_failure_test",
+            "package_install": ["git+https://github.com/TIGER-AI-Lab/Mantis.git"],
+            "steps": [
+                "pytest -v -s models/multimodal/generation/test_voxtral.py",
+                "pytest -v -s models/multimodal/pooling",
+            ],
+        },
        "vllm_pytorch_compilation_unit_tests": {
            "title": "PyTorch Compilation Unit Tests",
            "id": "vllm_pytorch_compilation_unit_tests",
@ -129,6 +147,28 @@ def sample_vllm_test_library():
                "pytest -v -s compile/test_decorator.py",
            ],
        },
+        "vllm_languagde_model_test_extended_generation_28_failure_test": {
+            "title": "Language Models Test (Extended Generation) 2.8 release failure",
+            "id": "vllm_languagde_model_test_extended_generation_28_failure_test",
+            "package_install": [
+                "--no-build-isolation",
+                "git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8",
+            ],
+            "steps": [
+                "pytest -v -s models/language/generation/test_mistral.py",
+            ],
+        },
+        "vllm_distributed_test_2_gpu_28_failure_test": {
+            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
+            "id": "vllm_distributed_test_2_gpu_28_failure_test",
+            "env_vars": {
+                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
+            },
+            "num_gpus": 4,
+            "steps": [
+                "pytest -v -s distributed/test_sequence_parallel.py",
+            ],
+        },
        # TODO(elainewy):need to add g6 with 4 gpus to run this test
        "vllm_lora_test": {
            "title": "LoRA Test %N",
--- a/.ci/lumen_cli/cli/lib/core/vllm/vllm_test.py
+++ b/.ci/lumen_cli/cli/lib/core/vllm/vllm_test.py
@ -104,20 +104,26 @@ class VllmTestRunner(BaseRunner):
        main function to run vllm test
        """
        self.prepare()
-        with working_directory(self.work_directory):
-            if self.test_type == TestInpuType.TEST_PLAN:
-                if self.num_shards > 1:
-                    run_test_plan(
-                        self.test_plan,
-                        "vllm",
-                        sample_vllm_test_library(),
-                        self.shard_id,
-                        self.num_shards,
-                    )
+        try:
+            with working_directory(self.work_directory):
+                if self.test_type == TestInpuType.TEST_PLAN:
+                    if self.num_shards > 1:
+                        run_test_plan(
+                            self.test_plan,
+                            "vllm",
+                            sample_vllm_test_library(),
+                            self.shard_id,
+                            self.num_shards,
+                        )
+                    else:
+                        run_test_plan(
+                            self.test_plan, "vllm", sample_vllm_test_library()
+                        )
                else:
-                    run_test_plan(self.test_plan, "vllm", sample_vllm_test_library())
-            else:
-                raise ValueError(f"Unknown test type {self.test_type}")
+                    raise ValueError(f"Unknown test type {self.test_type}")
+        finally:
+            # double check the torches are not overridden by other packages
+            check_versions()

    def _install_wheels(self, params: VllmTestParameters):
        logger.info("Running vllm test with inputs: %s", params)
--- a/.ci/manywheel/build_cuda.sh
+++ b/.ci/manywheel/build_cuda.sh
@ -124,6 +124,7 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
    fi
    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
        echo "Bundling with cudnn and cublas."
+
        DEPS_LIST+=(
            "/usr/local/cuda/lib64/libcudnn_adv.so.9"
            "/usr/local/cuda/lib64/libcudnn_cnn.so.9"
@ -133,16 +134,11 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9"
            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9"
            "/usr/local/cuda/lib64/libcudnn.so.9"
-            "/usr/local/cuda/lib64/libcublas.so.12"
-            "/usr/local/cuda/lib64/libcublasLt.so.12"
            "/usr/local/cuda/lib64/libcusparseLt.so.0"
-            "/usr/local/cuda/lib64/libcudart.so.12"
-            "/usr/local/cuda/lib64/libnvrtc.so.12"
            "/usr/local/cuda/lib64/libnvrtc-builtins.so"
            "/usr/local/cuda/lib64/libcufile.so.0"
            "/usr/local/cuda/lib64/libcufile_rdma.so.1"
            "/usr/local/cuda/lib64/libnvshmem_host.so.3"
-            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12"
            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
        )
        DEPS_SONAME+=(
@ -154,22 +150,56 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
            "libcudnn_engines_precompiled.so.9"
            "libcudnn_heuristic.so.9"
            "libcudnn.so.9"
-            "libcublas.so.12"
-            "libcublasLt.so.12"
            "libcusparseLt.so.0"
-            "libcudart.so.12"
-            "libnvrtc.so.12"
            "libnvrtc-builtins.so"
            "libnvshmem_host.so.3"
            "libcufile.so.0"
            "libcufile_rdma.so.1"
-            "libcupti.so.12"
            "libnvperf_host.so"
        )
        # Add libnvToolsExt only if CUDA version is not 12.9
-        if [[ $CUDA_VERSION != 12.9* ]]; then
-            DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
-            DEPS_SONAME+=("libnvToolsExt.so.1")
+        if [[ $CUDA_VERSION == 13* ]]; then
+            DEPS_LIST+=(
+                "/usr/local/cuda/lib64/libcublas.so.13"
+                "/usr/local/cuda/lib64/libcublasLt.so.13"
+                "/usr/local/cuda/lib64/libcudart.so.13"
+                "/usr/local/cuda/lib64/libnvrtc.so.13"
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13"
+                "/usr/local/cuda/lib64/libibverbs.so.1"
+                "/usr/local/cuda/lib64/librdmacm.so.1"
+                "/usr/local/cuda/lib64/libmlx5.so.1"
+                "/usr/local/cuda/lib64/libnl-3.so.200"
+                "/usr/local/cuda/lib64/libnl-route-3.so.200")
+            DEPS_SONAME+=(
+                "libcublas.so.13"
+                "libcublasLt.so.13"
+                "libcudart.so.13"
+                "libnvrtc.so.13"
+                "libcupti.so.13"
+                "libibverbs.so.1"
+                "librdmacm.so.1"
+                "libmlx5.so.1"
+                "libnl-3.so.200"
+                "libnl-route-3.so.200")
+            export USE_CUPTI_SO=1
+            export ATEN_STATIC_CUDA=0
+            export USE_CUDA_STATIC_LINK=0
+            export USE_CUFILE=0
+        else
+            DEPS_LIST+=(
+                "/usr/local/cuda/lib64/libnvToolsExt.so.1"
+                "/usr/local/cuda/lib64/libcublas.so.12"
+                "/usr/local/cuda/lib64/libcublasLt.so.12"
+                "/usr/local/cuda/lib64/libcudart.so.12"
+                "/usr/local/cuda/lib64/libnvrtc.so.12"
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
+            DEPS_SONAME+=(
+                "libnvToolsExt.so.1"
+                "libcublas.so.12"
+                "libcublasLt.so.12"
+                "libcudart.so.12"
+                "libnvrtc.so.12"
+                "libcupti.so.12")
        fi
    else
        echo "Using nvidia libs from pypi."
--- a/.ci/pytorch/build.sh
+++ b/.ci/pytorch/build.sh
@ -89,7 +89,7 @@ fi
 if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
  export USE_MKLDNN=1
  export USE_MKLDNN_ACL=1
-  export ACL_ROOT_DIR=/ComputeLibrary
+  export ACL_ROOT_DIR=/acl
 fi

 if [[ "$BUILD_ENVIRONMENT" == *riscv64* ]]; then
--- a/.ci/pytorch/macos-test.sh
+++ b/.ci/pytorch/macos-test.sh
@ -177,9 +177,6 @@ checkout_install_torchbench() {
  popd

  pip install -r .ci/docker/ci_commit_pins/huggingface-requirements.txt
-  # https://github.com/pytorch/pytorch/issues/160689 to remove torchao because
-  # its current version 0.12.0 doesn't work with transformers 4.54.0
-  pip uninstall -y torchao

  echo "Print all dependencies after TorchBench is installed"
  python -mpip freeze
@ -195,7 +192,7 @@ torchbench_setup_macos() {
  git checkout "$(cat ../.github/ci_commit_pins/vision.txt)"
  git submodule update --init --recursive
  python setup.py clean
-  python setup.py develop
+  python -m pip install -e . -v --no-build-isolation
  popd

  pushd torchaudio
@ -204,7 +201,7 @@ torchbench_setup_macos() {
  git submodule update --init --recursive
  python setup.py clean
  #TODO: Remove me, when figure out how to make TorchAudio find brew installed openmp
-  USE_OPENMP=0 python setup.py develop
+  USE_OPENMP=0 python -m pip install -e . -v --no-build-isolation
  popd

  checkout_install_torchbench
--- a/.ci/pytorch/test.sh
+++ b/.ci/pytorch/test.sh
@ -496,6 +496,14 @@ test_inductor_cpp_wrapper_shard() {
    -k 'take' \
    --shard "$1" "$NUM_TEST_SHARDS" \
    --verbose
+
+  if [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
+    python test/run_test.py \
+      --include inductor/test_mkldnn_pattern_matcher \
+      -k 'xpu' \
+      --shard "$1" "$NUM_TEST_SHARDS" \
+      --verbose
+  fi
 }

 # "Global" flags for inductor benchmarking controlled by TEST_CONFIG
@ -770,11 +778,6 @@ test_single_dynamo_benchmark() {
 }

 test_inductor_micro_benchmark() {
-  # torchao requires cuda 8.0 or above for bfloat16 support
-  if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;8.6"
-  fi
-  install_torchao
  TEST_REPORTS_DIR=$(pwd)/test/test-reports
  if [[ "${TEST_CONFIG}" == *cpu* ]]; then
    test_inductor_set_cpu_affinity
@ -1656,37 +1659,50 @@ elif [[ "${TEST_CONFIG}" == *operator_benchmark* ]]; then
    elif [[ "${TEST_CONFIG}" == *all* ]]; then
      TEST_MODE="all"
    fi
-
    test_operator_benchmark cpu ${TEST_MODE}
-
  fi
 elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
+  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
+    install_torchao
+  fi
  test_inductor_distributed
 elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
  test_inductor_halide
 elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
  test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
+  install_torchao
  test_inductor_micro_benchmark
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
  install_torchvision
+  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
+    install_torchao
+  fi
  id=$((SHARD_NUMBER-1))
  test_dynamo_benchmark huggingface "$id"
 elif [[ "${TEST_CONFIG}" == *timm* ]]; then
  install_torchvision
+  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
+    install_torchao
+  fi
  id=$((SHARD_NUMBER-1))
  test_dynamo_benchmark timm_models "$id"
 elif [[ "${TEST_CONFIG}" == cachebench ]]; then
  install_torchaudio
  install_torchvision
+  install_torchao
  PYTHONPATH=/torchbench test_cachebench
 elif [[ "${TEST_CONFIG}" == verify_cachebench ]]; then
  install_torchaudio
  install_torchvision
+  install_torchao
  PYTHONPATH=/torchbench test_verify_cachebench
 elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
  install_torchaudio
  install_torchvision
+  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
+    install_torchao
+  fi
  id=$((SHARD_NUMBER-1))
  # https://github.com/opencv/opencv-python/issues/885
  pip_install opencv-python==4.8.0.74
@ -1706,12 +1722,18 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
  fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
  install_torchvision
+  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
+    install_torchao
+  fi
  PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
  if [[ "$SHARD_NUMBER" -eq "1" ]]; then
    test_inductor_aoti
  fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
  install_torchvision
+  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
+    install_torchao
+  fi
  test_inductor_shard "${SHARD_NUMBER}"
  if [[ "${SHARD_NUMBER}" == 1 ]]; then
    if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then
--- a/.ci/pytorch/windows/internal/copy.bat
+++ b/.ci/pytorch/windows/internal/copy.bat
@ -1,12 +1,20 @@
-copy "%CUDA_PATH%\bin\cusparse*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\bin\cublas*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\bin\cudart*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\bin\curand*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\bin\cufft*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\bin\cusolver*64_*.dll*" pytorch\torch\lib
+
+if %CUDA_VERSION% geq 130 (
+    set "dll_path=bin\x64"
+) else (
+    set "dll_path=bin"
+)
+
+copy "%CUDA_PATH%\%dll_path%\cusparse*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\%dll_path%\cublas*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\%dll_path%\cudart*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\%dll_path%\curand*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\%dll_path%\cufft*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\%dll_path%\cusolver*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\%dll_path%\nvrtc*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\%dll_path%\nvJitLink_*.dll*"  pytorch\torch\lib

 copy "%CUDA_PATH%\bin\cudnn*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\bin\nvrtc*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\cupti64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\nvperf_host*.dll*" pytorch\torch\lib

@ -20,8 +28,3 @@ copy "%libuv_ROOT%\bin\uv.dll" pytorch\torch\lib
 if exist "C:\Windows\System32\zlibwapi.dll" (
    copy "C:\Windows\System32\zlibwapi.dll"  pytorch\torch\lib
 )
-
-::copy nvJitLink dll is requires for cuda 12+
-if exist "%CUDA_PATH%\bin\nvJitLink_*.dll*" (
-    copy "%CUDA_PATH%\bin\nvJitLink_*.dll*"  pytorch\torch\lib
-)
--- a/.ci/pytorch/windows/internal/driver_update.bat
+++ b/.ci/pytorch/windows/internal/driver_update.bat
@ -1,9 +1,9 @@
-set WIN_DRIVER_VN=528.89
-set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe" & REM @lint-ignore
-curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe
+set WIN_DRIVER_VN=580.88
+set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe" & REM @lint-ignore
+curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe
 if errorlevel 1 exit /b 1

-start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe -s -noreboot
+start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe -s -noreboot
 if errorlevel 1 exit /b 1

-del %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe || ver > NUL
+del %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe || ver > NUL
--- a/.ci/wheel/build_wheel.sh
+++ b/.ci/wheel/build_wheel.sh
@ -124,19 +124,15 @@ popd

 export TH_BINARY_BUILD=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
-export MACOSX_DEPLOYMENT_TARGET=10.15
+export MACOSX_DEPLOYMENT_TARGET=11.0
 export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}

-SETUPTOOLS_PINNED_VERSION="==70.1.0"
-PYYAML_PINNED_VERSION="==5.3"
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
 RENAME_WHEEL=true
 case $desired_python in
    3.14t)
        echo "Using 3.14 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
        NUMPY_PINNED_VERSION="==2.1.0"
        CONDA_ENV_CREATE_FLAGS="python-freethreading"
        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
@ -145,8 +141,6 @@ case $desired_python in
        ;;
    3.14)
        echo "Using 3.14t deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
        NUMPY_PINNED_VERSION="==2.1.0"
        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
        desired_python="3.14.0rc1"
@ -154,8 +148,6 @@ case $desired_python in
        ;;
    3.13t)
        echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
        NUMPY_PINNED_VERSION="==2.1.0"
        CONDA_ENV_CREATE_FLAGS="python-freethreading"
        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
@ -164,37 +156,23 @@ case $desired_python in
        ;;
    3.13)
        echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
        NUMPY_PINNED_VERSION="==2.1.0"
        ;;
    3.12)
        echo "Using 3.12 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
        NUMPY_PINNED_VERSION="==2.0.2"
        ;;
    3.11)
        echo "Using 3.11 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=5.3"
        NUMPY_PINNED_VERSION="==2.0.2"
        ;;
    3.10)
        echo "Using 3.10 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=5.3"
-        NUMPY_PINNED_VERSION="==2.0.2"
-        ;;
-    3.9)
-        echo "Using 3.9 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=5.3"
        NUMPY_PINNED_VERSION="==2.0.2"
        ;;
    *)
-        echo "Using default deps"
-        NUMPY_PINNED_VERSION="==1.11.3"
+        echo "Unsupported version $desired_python"
+        exit 1
        ;;
 esac

@ -204,8 +182,6 @@ conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_p
 source activate "$tmp_env_name"

 PINNED_PACKAGES=(
-    "setuptools${SETUPTOOLS_PINNED_VERSION}"
-    "pyyaml${PYYAML_PINNED_VERSION}"
    "numpy${NUMPY_PINNED_VERSION}"
 )
 retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements-build.txt"
@ -223,7 +199,7 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"

-python setup.py bdist_wheel -d "$whl_tmp_dir"
+python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name ${mac_version}

 echo "Finished setup.py bdist_wheel at $(date)"

--- a/.circleci/scripts/binary_populate_env.sh
+++ b/.circleci/scripts/binary_populate_env.sh
@ -75,8 +75,8 @@ TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_version.txt)
 # Here PYTORCH_EXTRA_INSTALL_REQUIREMENTS is already set for the all the wheel builds hence append TRITON_CONSTRAINT
 TRITON_CONSTRAINT="platform_system == 'Linux' and platform_machine == 'x86_64'"

-# CUDA 12.9 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]]; then
+# CUDA 12.9/13.0 builds have triton for Linux and Linux aarch64 binaries.
+if [[ "$DESIRED_CUDA" == "cu129" ]] || [[ "$DESIRED_CUDA" == "cu130" ]]; then
  TRITON_CONSTRAINT="platform_system == 'Linux'"
 fi

--- a/.github/actionlint.yaml
+++ b/.github/actionlint.yaml
@ -12,7 +12,9 @@ self-hosted-runner:
    - linux.9xlarge.ephemeral
    - am2.linux.9xlarge.ephemeral
    - linux.12xlarge
+    - linux.12xlarge.memory
    - linux.24xlarge
+    - linux.24xlarge.memory
    - linux.24xlarge.ephemeral
    - linux.24xlarge.amd
    - linux.arm64.2xlarge
--- a/.github/actions/build-external-packages/action.yml
+++ b/.github/actions/build-external-packages/action.yml
@ -4,6 +4,11 @@ name: Build External packages
 description: build external packages for PyTorch

 inputs:
+  cuda-version:
+    description: CUDA version to use
+    type: string
+    required: true
+    default: '12.8.1'
  cuda-arch-list:
    description: TORCH_CUDA_ARCH_LIST (e.g., "8.0;8.9;9.0")
    type: string
@ -44,11 +49,12 @@ runs:
      env:
        SCCACHE_BUCKET: ossci-compiler-cache-circleci-v2
        SCCACHE_REGION: us-east-1
+        CUDA_VERSION: ${{ inputs.cuda-version }}
        TORCH_CUDA_ARCH_LIST: ${{ inputs.cuda-arch-list }}
        BASE_IMAGE: ${{ inputs.docker-image }}
        BUILD_TARGETS: ${{ inputs.build-targets }}
-        PARENT_OUTPUT_DIR: ${{ inputs.output-dir}}
-
+        PARENT_OUTPUT_DIR: ${{ inputs.output-dir }}
+        TORCH_WHEELS_PATH: ${{ inputs.torch-wheel-dir }}
      shell: bash
      run: |
        set -euo pipefail
@ -69,7 +75,6 @@ runs:
          export OUTPUT_DIR
          echo "Building external package: $target in directory $OUTPUT_DIR"
          python3 -m cli.run build external "$target"
-
        done

        END_TIME=$(date +%s)
--- a/.github/actions/checkout-pytorch/action.yml
+++ b/.github/actions/checkout-pytorch/action.yml
@ -57,6 +57,21 @@ runs:
        submodules: ${{ inputs.submodules }}
        show-progress: false

+    - name: Clean submodules post checkout
+      id: clean-submodules
+      if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' }}
+      shell: bash
+      env:
+        NO_SUDO: ${{ inputs.no-sudo }}
+      run: |
+        cd "${GITHUB_WORKSPACE}"
+        # Clean stale submodule dirs
+        if [ -z "${NO_SUDO}" ]; then
+          sudo git submodule foreach --recursive git clean -ffdx
+        else
+          git submodule foreach --recursive git clean -ffdx
+        fi
+
    - name: Clean workspace (try again)
      if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' &&
        (steps.first-clean.outcome != 'success' || steps.first-checkout-attempt.outcome != 'success') }}
--- a/.github/ci_commit_pins/audio.txt
+++ b/.github/ci_commit_pins/audio.txt
@ -1 +1 @@
-10a5002c6195bd95e34df8fe28ff8a2d55a2a922
+fa5142928ee157aa65137c4ecff2fe9b1a9e0648
--- a/.github/ci_commit_pins/fbgemm_rocm.txt
+++ b/.github/ci_commit_pins/fbgemm_rocm.txt
@ -1 +1 @@
-7f1de94a4c2d14f59ad4ca84538c36084ea6b2c8
+08ae0af1395c8d8471f4025deb6af9aef90b342f
--- a/.github/ci_commit_pins/torchao.txt
+++ b/.github/ci_commit_pins/torchao.txt
@ -1 +1 @@
-51c87b6ead6b7e098ada95d6a7609ee873b854cf
+f32431e593d0e9db86c502d3872dd67ee40a005f
--- a/.github/ci_commit_pins/vllm.txt
+++ b/.github/ci_commit_pins/vllm.txt
@ -1 +1 @@
-321938e9ac4000e0cb37e328359a7fd3026bc672
+cc99baf14dacc2497d0c5ed84e076ef2c37f6a4d
--- a/.github/ci_commit_pins/xla.txt
+++ b/.github/ci_commit_pins/xla.txt
@ -1 +1 @@
-a1c6ee92c85e8b0955c20892ed68f032a6015c09
+6c5478ff7c3d50dd1e3047d72ec5909bea474073
--- a/.github/ci_configs/vllm/Dockerfile.tmp_vllm
+++ b/.github/ci_configs/vllm/Dockerfile.tmp_vllm
@ -12,54 +12,46 @@ ARG BUILD_BASE_IMAGE=torch-nightly-base
 # by default, it uses devel-ubuntu22.04 official image.
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04

+# The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
+ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"

-#################### TORCH NIGHTLY  BASE IMAGE ####################
+
+#################### TORCH NIGHTLY BASE IMAGE ####################
 # A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
-From nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
-ARG CUDA_VERSION=12.8.1
-ARG PYTHON_VERSION=3.12
-ARG TARGETPLATFORM
-ENV DEBIAN_FRONTEND=noninteractive
+FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base

-RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
-    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
+ARG CUDA_VERSION
+ARG PYTHON_VERSION
+ARG GET_PIP_URL

-# Install Python and other dependencies if it does not existed
-RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
-      echo "Installing Python ${PYTHON_VERSION}..." && \
-      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
-      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
-      apt-get update -y && \
-      apt-get install -y ccache software-properties-common git curl sudo && \
-      for i in 1 2 3; do \
-        add-apt-repository -y ppa:deadsnakes/ppa && break || \
-        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
-      done && \
-      apt-get update -y && \
-      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
-      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
-      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
-      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
-      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
-   else \
-      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
-   fi \
-   && python3 --version && python3 -m pip --version
+# Install Python and other dependencies
+RUN apt-get update -y \
+    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
+    && add-apt-repository -y ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
+    && python3 --version && python3 -m pip --version

 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
 # Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
 RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if [ "$current_gcc_version" -lt 10 ]; then \
-      echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-      apt-get update && \
-      apt-get install -y gcc-10 g++-10 && \
-      update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 && \
-      update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-    else \
-      echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-    fi && \
-    gcc --version && g++ --version
+    if command -v apt-get >/dev/null; then \
+        if [ "$current_gcc_version" -lt 10 ]; then \
+            echo "GCC version is $current_gcc_version, installing gcc-10..."; \
+            apt-get update \
+            && apt-get install -y gcc-10 g++-10 \
+            && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
+            && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
+        else \
+            echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
+        fi \
+    fi \
+    && gcc --version && g++ --version

 # install uv for faster pip installs
 RUN --mount=type=cache,target=/root/.cache/uv \
@ -79,6 +71,21 @@ ENV UV_LINK_MODE=copy
 FROM ${BUILD_BASE_IMAGE} AS base
 USER root

+ARG CUDA_VERSION
+ARG PYTHON_VERSION
+
+# TODO (huydhn): Only work with PyTorch manylinux builder
+ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
+
+# Install some system dependencies and double check python version
+RUN if command -v apt-get >/dev/null; then \
+        apt-get update -y \
+        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
+    else \
+        dnf install -y git curl wget sudo vim; \
+    fi \
+    && python3 --version && python3 -m pip --version
+
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
 # this won't be needed for future versions of this docker image
@ -118,17 +125,15 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
    if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
        echo "[INFO] Installing torch wheels to build vllm"; \
        torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
-        uv pip install --system "${torch_whl}[opt-einsum]"; \
-        uv pip install --system "${vision_whl}"; \
-        uv pip install --system "${audio_whl}"; \
+        vision_whl=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
+        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
+        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
    elif [ -n "$PINNED_TORCH_VERSION" ]; then \
        echo "[INFO] Installing pinned torch nightly version to build vllm: $PINNED_TORCH_VERSION"; \
-        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu128; \
+        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
    else \
        echo "[INFO] Installing torch nightly with latest one to build vllm"; \
-        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128; \
+        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
    fi

 # Install numba 0.61.2 for cuda environment
@ -137,12 +142,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \

 # Install common dependencies from vllm common.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-uv pip install --system -r requirements/common.txt
-
+    uv pip install --system -r requirements/common.txt

 # Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG exformer_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${exformer_cuda_arch_list}
+ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
+ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}

 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
@ -153,8 +157,8 @@ RUN pip freeze | grep -E 'ninja'

 # Build xformers with cuda and torch nightly/wheel
 # following official xformers guidance: https://github.com/facebookresearch/xformers#build
-# sha for https://github.com/facebookresearch/xformers/tree/v0.0.31
-ARG XFORMERS_COMMIT=eb0946a363464da96ea40afd1a7f72a907c25497
+# sha for https://github.com/facebookresearch/xformers/tree/v0.0.32.post2
+ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
 ENV CCACHE_DIR=/root/.cache/ccache

 RUN --mount=type=cache,target=/root/.cache/ccache \
@ -188,11 +192,6 @@ RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
 FROM base AS build
 ARG TARGETPLATFORM

-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
-
 COPY . .

 RUN python3 use_existing_torch.py
@ -251,9 +250,9 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
        python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
    fi

-RUN echo "[DEBUG] Listing  current directory:" && \
+RUN echo "[INFO] Listing current directory:" && \
    ls -al && \
-    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
+    echo "[INFO] Showing torch_build_versions.txt content:" && \
    cat torch_build_versions.txt

 #################### WHEEL BUILD IMAGE ####################
@ -263,42 +262,40 @@ RUN echo "[DEBUG] Listing  current directory:" && \
 # Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 USER root
+
+ARG CUDA_VERSION
+ARG PYTHON_VERSION
+ARG GET_PIP_URL
+
+# TODO (huydhn): Only work with PyTorch manylinux builder
+ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
+
 # prepare for environment starts
 WORKDIR /workspace

-RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
-    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
-
-# Install Python and other dependencies if it does not existed
-RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
-      echo "Installing Python ${PYTHON_VERSION}..." && \
-      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
-      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
-      apt-get update -y && \
-      apt-get install -y ccache software-properties-common git curl sudo && \
-      for i in 1 2 3; do \
-        add-apt-repository -y ppa:deadsnakes/ppa && break || \
-        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
-      done && \
-      apt-get update -y && \
-      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
-      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
-      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
-      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
-      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
-   else \
-      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
-   fi \
-   && python3 --version && python3 -m pip --version
-
+# Install Python and other dependencies
+RUN if command -v apt-get >/dev/null; then \
+        apt-get update -y \
+        && apt-get install -y ccache software-properties-common git curl wget sudo vim \
+        && add-apt-repository -y ppa:deadsnakes/ppa \
+        && apt-get update -y \
+        && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+        && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+        && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+        && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+        && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
+    else \
+        dnf install -y git curl wget sudo vim; \
+    fi \
+    && python3 --version && python3 -m pip --version

 # Get the torch versions, and whls used in previous stagtes for consistency
 COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
-RUN echo "[DEBUG] Listing current directory before torch install step:" && \
+RUN echo "[INFO] Listing current directory before torch install step:" && \
    ls -al && \
-    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
+    echo "[INFO] Showing torch_build_versions.txt content:" && \
    cat torch_build_versions.txt

 # Workaround for https://github.com/openai/triton/issues/2507 and
@ -307,7 +304,6 @@ RUN echo "[DEBUG] Listing current directory before torch install step:" && \
 # or future versions of triton.
 RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/

-
 # 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 \
@ -327,15 +323,13 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
    --mount=type=cache,target=/root/.cache/uv \
    if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
        torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
+        vision_whl=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
+        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
        echo "[INFO] Use wheels to build : '${torch_whl}' '${audio_whl}' '${vision_whl}'"; \
-        uv pip install --system "${torch_whl}[opt-einsum]"; \
-        uv pip install --system "${vision_whl}"; \
-        uv pip install --system "${audio_whl}"; \
+        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
    else \
        echo "[INFO] Installing torch versions from torch_build_versions.txt"; \
-        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu128; \
+        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
    fi

 # Install the vllm wheel from previous stage
@ -346,9 +340,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 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'
+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

@ -416,11 +409,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --system -r requirements/nightly_torch_test.txt

-# Workaround for #17068
-# pinned commit for v2.2.4
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@95d8aba8a8c75aedcaa6143713b11e745e7cd0d9#egg=mamba-ssm"
-
 # Logging to confirm the torch versions
 RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'

--- a/.github/scripts/build_triton_wheel.py
+++ b/.github/scripts/build_triton_wheel.py
@ -84,6 +84,7 @@ def build_triton(
                ["git", "checkout", f"release/{ver}.{rev}.x"], cwd=triton_basedir
            )
        else:
+            check_call(["git", "fetch", "origin", commit_hash], cwd=triton_basedir)
            check_call(["git", "checkout", commit_hash], cwd=triton_basedir)

        # change built wheel name and version
--- a/.github/scripts/generate_binary_build_matrix.py
+++ b/.github/scripts/generate_binary_build_matrix.py
@ -16,18 +16,16 @@ from typing import Optional


 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
    "12.6": "12.6.3",
    "12.8": "12.8.1",
-    "12.9": "12.9.1",
    "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
    "12.6": "9",
    "12.8": "9",
-    "12.9": "9",
    "13.0": "9",
 }

@ -40,77 +38,60 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]

 CPU_S390X_ARCH = ["cpu-s390x"]

-CUDA_AARCH64_ARCHES = ["12.9-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]


 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
    "12.6": (
-        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'"
+        "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'"
    ),
    "12.8": (
-        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'"
-    ),
-    "12.9": (
-        "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'"
+        "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'"
    ),
    "13.0": (
-        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'"
+        "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'"
    ),
    "xpu": (
        "intel-cmplr-lib-rt==2025.2.1 | "
@ -240,12 +221,8 @@ def generate_libtorch_matrix(
        if os == "linux":
            arches += CUDA_ARCHES
            arches += ROCM_ARCHES
-            if "13.0" in arches:
-                arches.remove("13.0")
        elif os == "windows":
            arches += CUDA_ARCHES
-            if "13.0" in arches:
-                arches.remove("13.0")
    if libtorch_variants is None:
        libtorch_variants = [
            "shared-with-deps",
@ -310,8 +287,6 @@ def generate_wheels_matrix(
            arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
        elif os == "windows":
            arches += CUDA_ARCHES + XPU_ARCHES
-            if "13.0" in arches:
-                arches.remove("13.0")
        elif os == "linux-aarch64":
            # Separate new if as the CPU type is different and
            # uses different build/test scripts
@ -334,19 +309,20 @@ def generate_wheels_matrix(
                else arch_version
            )

-            # TODO: Enable python 3.13t on cpu-s390x
-            if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
-                continue
            # TODO: Enable python 3.14 for rest
-            if os not in ["linux", "linux-aarch64", "macos-arm64", "windows"] and (
-                python_version == "3.14" or python_version == "3.14t"
-            ):
+            if os not in [
+                "linux",
+                "linux-aarch64",
+                "linux-s390x",
+                "macos-arm64",
+                "windows",
+            ] and (python_version == "3.14" or python_version == "3.14t"):
                continue

            # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install

            if (
-                arch_version in ["13.0", "12.9", "12.8", "12.6"]
+                arch_version in ["13.0", "12.8", "12.6"]
                and os == "linux"
                or arch_version in CUDA_AARCH64_ARCHES
            ):
@ -410,6 +386,5 @@ def generate_wheels_matrix(


 validate_nccl_dep_consistency("13.0")
-validate_nccl_dep_consistency("12.9")
 validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")
--- a/.github/scripts/prepare_vllm_wheels.sh
+++ b/.github/scripts/prepare_vllm_wheels.sh
@ -0,0 +1,91 @@
+#!/usr/bin/env bash
+
+set -eux
+
+torch_version=$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
+nightly=$(echo ${torch_version} | cut -d'.' -f4)
+
+# Copied from .ci/manywheel/build_common.sh
+make_wheel_record() {
+  fpath=$1
+  if echo $fpath | grep RECORD >/dev/null 2>&1; then
+    echo "$fpath,,"
+  else
+    fhash=$(openssl dgst -sha256 -binary $fpath | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
+    fsize=$(ls -nl $fpath | awk '{print $5}')
+    echo "$fpath,sha256=$fhash,$fsize"
+  fi
+}
+
+change_wheel_version() {
+  local package=$1
+  local wheel=$2
+  local f_version=$3
+  local t_version=$4
+
+  # Extract the wheel
+  ${PYTHON_EXECUTABLE} -mwheel unpack $wheel
+
+  mv "${package}-${f_version}" "${package}-${t_version}"
+  # Change the version from f_version to t_version in the dist-info dir
+  pushd "${package}-${t_version}"
+  mv "${package}-${f_version}.dist-info" "${package}-${t_version}.dist-info"
+
+  pushd "${package}-${t_version}.dist-info"
+  sed -i "s/${package}-${f_version}.dist-info/${package}-${t_version}.dist-info/g" RECORD
+
+  # Update the version in METADATA and its SHA256 hash
+  sed -i "s/Version: ${f_version}/Version: ${t_version}/g" METADATA
+  # then add PyTorch nightly dependency of vLLM
+  if [[ "${package}" == vllm ]] || [[ "${package}" == xformers ]]; then
+    sed -i "/License-File/a\Requires-Dist: torch==${torch_version}" METADATA
+  fi
+  sed -i '/METADATA,sha256/d' RECORD
+  popd
+
+  make_wheel_record "${package}-${t_version}.dist-info/METADATA" >> "${package}-${t_version}.dist-info/RECORD"
+  popd
+
+  # Repack the wheel
+  ${PYTHON_EXECUTABLE} -mwheel pack "${package}-${t_version}"
+
+  # Clean up
+  rm -rf "${package}-${t_version}"
+}
+
+repackage_wheel() {
+  local package=$1
+  pushd $package
+
+  local orig_wheel=$(find . -name *${package//-/_}*)
+  local orig_version=$(unzip -p $orig_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
+
+  local version=""
+  if [[ "${package}" == vllm ]]; then
+    # Copied from vllm/.buildkite/scripts/upload-wheels.sh
+    version=1.0.0
+  else
+    version=$(echo $orig_version | tr '.+' '.' | cut -d'.' -f1-3)
+  fi
+  local nightly_version=$version.$nightly
+
+  # Use nightly version
+  change_wheel_version ${package//-/_} $orig_wheel $orig_version $nightly_version
+  # Clean up
+  rm "${orig_wheel}"
+
+  auditwheel repair --plat $PLATFORM *.whl \
+    --exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
+  local repair_wheel=$(find wheelhouse -name *${PLATFORM}*)
+  local repair_wheel=$(basename ${repair_wheel})
+  popd
+
+  cp ${package}/wheelhouse/${repair_wheel} .
+  rm -rf $package
+}
+
+pushd externals/vllm/wheels
+for package in xformers flashinfer-python vllm; do
+  repackage_wheel $package
+done
+popd
--- a/.github/scripts/test_trymerge.py
+++ b/.github/scripts/test_trymerge.py
@ -27,6 +27,7 @@ from trymerge import (
    get_drci_classifications,
    gh_get_team_members,
    GitHubPR,
+    iter_issue_timeline_until_comment,
    JobCheckState,
    main as trymerge_main,
    MandatoryChecksMissingError,
@ -34,6 +35,8 @@ from trymerge import (
    RE_GHSTACK_DESC,
    read_merge_rules,
    remove_job_name_suffix,
+    sha_from_committed_event,
+    sha_from_force_push_after,
    validate_revert,
 )

@ -1138,5 +1141,176 @@ Pull Request resolved: https://github.com/pytorch/pytorch/pull/154394"""
        )


+@mock.patch("trymerge.gh_graphql", side_effect=mocked_gh_graphql)
+@mock.patch("trymerge.gh_fetch_merge_base", return_value="")
+@mock.patch(
+    "trymerge.get_drci_classifications", side_effect=mocked_drci_classifications
+)
+class TestTimelineFunctions(TestCase):
+    """Tests for the new timeline-related functions"""
+
+    def test_sha_from_committed_event(self, *args: Any) -> None:
+        """Test extracting SHA from committed event"""
+        # Based on actual GitHub API format - committed events have "sha" at top level
+        event = {
+            "event": "committed",
+            "sha": "fb21ce932ded6670c918804a0d9151b773770a7c",
+        }
+        self.assertEqual(
+            sha_from_committed_event(event), "fb21ce932ded6670c918804a0d9151b773770a7c"
+        )
+
+        # Test with missing SHA
+        event_no_sha = {"event": "committed"}
+        self.assertIsNone(sha_from_committed_event(event_no_sha))
+
+    def test_sha_from_force_push_after(self, *args: Any) -> None:
+        """Test extracting SHA from force push event"""
+        # NOTE: The current function doesn't handle the actual GitHub API format
+        # Real force push events have "commit_id" at top level, but this function
+        # looks for "after", "after_commit", "after_sha", or "head_sha" fields
+
+        # Test with the legacy format the current function handles
+        event_legacy = {
+            "event": "head_ref_force_pushed",
+            "after": {"sha": "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e"},
+        }
+        self.assertEqual(
+            sha_from_force_push_after(event_legacy),
+            "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
+        )
+
+        # Test with current GitHub API format (should return None with current implementation)
+        event_real_api = {
+            "event": "head_ref_force_pushed",
+            "commit_id": "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
+        }
+        self.assertEqual(
+            sha_from_force_push_after(event_real_api),
+            "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
+        )  # Current function doesn't handle commit_id
+
+        # Test with missing SHA
+        event_no_sha = {"event": "head_ref_force_pushed"}
+        self.assertIsNone(sha_from_force_push_after(event_no_sha))
+
+    @mock.patch("trymerge.gh_fetch_json_list")
+    def test_iter_issue_timeline_until_comment(
+        self, mock_gh_fetch_json_list: Any, *args: Any
+    ) -> None:
+        """Test timeline iteration until target comment"""
+        # Mock timeline data based on actual GitHub API format
+        timeline_data = [
+            {"event": "commented", "id": 100, "body": "first comment"},
+            {"event": "committed", "sha": "fb21ce932ded6670c918804a0d9151b773770a7c"},
+            {"event": "commented", "id": 200, "body": "target comment"},
+            {"event": "commented", "id": 300, "body": "after target"},
+        ]
+        mock_gh_fetch_json_list.return_value = timeline_data
+
+        # Test iteration stops at target comment
+        events = list(iter_issue_timeline_until_comment("pytorch", "pytorch", 123, 200))
+        self.assertEqual(len(events), 3)  # Should stop at target comment
+        self.assertEqual(events[0]["event"], "commented")
+        self.assertEqual(events[0]["id"], 100)
+        self.assertEqual(events[1]["event"], "committed")
+        self.assertEqual(events[1]["sha"], "fb21ce932ded6670c918804a0d9151b773770a7c")
+        self.assertEqual(events[2]["event"], "commented")
+        self.assertEqual(events[2]["id"], 200)
+
+    @mock.patch("trymerge.gh_fetch_json_list")
+    def test_iter_issue_timeline_until_comment_not_found(
+        self, mock_gh_fetch_json_list: Any, *args: Any
+    ) -> None:
+        """Test timeline iteration when target comment is not found"""
+        # Mock empty timeline
+        mock_gh_fetch_json_list.return_value = []
+
+        events = list(iter_issue_timeline_until_comment("pytorch", "pytorch", 123, 999))
+        self.assertEqual(len(events), 0)
+
+    @mock.patch("trymerge.iter_issue_timeline_until_comment")
+    def test_get_commit_sha_at_comment_commit_after_comment(
+        self, mock_iter_timeline: Any, *args: Any
+    ) -> None:
+        """Test get_commit_sha_at_comment returns correct SHA after comment"""
+        mock_iter_timeline.return_value = [
+            {"event": "committed", "sha": "commit1"},
+            {"event": "committed", "sha": "commit2"},
+            {"event": "commented", "id": 100},
+            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
+        ]
+        pr = GitHubPR("pytorch", "pytorch", 77700)
+        sha = pr.get_commit_sha_at_comment(100)
+        self.assertEqual(sha, "commit2")
+
+    @mock.patch("trymerge.iter_issue_timeline_until_comment")
+    def test_get_commit_sha_at_comment_force_push_before_comment(
+        self, mock_iter_timeline: Any, *args: Any
+    ) -> None:
+        mock_iter_timeline.return_value = [
+            {"event": "committed", "sha": "commit1"},
+            {"event": "committed", "sha": "commit2"},
+            {"event": "head_ref_force_pushed", "commit_id": "commit3"},
+            {"event": "commented", "id": 100},
+        ]
+        pr = GitHubPR("pytorch", "pytorch", 77700)
+        sha = pr.get_commit_sha_at_comment(100)
+        self.assertEqual(sha, "commit3")
+
+    @mock.patch("trymerge.iter_issue_timeline_until_comment")
+    def test_get_commit_sha_at_comment_force_push_before_comment_legacy_mode(
+        self, mock_iter_timeline: Any, *args: Any
+    ) -> None:
+        mock_iter_timeline.return_value = [
+            {"event": "committed", "sha": "commit1"},
+            {"event": "committed", "sha": "commit2"},
+            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
+            {"event": "commented", "id": 100},
+        ]
+        pr = GitHubPR("pytorch", "pytorch", 77700)
+        sha = pr.get_commit_sha_at_comment(100)
+        self.assertEqual(sha, "commit3")
+
+    @mock.patch("trymerge.iter_issue_timeline_until_comment")
+    def test_get_commit_sha_at_comment_multiple_comments(
+        self, mock_iter_timeline: Any, *args: Any
+    ) -> None:
+        mock_iter_timeline.return_value = [
+            {"event": "committed", "sha": "commit1"},
+            {"event": "commented", "id": 100},
+            {"event": "committed", "sha": "commit2"},
+            {"event": "commented", "id": 200},
+            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
+            {"event": "commented", "id": 300},
+        ]
+        pr = GitHubPR("pytorch", "pytorch", 77700)
+        sha = pr.get_commit_sha_at_comment(200)
+        self.assertEqual(sha, "commit2")
+        sha = pr.get_commit_sha_at_comment(300)
+        self.assertEqual(sha, "commit3")
+
+    @mock.patch("trymerge.iter_issue_timeline_until_comment")
+    def test_get_commit_sha_at_comment_no_events(
+        self, mock_iter_timeline: Any, *args: Any
+    ) -> None:
+        mock_iter_timeline.return_value = [
+            {"event": "commented", "id": 100},
+            {"event": "labeled", "label": {"name": "test"}},
+        ]
+        pr = GitHubPR("pytorch", "pytorch", 77700)
+        sha = pr.get_commit_sha_at_comment(100)
+        self.assertIsNone(sha)
+
+    @mock.patch("trymerge.iter_issue_timeline_until_comment")
+    def test_get_commit_sha_at_comment_exception(
+        self, mock_iter_timeline: Any, *args: Any
+    ) -> None:
+        mock_iter_timeline.side_effect = Exception("API error")
+        pr = GitHubPR("pytorch", "pytorch", 77700)
+        sha = pr.get_commit_sha_at_comment(100)
+        self.assertIsNone(sha)
+
+
 if __name__ == "__main__":
    main()
--- a/.github/scripts/trymerge.py
+++ b/.github/scripts/trymerge.py
@ -450,6 +450,63 @@ HAS_NO_CONNECTED_DIFF_TITLE = (
 IGNORABLE_FAILED_CHECKS_THESHOLD = 10


+def iter_issue_timeline_until_comment(
+    org: str, repo: str, issue_number: int, target_comment_id: int, max_pages: int = 200
+) -> Any:
+    """
+    Yield timeline entries in order until (and including) the entry whose id == target_comment_id
+    for a 'commented' event. Stops once the target comment is encountered.
+    """
+    page = 1
+
+    while page <= max_pages:
+        url = (
+            f"https://api.github.com/repos/{org}/{repo}/issues/{issue_number}/timeline"
+        )
+        params = {"per_page": 100, "page": page}
+
+        batch = gh_fetch_json_list(url, params)
+
+        if not batch:
+            return
+        for ev in batch:
+            # The target is the issue comment row with event == "commented" and id == issue_comment_id
+            if ev.get("event") == "commented" and ev.get("id") == target_comment_id:
+                yield ev  # nothing in the timeline after this matters, so stop early
+                return
+            yield ev
+        if len(batch) < 100:
+            return
+        page += 1
+
+    # If we got here without finding the comment, then we either hit a bug or some github PR
+    # has a _really_ long timeline.
+    # The max # of pages found on any pytorch/pytorch PR at the time of this change was 41
+    raise RuntimeError(
+        f"Could not find a merge commit in the first {max_pages} pages of the timeline at url {url}."
+        f"This is most likely a bug, please report it to the @pytorch/pytorch-dev-infra team."
+    )
+
+
+def sha_from_committed_event(ev: dict[str, Any]) -> Optional[str]:
+    """Extract SHA from committed event in timeline"""
+    return ev.get("sha")
+
+
+def sha_from_force_push_after(ev: dict[str, Any]) -> Optional[str]:
+    """Extract SHA from force push event in timeline"""
+    # The current GitHub API format
+    commit_id = ev.get("commit_id")
+    if commit_id:
+        return str(commit_id)
+
+    # Legacy format
+    after = ev.get("after") or ev.get("after_commit") or {}
+    if isinstance(after, dict):
+        return after.get("sha") or after.get("oid")
+    return ev.get("after_sha") or ev.get("head_sha")
+
+
 def gh_get_pr_info(org: str, proj: str, pr_no: int) -> Any:
    rc = gh_graphql(GH_GET_PR_INFO_QUERY, name=proj, owner=org, number=pr_no)
    return rc["data"]["repository"]["pullRequest"]
@ -843,6 +900,44 @@ class GitHubPR:
    def get_commit_count(self) -> int:
        return int(self.info["commits_with_authors"]["totalCount"])

+    def get_commit_sha_at_comment(self, comment_id: int) -> Optional[str]:
+        """
+        Get the PR head commit SHA that was present when a specific comment was posted.
+        This ensures we only merge the state of the PR at the time the merge command was issued,
+        not any subsequent commits that may have been pushed after.
+
+        Returns None if no head-changing events found before the comment or if the comment was not found.
+        """
+        head = None
+
+        try:
+            for event in iter_issue_timeline_until_comment(
+                self.org, self.project, self.pr_num, comment_id
+            ):
+                etype = event.get("event")
+                if etype == "committed":
+                    sha = sha_from_committed_event(event)
+                    if sha:
+                        head = sha
+                        print(f"Timeline: Found commit event for SHA {sha}")
+                elif etype == "head_ref_force_pushed":
+                    sha = sha_from_force_push_after(event)
+                    if sha:
+                        head = sha
+                        print(f"Timeline: Found force push event for SHA {sha}")
+                elif etype == "commented":
+                    if event.get("id") == comment_id:
+                        print(f"Timeline: Found final comment with sha {sha}")
+                        return head
+        except Exception as e:
+            print(
+                f"Warning: Failed to reconstruct timeline for comment {comment_id}: {e}"
+            )
+            return None
+
+        print(f"Did not find comment with id {comment_id} in the PR timeline")
+        return None
+
    def get_pr_creator_login(self) -> str:
        return cast(str, self.info["author"]["login"])

@ -1234,11 +1329,14 @@ class GitHubPR:
        skip_all_rule_checks: bool = False,
    ) -> list["GitHubPR"]:
        """
-        :param skip_all_rule_checks: If true, skips all rule checks, useful for dry-running merge locally
+        :param skip_all_rule_checks: If true, skips all rule checks on ghstack PRs, useful for dry-running merge locally
        """
        branch_to_merge_into = self.default_branch() if branch is None else branch
        if repo.current_branch() != branch_to_merge_into:
            repo.checkout(branch_to_merge_into)
+
+        # It's okay to skip the commit SHA check for ghstack PRs since
+        # authoring requires write access to the repo.
        if self.is_ghstack_pr():
            return self.merge_ghstack_into(
                repo,
@ -1249,14 +1347,41 @@ class GitHubPR:

        msg = self.gen_commit_message()
        pr_branch_name = f"__pull-request-{self.pr_num}__init__"
-        repo.fetch(self.last_commit_sha(), pr_branch_name)
+
+        # Determine which commit SHA to merge
+        commit_to_merge = None
+        if not comment_id:
+            raise ValueError("Must provide --comment-id when merging regular PRs")
+
+        # Get the commit SHA that was present when the comment was made
+        commit_to_merge = self.get_commit_sha_at_comment(comment_id)
+        if not commit_to_merge:
+            raise RuntimeError(
+                f"Could not find commit that was pushed before comment {comment_id}"
+            )
+
+        # Validate that this commit is the latest commit on the PR
+        latest_commit = self.last_commit_sha()
+        if commit_to_merge != latest_commit:
+            raise RuntimeError(
+                f"Commit {commit_to_merge} was HEAD when comment {comment_id} was posted "
+                f"but now the latest commit on the PR is {latest_commit}. "
+                f"Please re-issue the merge command to merge the latest commit."
+            )
+
+        print(f"Merging commit {commit_to_merge} locally")
+
+        repo.fetch(commit_to_merge, pr_branch_name)
        repo._run_git("merge", "--squash", pr_branch_name)
        repo._run_git("commit", f'--author="{self.get_author()}"', "-m", msg)

        # Did the PR change since we started the merge?
        pulled_sha = repo.show_ref(pr_branch_name)
        latest_pr_status = GitHubPR(self.org, self.project, self.pr_num)
-        if pulled_sha != latest_pr_status.last_commit_sha():
+        if (
+            pulled_sha != latest_pr_status.last_commit_sha()
+            or pulled_sha != commit_to_merge
+        ):
            raise RuntimeError(
                "PR has been updated since CI checks last passed. Please rerun the merge command."
            )
--- a/.github/templates/linux_binary_build_workflow.yml.j2
+++ b/.github/templates/linux_binary_build_workflow.yml.j2
@ -135,7 +135,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
--- a/.github/templates/macos_binary_build_workflow.yml.j2
+++ b/.github/templates/macos_binary_build_workflow.yml.j2
@ -68,11 +68,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      !{{ common.checkout(deep_clone=False, directory="pytorch") }}
      - name: Populate binary env
        run: |
--- a/.github/templates/upload.yml.j2
+++ b/.github/templates/upload.yml.j2
@ -33,7 +33,7 @@
  {%- if is_windows %}
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
  {%- endif %}

 {%- else %}
--- a/.github/workflows/_win-build.yml
+++ b/.github/workflows/_win-build.yml
@ -77,6 +77,7 @@ jobs:
        run: |
          git config --global core.longpaths true
          git config --global core.symlinks true
+          git config --global core.ignorecase false

          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
          # the directory on Windows and prevent GHA from checking out as reported
--- a/.github/workflows/_win-test.yml
+++ b/.github/workflows/_win-test.yml
@ -70,6 +70,7 @@ jobs:
        run: |
          git config --global core.longpaths true
          git config --global core.symlinks true
+          git config --global core.ignorecase false

          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
          # the directory on Windows and prevent GHA from checking out as reported
--- a/.github/workflows/_xpu-test.yml
+++ b/.github/workflows/_xpu-test.yml
@ -275,7 +275,7 @@ jobs:
      - name: Change permissions
        if: ${{ always() && steps.test.conclusion }}
        run: |
-          docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "sudo chown -R 1001:1001 test"
+          docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "sudo chown -R 1000:1000 test"

      - name: Print remaining test logs
        shell: bash
--- a/.github/workflows/build-manywheel-images.yml
+++ b/.github/workflows/build-manywheel-images.yml
@ -47,12 +47,11 @@ jobs:
      matrix:
        include: [
          { name: "manylinux2_28-builder",          tag: "cuda13.0",         runner: "linux.9xlarge.ephemeral" },
-          { name: "manylinux2_28-builder",          tag: "cuda12.9",         runner: "linux.9xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "cuda12.8",          runner: "linux.9xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "cuda12.6",          runner: "linux.9xlarge.ephemeral" },
          { name: "manylinuxaarch64-builder",       tag: "cuda13.0",          runner: "linux.arm64.2xlarge.ephemeral" },
-          { name: "manylinuxaarch64-builder",       tag: "cuda12.9",          runner: "linux.arm64.2xlarge.ephemeral" },
          { name: "manylinuxaarch64-builder",       tag: "cuda12.8",          runner: "linux.arm64.2xlarge.ephemeral" },
+          { name: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "rocm6.3",           runner: "linux.9xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "cpu",               runner: "linux.9xlarge.ephemeral" },
--- a/.github/workflows/build-vllm-wheel.yml
+++ b/.github/workflows/build-vllm-wheel.yml
@ -0,0 +1,219 @@
+name: Build vLLM wheels
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - .github/workflows/build-vllm-wheel.yml
+      - .github/ci_commit_pins/vllm.txt
+  workflow_dispatch:
+  pull_request:
+    paths:
+      - .github/workflows/build-vllm-wheel.yml
+      - .github/ci_commit_pins/vllm.txt
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  build-wheel:
+    if: github.repository_owner == 'pytorch'
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: [ '3.12' ]
+        # TODO (huydhn): Add cu130 https://github.com/pytorch/pytorch/pull/162000#issuecomment-3261541554
+        device: [ 'cu128', 'cu129' ]
+        runner: [ 'linux.12xlarge.memory' ]
+        include:
+          - device: cu128
+            manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.8'
+          - device: cu129
+            manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
+    name: "Build ${{ matrix.device }} vLLM wheel"
+    runs-on: ${{ matrix.runner }}
+    timeout-minutes: 480
+    env:
+      PY_VERS: ${{ matrix.python-version }}
+      MANYLINUX_IMAGE: ${{ matrix.manylinux-image }}
+      PLATFORM: 'manylinux_2_28_x86_64'
+      BUILD_DEVICE: ${{ matrix.device }}
+    steps:
+      - name: Setup SSH (Click me for login details)
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Checkout PyTorch
+        uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
+        with:
+          submodules: false
+
+      - name: Setup Linux
+        uses: ./.github/actions/setup-linux
+
+      - name: Get latest PyTorch nightly
+        shell: bash
+        run: |
+          set -eux
+
+          # Determine python executable for given version (copied from build-triton-wheel)
+          case $PY_VERS in
+          3.10)
+            PYTHON_EXECUTABLE=/opt/python/cp310-cp310/bin/python
+            ;;
+          3.11)
+            PYTHON_EXECUTABLE=/opt/python/cp311-cp311/bin/python
+            ;;
+          3.12)
+            PYTHON_EXECUTABLE=/opt/python/cp312-cp312/bin/python
+            ;;
+          3.13)
+            PYTHON_EXECUTABLE=/opt/python/cp313-cp313/bin/python
+            ;;
+          3.13t)
+            PYTHON_EXECUTABLE=/opt/python/cp313-cp313t/bin/python
+            ;;
+          3.14)
+            PYTHON_EXECUTABLE=/opt/python/cp314-cp314/bin/python
+            ;;
+          3.14t)
+            PYTHON_EXECUTABLE=/opt/python/cp314-cp314t/bin/python
+            ;;
+          *)
+            echo "Unsupported python version ${PY_VERS}"
+            exit 1
+            ;;
+          esac
+
+          # Keep PyTorch nightly wheel here so that we can install it later during
+          # vLLM build process
+          mkdir -p "${RUNNER_TEMP}/artifacts/"
+
+          container_name=$(docker run \
+            --tty \
+            --detach \
+            -e PLATFORM \
+            -e PYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" \
+            -v "${GITHUB_WORKSPACE}:/pytorch" \
+            -v "${RUNNER_TEMP}/artifacts:/artifacts" \
+            -w /artifacts/ \
+            "${MANYLINUX_IMAGE}"
+          )
+
+          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip install \
+            --pre torch torchvision torchaudio \
+            --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
+
+          # I wonder if there is a command to both download and install the wheels
+          # in one go
+          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip download \
+            --pre torch torchvision torchaudio \
+            --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
+
+          # Save this for later
+          echo "container_name=${container_name}" >> "$GITHUB_ENV"
+
+      - name: Build vLLM wheel
+        uses: ./.github/actions/build-external-packages
+        with:
+          build-targets: vllm
+          docker-image: ${{ env.MANYLINUX_IMAGE }}
+          cuda-arch-list: '8.0;8.9;9.0;10.0;12.0'
+          torch-wheel-dir: ${{ runner.temp }}/artifacts
+          output-dir: ${{ runner.temp }}/artifacts/externals
+
+      - name: Prepare vLLM wheel
+        shell: bash
+        run: |
+          set -eux
+
+          # Get these wheels ready, the vllm renaming logic is copied from its .buildkite/scripts/upload-wheels.sh
+          docker exec -t "${container_name}" bash -c /pytorch/.github/scripts/prepare_vllm_wheels.sh
+          docker exec -t "${container_name}" chown -R 1000:1000 /artifacts
+
+      - uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
+        with:
+          name: vllm-wheel-${{ matrix.device }}-${{ matrix.python-version }}-${{ env.PLATFORM }}
+          if-no-files-found: error
+          path: ${{ runner.temp }}/artifacts/externals/vllm/wheels/*.whl
+
+      - name: Teardown Linux
+        uses: pytorch/test-infra/.github/actions/teardown-linux@main
+        if: always()
+
+  # Copied from build-triton-wheel workflow (mostly)
+  upload-wheel:
+    name: "Upload ${{ matrix.device }} vLLM wheel"
+    needs:
+      - build-wheel
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        device: [ 'cu128', 'cu129' ]
+    env:
+      BUILD_DEVICE: ${{ matrix.device }}
+    permissions:
+      id-token: write
+      contents: read
+    container:
+      image: continuumio/miniconda3:4.12.0
+    environment: ${{ (github.event_name == 'push' && github.event.ref == 'refs/heads/main') && 'nightly-wheel-upload' || '' }}
+    steps:
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+
+      - name: Configure AWS credentials(PyTorch account) for main
+        if: ${{ github.event_name == 'push' && github.event.ref == 'refs/heads/main' }}
+        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+        with:
+          role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_nightly_build_wheels
+          aws-region: us-east-1
+
+      - name: Configure AWS credentials(PyTorch account) for RC builds
+        if: ${{ github.event_name == 'push' &&  (startsWith(github.event.ref, 'refs/tags/') && !startsWith(github.event.ref, 'refs/tags/ciflow/')) }}
+        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+        with:
+          role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_test_build_wheels
+          aws-region: us-east-1
+
+      - name: Download Build Artifacts
+        uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
+        with:
+          # Download all available artifacts
+          path: ${{ runner.temp }}/artifacts-all
+
+      - name: Select Wheel Artifacts
+        shell: bash
+        run: |
+          set -eux
+          mkdir -p "${RUNNER_TEMP}/artifacts/"
+          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-*/* "${RUNNER_TEMP}/artifacts/"
+
+      - name: Set DRY_RUN (only for tagged pushes)
+        if: ${{ github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v')) }}
+        shell: bash
+        run: |
+          echo "DRY_RUN=disabled" >> "$GITHUB_ENV"
+
+      - name: Set UPLOAD_CHANNEL (only for tagged pushes)
+        if: ${{ github.event_name == 'push' && startsWith(github.event.ref, 'refs/tags/v') }}
+        shell: bash
+        run: |
+          set -ex
+
+          if [[ "${GITHUB_REF_NAME}" = *-rc[0-9]* ]]; then
+            echo "UPLOAD_CHANNEL=test" >> "$GITHUB_ENV"
+          fi
+
+      - name: Upload binaries
+        env:
+          PACKAGE_TYPE: wheel
+          UPLOAD_SUBFOLDER: ${{ env.BUILD_DEVICE }}
+          PKG_DIR: ${{ runner.temp }}/artifacts
+        shell: bash
+        run: |
+          set -ex
+          bash .circleci/scripts/binary_upload.sh
--- a/.github/workflows/docker-builds.yml
+++ b/.github/workflows/docker-builds.yml
@ -50,10 +50,9 @@ jobs:
        runner: [linux.12xlarge]
        docker-image-name: [
          pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11,
+          pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11,
          pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm,
          pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks,
          pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9,
          pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
          pytorch-linux-jammy-py3.10-clang12,
@ -64,7 +63,7 @@ jobs:
          pytorch-linux-jammy-rocm-n-py3-benchmarks,
          pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12,
          pytorch-linux-jammy-py3.10-gcc11,
-          pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks,
+          pytorch-linux-jammy-py3-gcc11-inductor-benchmarks,
          pytorch-linux-jammy-py3.12-halide,
          pytorch-linux-jammy-xpu-n-1-py3,
          pytorch-linux-jammy-xpu-n-py3,
--- a/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
@ -112,7 +112,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_10-cuda-aarch64-12_9-build:
+  manywheel-py3_10-cuda-aarch64-12_6-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -121,39 +121,85 @@ jobs:
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.10"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_10-cuda-aarch64-12_9
+      build_name: manywheel-py3_10-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_10-cuda-aarch64-12_9-upload:  # Uploading
+  manywheel-py3_10-cuda-aarch64-12_6-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: manywheel-py3_10-cuda-aarch64-12_9-build
+    needs: manywheel-py3_10-cuda-aarch64-12_6-build
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.10"
-      build_name: manywheel-py3_10-cuda-aarch64-12_9
+      build_name: manywheel-py3_10-cuda-aarch64-12_6
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_10-cuda-aarch64-12_8-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda-aarch64-12_8-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda-aarch64-12_8-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda-aarch64-12_8
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -178,7 +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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -269,7 +315,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_11-cuda-aarch64-12_9-build:
+  manywheel-py3_11-cuda-aarch64-12_6-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -278,39 +324,85 @@ jobs:
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.11"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_11-cuda-aarch64-12_9
+      build_name: manywheel-py3_11-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_11-cuda-aarch64-12_9-upload:  # Uploading
+  manywheel-py3_11-cuda-aarch64-12_6-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: manywheel-py3_11-cuda-aarch64-12_9-build
+    needs: manywheel-py3_11-cuda-aarch64-12_6-build
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.11"
-      build_name: manywheel-py3_11-cuda-aarch64-12_9
+      build_name: manywheel-py3_11-cuda-aarch64-12_6
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_11-cuda-aarch64-12_8-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda-aarch64-12_8-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda-aarch64-12_8-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda-aarch64-12_8
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -335,7 +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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -426,7 +518,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_12-cuda-aarch64-12_9-build:
+  manywheel-py3_12-cuda-aarch64-12_6-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -435,39 +527,85 @@ jobs:
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.12"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_12-cuda-aarch64-12_9
+      build_name: manywheel-py3_12-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda-aarch64-12_9-upload:  # Uploading
+  manywheel-py3_12-cuda-aarch64-12_6-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: manywheel-py3_12-cuda-aarch64-12_9-build
+    needs: manywheel-py3_12-cuda-aarch64-12_6-build
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda-aarch64-12_9
+      build_name: manywheel-py3_12-cuda-aarch64-12_6
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_12-cuda-aarch64-12_8-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda-aarch64-12_8-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda-aarch64-12_8-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda-aarch64-12_8
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -492,7 +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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -583,7 +721,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_13-cuda-aarch64-12_9-build:
+  manywheel-py3_13-cuda-aarch64-12_6-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -592,39 +730,85 @@ jobs:
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.13"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13-cuda-aarch64-12_9
+      build_name: manywheel-py3_13-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13-cuda-aarch64-12_9-upload:  # Uploading
+  manywheel-py3_13-cuda-aarch64-12_6-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: manywheel-py3_13-cuda-aarch64-12_9-build
+    needs: manywheel-py3_13-cuda-aarch64-12_6-build
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.13"
-      build_name: manywheel-py3_13-cuda-aarch64-12_9
+      build_name: manywheel-py3_13-cuda-aarch64-12_6
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_13-cuda-aarch64-12_8-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda-aarch64-12_8-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda-aarch64-12_8-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda-aarch64-12_8
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -649,7 +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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -740,7 +924,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_13t-cuda-aarch64-12_9-build:
+  manywheel-py3_13t-cuda-aarch64-12_6-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -749,39 +933,85 @@ jobs:
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.13t"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13t-cuda-aarch64-12_9
+      build_name: manywheel-py3_13t-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13t-cuda-aarch64-12_9-upload:  # Uploading
+  manywheel-py3_13t-cuda-aarch64-12_6-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: manywheel-py3_13t-cuda-aarch64-12_9-build
+    needs: manywheel-py3_13t-cuda-aarch64-12_6-build
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.13t"
-      build_name: manywheel-py3_13t-cuda-aarch64-12_9
+      build_name: manywheel-py3_13t-cuda-aarch64-12_6
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_13t-cuda-aarch64-12_8-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda-aarch64-12_8-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda-aarch64-12_8-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda-aarch64-12_8
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -806,7 +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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -897,7 +1127,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_14-cuda-aarch64-12_9-build:
+  manywheel-py3_14-cuda-aarch64-12_6-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -906,39 +1136,85 @@ jobs:
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.14"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14-cuda-aarch64-12_9
+      build_name: manywheel-py3_14-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cuda-aarch64-12_9-upload:  # Uploading
+  manywheel-py3_14-cuda-aarch64-12_6-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: manywheel-py3_14-cuda-aarch64-12_9-build
+    needs: manywheel-py3_14-cuda-aarch64-12_6-build
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cuda-aarch64-12_9
+      build_name: manywheel-py3_14-cuda-aarch64-12_6
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_14-cuda-aarch64-12_8-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda-aarch64-12_8-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda-aarch64-12_8-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda-aarch64-12_8
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -963,7 +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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1054,7 +1330,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_14t-cuda-aarch64-12_9-build:
+  manywheel-py3_14t-cuda-aarch64-12_6-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -1063,39 +1339,85 @@ jobs:
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.14t"
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.arm64.m7g.4xlarge.ephemeral
      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14t-cuda-aarch64-12_9
+      build_name: manywheel-py3_14t-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cuda-aarch64-12_9-upload:  # Uploading
+  manywheel-py3_14t-cuda-aarch64-12_6-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: manywheel-py3_14t-cuda-aarch64-12_9-build
+    needs: manywheel-py3_14t-cuda-aarch64-12_6-build
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: manywheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: "12.6-aarch64"
      GPU_ARCH_TYPE: cuda-aarch64
      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cuda-aarch64-12_9
+      build_name: manywheel-py3_14t-cuda-aarch64-12_6
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_14t-cuda-aarch64-12_8-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda-aarch64-12_8-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda-aarch64-12_8-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda-aarch64-12_8
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -1120,7 +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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-linux-binary-libtorch-nightly.yml
+++ b/.github/workflows/generated-linux-binary-libtorch-nightly.yml
@ -248,7 +248,7 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  libtorch-cuda12_9-shared-with-deps-release-build:
+  libtorch-cuda13_0-shared-with-deps-release-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
    needs: get-label-type
@ -257,22 +257,22 @@ jobs:
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
      LIBTORCH_CONFIG: release
      LIBTORCH_VARIANT: shared-with-deps
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_name: libtorch-cuda13_0-shared-with-deps-release
      build_environment: linux-binary-libtorch
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+  libtorch-cuda13_0-shared-with-deps-release-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - libtorch-cuda12_9-shared-with-deps-release-build
+      - libtorch-cuda13_0-shared-with-deps-release-build
      - get-label-type
    uses: ./.github/workflows/_binary-test-linux.yml
    with:
@ -280,38 +280,38 @@ jobs:
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
      LIBTORCH_CONFIG: release
      LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_name: libtorch-cuda13_0-shared-with-deps-release
      build_environment: linux-binary-libtorch
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+  libtorch-cuda13_0-shared-with-deps-release-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    needs: libtorch-cuda13_0-shared-with-deps-release-test
    with:
      PYTORCH_ROOT: /pytorch
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
      LIBTORCH_CONFIG: release
      LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_name: libtorch-cuda13_0-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
--- a/.github/workflows/generated-linux-binary-manywheel-main.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-main.yml
@ -60,7 +60,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_8-test:  # Testing
--- a/.github/workflows/generated-linux-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-nightly.yml
@ -127,7 +127,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda12_8-test:  # Testing
@ -241,72 +241,6 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_10-cuda12_9-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.10"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_10-cuda12_9
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_10-cuda12_9-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_10-cuda12_9-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.10"
-      build_name: manywheel-py3_10-cuda12_9
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_10-cuda12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_10-cuda12_9-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.10"
-      build_name: manywheel-py3_10-cuda12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
  manywheel-py3_10-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -325,7 +259,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda13_0-test:  # Testing
@ -638,7 +572,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
@ -785,7 +719,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda12_6-test:  # Testing
@ -851,7 +785,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda12_8-test:  # Testing
@ -899,72 +833,6 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_11-cuda12_9-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.11"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_11-cuda12_9
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_11-cuda12_9-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_11-cuda12_9-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.11"
-      build_name: manywheel-py3_11-cuda12_9
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_11-cuda12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_11-cuda12_9-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.11"
-      build_name: manywheel-py3_11-cuda12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
  manywheel-py3_11-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -983,7 +851,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda13_0-test:  # Testing
@ -1296,7 +1164,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
@ -1443,7 +1311,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_6-test:  # Testing
@ -1509,7 +1377,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_8-test:  # Testing
@ -1557,72 +1425,6 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_12-cuda12_9-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.12"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_12-cuda12_9
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda12_9-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_12-cuda12_9-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda12_9
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_12-cuda12_9-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
  manywheel-py3_12-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -1641,7 +1443,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda13_0-test:  # Testing
@ -1954,7 +1756,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
@ -2101,7 +1903,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda12_6-test:  # Testing
@ -2167,7 +1969,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda12_8-test:  # Testing
@ -2215,72 +2017,6 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_13-cuda12_9-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.13"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_13-cuda12_9
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13-cuda12_9-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_13-cuda12_9-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.13"
-      build_name: manywheel-py3_13-cuda12_9
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13-cuda12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_13-cuda12_9-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.13"
-      build_name: manywheel-py3_13-cuda12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
  manywheel-py3_13-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -2299,7 +2035,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda13_0-test:  # Testing
@ -2612,7 +2348,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
@ -2759,7 +2495,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda12_6-test:  # Testing
@ -2825,7 +2561,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda12_8-test:  # Testing
@ -2873,72 +2609,6 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_13t-cuda12_9-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.13t"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_13t-cuda12_9
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13t-cuda12_9-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_13t-cuda12_9-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.13t"
-      build_name: manywheel-py3_13t-cuda12_9
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13t-cuda12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_13t-cuda12_9-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.13t"
-      build_name: manywheel-py3_13t-cuda12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
  manywheel-py3_13t-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -2957,7 +2627,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda13_0-test:  # Testing
@ -3270,7 +2940,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
@ -3417,7 +3087,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda12_6-test:  # Testing
@ -3483,7 +3153,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda12_8-test:  # Testing
@ -3531,72 +3201,6 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_14-cuda12_9-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_14-cuda12_9
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cuda12_9-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_14-cuda12_9-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cuda12_9
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cuda12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14-cuda12_9-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cuda12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
  manywheel-py3_14-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -3615,7 +3219,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda13_0-test:  # Testing
@ -3928,7 +3532,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
@ -4075,7 +3679,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda12_6-test:  # Testing
@ -4141,7 +3745,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda12_8-test:  # Testing
@ -4189,72 +3793,6 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

-  manywheel-py3_14t-cuda12_9-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14t"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_14t-cuda12_9
-      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cuda12_9-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_14t-cuda12_9-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cuda12_9
-      build_environment: linux-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cuda12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14t-cuda12_9-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
-      GPU_ARCH_TYPE: cuda
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cuda12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
  manywheel-py3_14t-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -4273,7 +3811,7 @@ jobs:
      runner_prefix: "${{ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
+      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'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda13_0-test:  # Testing
@ -4586,7 +4124,7 @@ jobs:
      contents: read
    steps:
      - name: Setup XPU
-        uses: ./.github/actions/setup-xpu
+        uses: pytorch/pytorch/.github/actions/setup-xpu@main
      - name: configure aws credentials
        id: aws_creds
        uses: aws-actions/configure-aws-credentials@v4
--- a/.github/workflows/generated-linux-s390x-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-s390x-binary-manywheel-nightly.yml
@ -302,3 +302,195 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_13t-cpu-s390x-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.13t"
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+      timeout-minutes: 420
+      build_name: manywheel-py3_13t-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cpu-s390x-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13t-cpu-s390x-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cpu-s390x-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cpu-s390x-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cpu-s390x
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_14-cpu-s390x-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.14"
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+      timeout-minutes: 420
+      build_name: manywheel-py3_14-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cpu-s390x-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14-cpu-s390x-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cpu-s390x-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cpu-s390x-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cpu-s390x
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_14t-cpu-s390x-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.14t"
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+      timeout-minutes: 420
+      build_name: manywheel-py3_14t-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cpu-s390x-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14t-cpu-s390x-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cpu-s390x-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cpu-s390x-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cpu-s390x
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
--- a/.github/workflows/generated-macos-arm64-binary-libtorch-release-nightly.yml
+++ b/.github/workflows/generated-macos-arm64-binary-libtorch-release-nightly.yml
@ -46,7 +46,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -67,11 +67,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
--- a/.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml
+++ b/.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml
@ -63,11 +63,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
@ -208,11 +203,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
@ -353,11 +343,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
@ -498,11 +483,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
@ -643,11 +623,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
@ -788,11 +763,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
@ -933,11 +903,6 @@ jobs:
          chmod +x "${RUNNER_TEMP}/conda.sh"
          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
      - name: Checkout PyTorch
        uses: actions/checkout@v4
        with:
--- a/.github/workflows/generated-windows-arm64-binary-libtorch-debug-nightly.yml
+++ b/.github/workflows/generated-windows-arm64-binary-libtorch-debug-nightly.yml
@ -64,7 +64,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Populate binary env
        shell: cmd
@ -141,7 +141,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Populate binary env
        shell: cmd
@ -201,7 +201,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cpu-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-windows-arm64-binary-libtorch-release-nightly.yml
+++ b/.github/workflows/generated-windows-arm64-binary-libtorch-release-nightly.yml
@ -64,7 +64,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Populate binary env
        shell: cmd
@ -141,7 +141,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Populate binary env
        shell: cmd
@ -201,7 +201,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cpu-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-windows-binary-libtorch-debug-main.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-debug-main.yml
@ -51,7 +51,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -166,7 +166,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
--- a/.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml
@ -58,7 +58,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -173,7 +173,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -283,7 +283,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cpu-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -306,7 +306,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -422,7 +422,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -533,7 +533,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cuda12_6-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -556,7 +556,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -672,7 +672,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -783,12 +783,12 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cuda12_8-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  libtorch-cuda12_9-shared-with-deps-debug-build:
+  libtorch-cuda13_0-shared-with-deps-debug-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -798,15 +798,15 @@ jobs:
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      LIBTORCH_CONFIG: debug
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -884,7 +884,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: libtorch-cuda12_9-shared-with-deps-debug
+          name: libtorch-cuda13_0-shared-with-deps-debug
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -902,10 +902,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  libtorch-cuda12_9-shared-with-deps-debug-test:  # Testing
+  libtorch-cuda13_0-shared-with-deps-debug-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - libtorch-cuda12_9-shared-with-deps-debug-build
+      - libtorch-cuda13_0-shared-with-deps-debug-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -914,15 +914,15 @@ jobs:
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      LIBTORCH_CONFIG: debug
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -992,7 +992,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: libtorch-cuda12_9-shared-with-deps-debug
+          name: libtorch-cuda13_0-shared-with-deps-debug
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -1015,26 +1015,26 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  libtorch-cuda12_9-shared-with-deps-debug-upload:  # Uploading
+  libtorch-cuda13_0-shared-with-deps-debug-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: libtorch-cuda12_9-shared-with-deps-debug-test
+    needs: libtorch-cuda13_0-shared-with-deps-debug-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      LIBTORCH_CONFIG: debug
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
-      build_name: libtorch-cuda12_9-shared-with-deps-debug
+      DESIRED_PYTHON: "3.10"
+      build_name: libtorch-cuda13_0-shared-with-deps-debug
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
--- a/.github/workflows/generated-windows-binary-libtorch-release-main.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-release-main.yml
@ -51,7 +51,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -166,7 +166,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
--- a/.github/workflows/generated-windows-binary-libtorch-release-nightly.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-release-nightly.yml
@ -58,7 +58,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -173,7 +173,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -283,7 +283,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cpu-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -306,7 +306,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -422,7 +422,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -533,7 +533,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cuda12_6-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -556,7 +556,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -672,7 +672,7 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -783,12 +783,12 @@ jobs:
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
      build_name: libtorch-cuda12_8-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  libtorch-cuda12_9-shared-with-deps-release-build:
+  libtorch-cuda13_0-shared-with-deps-release-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -798,15 +798,15 @@ jobs:
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      LIBTORCH_CONFIG: release
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      # NOTE: These environment variables are put here so that they can be applied on every job equally
      #       They are also here because setting them at a workflow level doesn't give us access to the
@ -884,7 +884,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: libtorch-cuda12_9-shared-with-deps-release
+          name: libtorch-cuda13_0-shared-with-deps-release
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -902,10 +902,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+  libtorch-cuda13_0-shared-with-deps-release-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - libtorch-cuda12_9-shared-with-deps-release-build
+      - libtorch-cuda13_0-shared-with-deps-release-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -914,15 +914,15 @@ jobs:
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      LIBTORCH_CONFIG: release
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
+      DESIRED_PYTHON: "3.10"
    steps:
      - name: Display EC2 information
        shell: bash
@ -992,7 +992,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: libtorch-cuda12_9-shared-with-deps-release
+          name: libtorch-cuda13_0-shared-with-deps-release
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -1015,26 +1015,26 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+  libtorch-cuda13_0-shared-with-deps-release-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    needs: libtorch-cuda13_0-shared-with-deps-release-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: libtorch
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      LIBTORCH_CONFIG: release
      LIBTORCH_VARIANT: shared-with-deps
      # This is a dummy value for libtorch to work correctly with our batch scripts
      # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.9"
-      build_name: libtorch-cuda12_9-shared-with-deps-release
+      DESIRED_PYTHON: "3.10"
+      build_name: libtorch-cuda13_0-shared-with-deps-release
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
--- a/.github/workflows/generated-windows-binary-wheel-nightly.yml
+++ b/.github/workflows/generated-windows-binary-wheel-nightly.yml
@ -752,7 +752,7 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_10-cuda12_9-build:
+  wheel-py3_10-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -762,8 +762,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.10"
@ -844,7 +844,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: wheel-py3_10-cuda12_9
+          name: wheel-py3_10-cuda13_0
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -862,10 +862,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  wheel-py3_10-cuda12_9-test:  # Testing
+  wheel-py3_10-cuda13_0-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - wheel-py3_10-cuda12_9-build
+      - wheel-py3_10-cuda13_0-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -874,8 +874,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.10"
@ -948,7 +948,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: wheel-py3_10-cuda12_9
+          name: wheel-py3_10-cuda13_0
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -971,22 +971,22 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  wheel-py3_10-cuda12_9-upload:  # Uploading
+  wheel-py3_10-cuda13_0-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: wheel-py3_10-cuda12_9-test
+    needs: wheel-py3_10-cuda13_0-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DESIRED_PYTHON: "3.10"
-      build_name: wheel-py3_10-cuda12_9
+      build_name: wheel-py3_10-cuda13_0
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -1937,7 +1937,7 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_11-cuda12_9-build:
+  wheel-py3_11-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -1947,8 +1947,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.11"
@ -2029,7 +2029,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: wheel-py3_11-cuda12_9
+          name: wheel-py3_11-cuda13_0
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -2047,10 +2047,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  wheel-py3_11-cuda12_9-test:  # Testing
+  wheel-py3_11-cuda13_0-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - wheel-py3_11-cuda12_9-build
+      - wheel-py3_11-cuda13_0-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -2059,8 +2059,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.11"
@ -2133,7 +2133,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: wheel-py3_11-cuda12_9
+          name: wheel-py3_11-cuda13_0
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -2156,22 +2156,22 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  wheel-py3_11-cuda12_9-upload:  # Uploading
+  wheel-py3_11-cuda13_0-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: wheel-py3_11-cuda12_9-test
+    needs: wheel-py3_11-cuda13_0-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DESIRED_PYTHON: "3.11"
-      build_name: wheel-py3_11-cuda12_9
+      build_name: wheel-py3_11-cuda13_0
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -3122,7 +3122,7 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_12-cuda12_9-build:
+  wheel-py3_12-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -3132,8 +3132,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.12"
@ -3214,7 +3214,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: wheel-py3_12-cuda12_9
+          name: wheel-py3_12-cuda13_0
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -3232,10 +3232,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  wheel-py3_12-cuda12_9-test:  # Testing
+  wheel-py3_12-cuda13_0-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - wheel-py3_12-cuda12_9-build
+      - wheel-py3_12-cuda13_0-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -3244,8 +3244,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.12"
@ -3318,7 +3318,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: wheel-py3_12-cuda12_9
+          name: wheel-py3_12-cuda13_0
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -3341,22 +3341,22 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  wheel-py3_12-cuda12_9-upload:  # Uploading
+  wheel-py3_12-cuda13_0-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: wheel-py3_12-cuda12_9-test
+    needs: wheel-py3_12-cuda13_0-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DESIRED_PYTHON: "3.12"
-      build_name: wheel-py3_12-cuda12_9
+      build_name: wheel-py3_12-cuda13_0
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -4307,7 +4307,7 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_13-cuda12_9-build:
+  wheel-py3_13-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -4317,8 +4317,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.13"
@ -4399,7 +4399,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: wheel-py3_13-cuda12_9
+          name: wheel-py3_13-cuda13_0
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -4417,10 +4417,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  wheel-py3_13-cuda12_9-test:  # Testing
+  wheel-py3_13-cuda13_0-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - wheel-py3_13-cuda12_9-build
+      - wheel-py3_13-cuda13_0-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -4429,8 +4429,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.13"
@ -4503,7 +4503,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: wheel-py3_13-cuda12_9
+          name: wheel-py3_13-cuda13_0
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -4526,22 +4526,22 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  wheel-py3_13-cuda12_9-upload:  # Uploading
+  wheel-py3_13-cuda13_0-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: wheel-py3_13-cuda12_9-test
+    needs: wheel-py3_13-cuda13_0-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DESIRED_PYTHON: "3.13"
-      build_name: wheel-py3_13-cuda12_9
+      build_name: wheel-py3_13-cuda13_0
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -5492,7 +5492,7 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_13t-cuda12_9-build:
+  wheel-py3_13t-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -5502,8 +5502,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.13t"
@ -5584,7 +5584,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: wheel-py3_13t-cuda12_9
+          name: wheel-py3_13t-cuda13_0
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -5602,10 +5602,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  wheel-py3_13t-cuda12_9-test:  # Testing
+  wheel-py3_13t-cuda13_0-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - wheel-py3_13t-cuda12_9-build
+      - wheel-py3_13t-cuda13_0-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -5614,8 +5614,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.13t"
@ -5688,7 +5688,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: wheel-py3_13t-cuda12_9
+          name: wheel-py3_13t-cuda13_0
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -5711,22 +5711,22 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  wheel-py3_13t-cuda12_9-upload:  # Uploading
+  wheel-py3_13t-cuda13_0-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: wheel-py3_13t-cuda12_9-test
+    needs: wheel-py3_13t-cuda13_0-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DESIRED_PYTHON: "3.13t"
-      build_name: wheel-py3_13t-cuda12_9
+      build_name: wheel-py3_13t-cuda13_0
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -6677,7 +6677,7 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_14-cuda12_9-build:
+  wheel-py3_14-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -6687,8 +6687,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.14"
@ -6769,7 +6769,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: wheel-py3_14-cuda12_9
+          name: wheel-py3_14-cuda13_0
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -6787,10 +6787,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  wheel-py3_14-cuda12_9-test:  # Testing
+  wheel-py3_14-cuda13_0-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - wheel-py3_14-cuda12_9-build
+      - wheel-py3_14-cuda13_0-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -6799,8 +6799,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.14"
@ -6873,7 +6873,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: wheel-py3_14-cuda12_9
+          name: wheel-py3_14-cuda13_0
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -6896,22 +6896,22 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  wheel-py3_14-cuda12_9-upload:  # Uploading
+  wheel-py3_14-cuda13_0-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: wheel-py3_14-cuda12_9-test
+    needs: wheel-py3_14-cuda13_0-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DESIRED_PYTHON: "3.14"
-      build_name: wheel-py3_14-cuda12_9
+      build_name: wheel-py3_14-cuda13_0
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
@ -7862,7 +7862,7 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_14t-cuda12_9-build:
+  wheel-py3_14t-cuda13_0-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    needs: get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
@ -7872,8 +7872,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.14t"
@ -7954,7 +7954,7 @@ jobs:
      - uses: actions/upload-artifact@v4.4.0
        if: always()
        with:
-          name: wheel-py3_14t-cuda12_9
+          name: wheel-py3_14t-cuda13_0
          retention-days: 14
          if-no-files-found: error
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
@ -7972,10 +7972,10 @@ jobs:
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1

-  wheel-py3_14t-cuda12_9-test:  # Testing
+  wheel-py3_14t-cuda13_0-test:  # Testing
    if: ${{ github.repository_owner == 'pytorch' }}
    needs:
-      - wheel-py3_14t-cuda12_9-build
+      - wheel-py3_14t-cuda13_0-build
      - get-label-type
    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
    timeout-minutes: 360
@ -7984,8 +7984,8 @@ jobs:
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      SKIP_ALL_TESTS: 1
      DESIRED_PYTHON: "3.14t"
@ -8058,7 +8058,7 @@ jobs:
      - uses: actions/download-artifact@v4.1.7
        name: Download Build Artifacts
        with:
-          name: wheel-py3_14t-cuda12_9
+          name: wheel-py3_14t-cuda13_0
          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
      - name: Populate binary env
        shell: bash
@ -8081,22 +8081,22 @@ jobs:
        if: always()
        run: |
          .github\scripts\kill_active_ssh_sessions.ps1
-  wheel-py3_14t-cuda12_9-upload:  # Uploading
+  wheel-py3_14t-cuda13_0-upload:  # Uploading
    if: ${{ github.repository_owner == 'pytorch' }}
    permissions:
      id-token: write
      contents: read
-    needs: wheel-py3_14t-cuda12_9-test
+    needs: wheel-py3_14t-cuda13_0-test
    with:
      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
      PACKAGE_TYPE: wheel
      # TODO: This is a legacy variable that we eventually want to get rid of in
      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      DESIRED_CUDA: cu130
+      GPU_ARCH_VERSION: "13.0"
      GPU_ARCH_TYPE: cuda
      DESIRED_PYTHON: "3.14t"
-      build_name: wheel-py3_14t-cuda12_9
+      build_name: wheel-py3_14t-cuda13_0
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
--- a/.github/workflows/inductor-micro-benchmark-x86.yml
+++ b/.github/workflows/inductor-micro-benchmark-x86.yml
@ -18,13 +18,13 @@ permissions:
  contents: read

 jobs:
-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
+  inductor-build:
    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+    name: inductor-build
    uses: ./.github/workflows/_linux-build.yml
    with:
      build-environment: linux-jammy-py3.9-gcc11
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      # Use metal host for benchmark jobs
      test-matrix: |
        { include: [
@ -32,13 +32,13 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-micro-benchmark-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-micro-benchmark-test:
+    name: inductor-micro-benchmark-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: inductor-build
    with:
      build-environment: linux-jammy-py3.9-gcc11
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
      timeout-minutes: 720
    secrets: inherit
--- a/.github/workflows/inductor-micro-benchmark.yml
+++ b/.github/workflows/inductor-micro-benchmark.yml
@ -35,6 +35,8 @@ jobs:
    needs:
      - get-default-label-prefix
    with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
@ -43,6 +45,7 @@ jobs:
        { include: [
          { config: "inductor-micro-benchmark", shard: 1, num_shards: 1, runner: "linux.aws.a100", owners: ["oncall:pt2"] },
        ]}
+      build-additional-packages: "vision audio fbgemm torchao"
    secrets: inherit

  test:
--- a/.github/workflows/inductor-nightly.yml
+++ b/.github/workflows/inductor-nightly.yml
@ -32,13 +32,13 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  linux-jammy-cpu-py3_9-gcc11-nightly-dynamo-benchmarks-build:
-    name: linux-jammy-cpu-py3.9-gcc11-nightly-dynamo-benchmarks
+  nightly-dynamo-benchmarks-build:
+    name: nightly-dynamo-benchmarks-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-default-label-prefix
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
      test-matrix: |
        { include: [
@ -51,13 +51,13 @@ jobs:
      build-additional-packages: "vision audio torchao"
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-nightly-dynamo-benchmarks-test:
-    name: linux-jammy-cpu-py3.9-gcc11-nightly-dynamo-benchmarks
+  nightly-dynamo-benchmarks-test:
+    name: nightly-dynamo-benchmarks-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-nightly-dynamo-benchmarks-build
+    needs: nightly-dynamo-benchmarks-build
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-nightly-dynamo-benchmarks-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-nightly-dynamo-benchmarks-build.outputs.test-matrix }}
+      docker-image: ${{ needs.nightly-dynamo-benchmarks-build.outputs.docker-image }}
+      test-matrix: ${{ needs.nightly-dynamo-benchmarks-build.outputs.test-matrix }}
      timeout-minutes: 720
    secrets: inherit
--- a/.github/workflows/inductor-perf-test-nightly-h100.yml
+++ b/.github/workflows/inductor-perf-test-nightly-h100.yml
@ -84,9 +84,8 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  # NB: Keep this in sync with trunk.yml
  build:
-    name: cuda12.8-py3.10-gcc9-sm90
+    name: build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
@ -128,7 +127,7 @@ jobs:
    secrets: inherit

  test-periodically:
-    name: cuda12.8-py3.10-gcc9-sm90
+    name: test-periodically
    uses: ./.github/workflows/_linux-test.yml
    needs: build
    if: github.event.schedule == '15 0,12 * * 1-6'
@ -138,14 +137,13 @@ jobs:
      docker-image: ${{ needs.build.outputs.docker-image }}
      test-matrix: ${{ needs.build.outputs.test-matrix }}
      timeout-minutes: 720
-      # disable monitor in perf tests, next step is to enable it
      disable-monitor: false
      monitor-log-interval: 15
      monitor-data-collect-interval: 4
    secrets: inherit

  test-weekly:
-    name: cuda12.8-py3.10-gcc9-sm90
+    name: test-weekly
    uses: ./.github/workflows/_linux-test.yml
    needs: build
    if: github.event.schedule == '0 7 * * 0'
@ -155,14 +153,13 @@ jobs:
      docker-image: ${{ needs.build.outputs.docker-image }}
      test-matrix: ${{ needs.build.outputs.test-matrix }}
      timeout-minutes: 1440
-      # disable monitor in perf tests, next step is to enable it
      disable-monitor: false
      monitor-log-interval: 15
      monitor-data-collect-interval: 4
    secrets: inherit

  test:
-    name: cuda12.8-py3.10-gcc9-sm90
+    name: test
    uses: ./.github/workflows/_linux-test.yml
    needs: build
    # The pull_request trigger is used in PR to bump transformers pin which always
@ -174,7 +171,6 @@ jobs:
      docker-image: ${{ needs.build.outputs.docker-image }}
      test-matrix: ${{ needs.build.outputs.test-matrix }}
      timeout-minutes: 720
-      # disable monitor in perf tests for more investigation
      disable-monitor: false
      monitor-log-interval: 15
      monitor-data-collect-interval: 4
--- a/.github/workflows/inductor-perf-test-nightly-x86-zen.yml
+++ b/.github/workflows/inductor-perf-test-nightly-x86-zen.yml
@ -69,14 +69,14 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  linux-jammy-zen-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-zen-cpu-py3.9-gcc11-inductor
+  inductor-build:
+    name: inductor-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      test-matrix: |
        { include: [
          { config: "inductor_huggingface_perf_cpu_x86_zen", shard: 1, num_shards: 3, runner: "linux.24xlarge.amd" },
@ -95,16 +95,16 @@ jobs:
      selected-test-configs: ${{ inputs.benchmark_configs }}
    secrets: inherit

-  linux-jammy-zen-cpu-py3_9-gcc11-inductor-test-nightly:
-    name: linux-jammy-zen-cpu-py3.9-gcc11-inductor
+  inductor-test-nightly:
+    name: inductor-test-nightly
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-zen-cpu-py3_9-gcc11-inductor-build
+    needs: inductor-build
    if: github.event.schedule == '0 7 * * *'
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
      dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true
-      docker-image: ${{ needs.linux-jammy-zen-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-zen-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
      timeout-minutes: 720
      # disable monitor in perf tests
      disable-monitor: false
@ -112,17 +112,16 @@ jobs:
      monitor-data-collect-interval: 4
    secrets: inherit

-
-  linux-jammy-zen-cpu-py3_9-gcc11-inductor-test:
-    name: linux-jammy-zen-cpu-py3.9-gcc11-inductor
+  inductor-test:
+    name: inductor-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-zen-cpu-py3_9-gcc11-inductor-build
+    needs: inductor-build
    if: github.event_name == 'workflow_dispatch'
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}
-      docker-image: ${{ needs.linux-jammy-zen-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-zen-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
      timeout-minutes: 720
      # disable monitor in perf tests
      disable-monitor: false
--- a/.github/workflows/inductor-perf-test-nightly-x86.yml
+++ b/.github/workflows/inductor-perf-test-nightly-x86.yml
@ -74,14 +74,14 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-build:
+    name: inductor-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      test-matrix: |
        { include: [
          { config: "inductor_huggingface_perf_cpu_x86", shard: 1, num_shards: 3, runner: "linux.24xl.spr-metal" },
@ -101,16 +101,16 @@ jobs:
      build-additional-packages: "vision audio torchao"
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-test-nightly-freezing:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-test-nightly-freezing:
+    name: inductor-test-nightly-freezing
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: inductor-build
    if: github.event.schedule == '0 7 * * *'
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
      dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true-freezing-true
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
      timeout-minutes: 720
      # disable monitor in perf tests
      disable-monitor: false
@ -118,16 +118,16 @@ jobs:
      monitor-data-collect-interval: 4
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-test:
+    name: inductor-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: inductor-build
    if: github.event_name == 'workflow_dispatch'
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-freezing-${{ inputs.freezing }}
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
      timeout-minutes: 720
      # disable monitor in perf tests
      disable-monitor: false
--- a/.github/workflows/inductor-perf-test-nightly.yml
+++ b/.github/workflows/inductor-perf-test-nightly.yml
@ -79,7 +79,6 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  # NB: Keep this in sync with trunk.yml
  build:
    name: cuda12.8-py3.10-gcc9-sm80
    uses: ./.github/workflows/_linux-build.yml
--- a/.github/workflows/inductor-periodic.yml
+++ b/.github/workflows/inductor-periodic.yml
@ -31,11 +31,13 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  linux-jammy-cuda12_8-py3_10-gcc9-periodic-dynamo-benchmarks-build:
-    name: cuda12.8-py3.10-gcc9-sm86-periodic-dynamo-benchmarks
+  periodic-dynamo-benchmarks-build:
+    name: periodic-dynamo-benchmarks-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-default-label-prefix
    with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
@ -57,23 +59,33 @@ jobs:
          { config: "dynamic_aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.g5.4xlarge.nvidia.gpu" },
          { config: "dynamic_aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
          { config: "dynamic_aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "dynamic_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "dynamic_inductor_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "dynamic_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "dynamic_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "dynamic_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "aot_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "aot_inductor_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "aot_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "aot_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "aot_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
        ]}
      build-additional-packages: "vision audio fbgemm torchao"
    secrets: inherit

-  linux-jammy-cuda12_8-py3_10-gcc9-periodic-dynamo-benchmarks-test:
-    name: cuda12.8-py3.10-gcc9-sm86-periodic-dynamo-benchmarks
+  periodic-dynamo-benchmarks-test:
+    name: periodic-dynamo-benchmarks-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-periodic-dynamo-benchmarks-build
+    needs: periodic-dynamo-benchmarks-build
    with:
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-periodic-dynamo-benchmarks-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-periodic-dynamo-benchmarks-build.outputs.test-matrix }}
+      docker-image: ${{ needs.periodic-dynamo-benchmarks-build.outputs.docker-image }}
+      test-matrix: ${{ needs.periodic-dynamo-benchmarks-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-rocm-py3_10-periodic-dynamo-benchmarks-build:
+  rocm-periodic-dynamo-benchmarks-build:
    if: github.repository_owner == 'pytorch'
-    name: rocm-py3_10-periodic-dynamo-benchmarks
+    name: rocm-periodic-dynamo-benchmarks-build
    uses: ./.github/workflows/_linux-build.yml
    with:
      build-environment: linux-jammy-rocm-py3_10
@ -99,25 +111,27 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-rocm-py3_10-periodic-dynamo-benchmarks-test:
+  rocm-periodic-dynamo-benchmarks-test:
    permissions:
      id-token: write
      contents: read
-    name: rocm-py3_10-periodic-dynamo-benchmarks
+    name: rocm-periodic-dynamo-benchmarks-test
    uses: ./.github/workflows/_rocm-test.yml
-    needs: linux-jammy-rocm-py3_10-periodic-dynamo-benchmarks-build
+    needs: rocm-periodic-dynamo-benchmarks-build
    with:
      build-environment: linux-jammy-rocm-py3_10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-periodic-dynamo-benchmarks-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-periodic-dynamo-benchmarks-build.outputs.test-matrix }}
+      docker-image: ${{ needs.rocm-periodic-dynamo-benchmarks-build.outputs.docker-image }}
+      test-matrix: ${{ needs.rocm-periodic-dynamo-benchmarks-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-smoke-build:
-    name: cuda12.8-py3.10-gcc9-sm80
+  inductor-smoke-build:
+    name: inductor-smoke-build
    uses: ./.github/workflows/_linux-build.yml
    needs:
      - get-default-label-prefix
    with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
@ -129,23 +143,23 @@ jobs:
      build-additional-packages: "vision audio fbgemm torchao"
    secrets: inherit

-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-smoke-test:
-    name: cuda12.8-py3.10-gcc9-sm80
+  inductor-smoke-test:
+    name: inductor-smoke-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-inductor-smoke-build
+    needs: inductor-smoke-build
    with:
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-smoke-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-smoke-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-smoke-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-smoke-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-periodic-dynamo-benchmarks-build:
-    name: linux-jammy-cpu-py3.9-gcc11-periodic-dynamo-benchmarks
+  periodic-dynamo-benchmarks-cpu-build:
+    name: periodic-dynamo-benchmarks-cpu-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-default-label-prefix
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
      test-matrix: |
        { include: [
@ -160,68 +174,6 @@ jobs:
          { config: "cpu_inductor_freezing_avx2_torchbench", shard: 2, num_shards: 2, runner: "linux.10xlarge.avx2" },
          { config: "cpu_inductor_freezing_avx2_timm", shard: 1, num_shards: 2, runner: "linux.10xlarge.avx2" },
          { config: "cpu_inductor_freezing_avx2_timm", shard: 2, num_shards: 2, runner: "linux.10xlarge.avx2" },
-        ]}
-      build-additional-packages: "vision audio torchao"
-    secrets: inherit
-
-  linux-jammy-cpu-py3_9-gcc11-periodic-dynamo-benchmarks-test:
-    name: linux-jammy-cpu-py3.9-gcc11-periodic-dynamo-benchmarks
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-periodic-dynamo-benchmarks-build
-    with:
-      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-periodic-dynamo-benchmarks-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-periodic-dynamo-benchmarks-build.outputs.test-matrix }}
-    secrets: inherit
-
-
-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-build:
-    name: cuda12.8-py3.10-gcc9-sm86
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-default-label-prefix
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
-      cuda-arch-list: '8.6'
-      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      sync-tag: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
-      test-matrix: |
-        { include: [
-          { config: "dynamic_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "dynamic_inductor_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "dynamic_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "dynamic_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "dynamic_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "aot_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "aot_inductor_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "aot_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "aot_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "aot_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-        ]}
-      build-additional-packages: "vision audio fbgemm torchao"
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
-    name: cuda12.8-py3.10-gcc9-sm86
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.test-matrix }}
-    secrets: inherit
-
-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-default-label-prefix
-    with:
-      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
-      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      sync-tag: linux-jammy-cpu-py3_9-gcc11-inductor-build
-      test-matrix: |
-        { include: [
          { config: "cpu_inductor_freezing_huggingface", shard: 1, num_shards: 1, runner: "linux.8xlarge.amx" },
          { config: "cpu_inductor_freezing_timm", shard: 1, num_shards: 2, runner: "linux.8xlarge.amx" },
          { config: "cpu_inductor_freezing_timm", shard: 2, num_shards: 2, runner: "linux.8xlarge.amx" },
@ -247,12 +199,12 @@ jobs:
      build-additional-packages: "vision audio torchao"
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  periodic-dynamo-benchmarks-cpu-test:
+    name: periodic-dynamo-benchmarks-cpu-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: periodic-dynamo-benchmarks-cpu-build
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.docker-image }}
+      test-matrix: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.test-matrix }}
    secrets: inherit
--- a/.github/workflows/inductor-unittest.yml
+++ b/.github/workflows/inductor-unittest.yml
@ -28,11 +28,13 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-build:
-    name: cuda12.8-py3.10-gcc9-sm86
+  inductor-build:
+    name: inductor-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
      cuda-arch-list: '8.6'
@ -45,46 +47,21 @@ jobs:
          { config: "inductor_cpp_wrapper", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
          { config: "inductor_cpp_wrapper", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
        ]}
+      build-additional-packages: "vision audio torchao"
    secrets: inherit

-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
-    name: cuda12.8-py3.10-gcc9-sm86
+  inductor-test:
+    name: inductor-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
+    needs: inductor-build
    with:
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-cuda12_8-py3_12-gcc9-inductor-build:
-    name: cuda12.8-py3.12-gcc9-sm86
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.12-gcc9-sm86
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks
-      cuda-arch-list: '8.6'
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      test-matrix: |
-        { include: [
-          { config: "inductor", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
-          { config: "inductor", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_12-gcc9-inductor-test:
-    name: cuda12.8-py3.12-gcc9-sm86
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_12-gcc9-inductor-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.12-gcc9-sm86
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_12-gcc9-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_12-gcc9-inductor-build.outputs.test-matrix }}
-    secrets: inherit
-
-  linux-jammy-cpu-py3_12-inductor-halide-build:
-    name: linux-jammy-cpu-py3.12-gcc11-inductor-halide
+  inductor-halide-build:
+    name: inductor-halide-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
@ -97,18 +74,18 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-cpu-py3_12-inductor-halide-test:
-    name: linux-jammy-cpu-py3.12-gcc11-inductor-halide
+  inductor-halide-test:
+    name: inductor-halide-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_12-inductor-halide-build
+    needs: inductor-halide-build
    with:
      build-environment: linux-jammy-py3.12-gcc11
-      docker-image: ${{ needs.linux-jammy-cpu-py3_12-inductor-halide-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_12-inductor-halide-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-halide-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-halide-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-cpu-py3_12-inductor-triton-cpu-build:
-    name: linux-jammy-cpu-py3.12-gcc11-inductor-triton-cpu
+  inductor-triton-cpu-build:
+    name: inductor-triton-cpu-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
@ -121,23 +98,23 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-cpu-py3_12-inductor-triton-cpu-test:
+  inductor-triton-cpu-test:
    name: linux-jammy-cpu-py3.12-gcc11-inductor-triton-cpu
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_12-inductor-triton-cpu-build
+    needs: inductor-triton-cpu-build
    with:
      build-environment: linux-jammy-py3.12-gcc11
-      docker-image: ${{ needs.linux-jammy-cpu-py3_12-inductor-triton-cpu-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_12-inductor-triton-cpu-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-triton-cpu-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-triton-cpu-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-cpu-build:
+    name: inductor-cpu-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      test-matrix: |
        { include: [
@ -148,37 +125,12 @@ jobs:
        ]}
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-cpu-test:
+    name: inductor-cpu-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: inductor-cpu-build
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_13-gcc9-inductor-build:
-    name: cuda12.8-py3.13-gcc9-sm86
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.13-gcc9-sm86
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks
-      cuda-arch-list: '8.6'
-      test-matrix: |
-        { include: [
-          { config: "inductor", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
-          { config: "inductor", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_13-gcc9-inductor-test:
-    name: cuda12.8-py3.13-gcc9-sm86
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_13-gcc9-inductor-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.13-gcc9-sm86
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_13-gcc9-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_13-gcc9-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
    secrets: inherit
--- a/.github/workflows/inductor.yml
+++ b/.github/workflows/inductor.yml
@ -44,16 +44,17 @@ jobs:
      curr_ref_type: ${{ github.ref_type }}
      opt_out_experiments: lf

-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-build:
-    name: cuda12.8-py3.10-gcc9-sm86
+  inductor-build:
+    name: inductor-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
      cuda-arch-list: '8.6'
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      sync-tag: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
      test-matrix: |
        { include: [
          { config: "inductor_huggingface", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
@ -65,25 +66,24 @@ jobs:
      build-additional-packages: "vision audio fbgemm torchao"
    secrets: inherit

-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
-    name: cuda12.8-py3.10-gcc9-sm86
+  inductor-test:
+    name: inductor-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
+    needs: inductor-build
    with:
      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-build:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-cpu-build:
+    name: inductor-cpu-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      sync-tag: linux-jammy-cpu-py3_9-gcc11-inductor-build
      test-matrix: |
        { include: [
          { config: "cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
@ -98,12 +98,12 @@ jobs:
      build-additional-packages: "vision audio torchao"
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-inductor-test:
-    name: linux-jammy-cpu-py3.9-gcc11-inductor
+  inductor-cpu-test:
+    name: inductor-cpu-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-inductor-build
+    needs: inductor-cpu-build
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-inductor-build.outputs.test-matrix }}
+      docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
+      test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
    secrets: inherit
--- a/.github/workflows/operator_benchmark.yml
+++ b/.github/workflows/operator_benchmark.yml
@ -24,38 +24,38 @@ permissions:
  contents: read

 jobs:
-  linux-jammy-cpu-py3_9-gcc11-opbenchmark-build:
+  opbenchmark-build:
    if: github.repository_owner == 'pytorch'
-    name: linux-jammy-cpu-py3.9-gcc11-opbenchmark
+    name: opbenchmark-build
    uses: ./.github/workflows/_linux-build.yml
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      test-matrix: |
        { include: [
          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
        ]}
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-opbenchmark-on-demand-build:
+  opbenchmark-on-demand-build:
    if: ${{ github.event_name == 'workflow_dispatch' && github.repository_owner == 'pytorch' }}
-    name: linux-jammy-cpu-py3.9-gcc11-opbenchmark
+    name: opbenchmark-on-demand-build
    uses: ./.github/workflows/_linux-build.yml
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      test-matrix: |
        { include: [
          { config: "cpu_operator_benchmark_${{ inputs.test_mode }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
        ]}
    secrets: inherit

-  linux-jammy-cpu-py3_9-gcc11-opbenchmark-test:
-    name: linux-jammy-cpu-py3.9-gcc11-opbenchmark
+  opbenchmark-test:
+    name: opbenchmark-test
    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cpu-py3_9-gcc11-opbenchmark-build
+    needs: opbenchmark-build
    with:
      build-environment: linux-jammy-py3.9-gcc11-build
-      docker-image: ${{ needs.linux-jammy-cpu-py3_9-gcc11-opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cpu-py3_9-gcc11-opbenchmark-build.outputs.test-matrix }}
+      docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
    secrets: inherit
--- a/.github/workflows/periodic.yml
+++ b/.github/workflows/periodic.yml
@ -170,6 +170,38 @@ jobs:
      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-debug-build.outputs.test-matrix }}
    secrets: inherit

+  linux-jammy-cuda13_0-py3_10-gcc11-build:
+    name: linux-jammy-cuda13.0-py3.10-gcc11
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      cuda-arch-list: 7.5
+      build-environment: linux-jammy-cuda13.0-py3.10-gcc11
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11
+      test-matrix: |
+        { include: [
+          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-cuda13_0-py3_10-gcc11-test:
+    name: linux-jammy-cuda13.0-py3.10-gcc11
+    uses: ./.github/workflows/_linux-test.yml
+    needs:
+      - linux-jammy-cuda13_0-py3_10-gcc11-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-cuda13.0-py3.10-gcc11
+      docker-image: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.test-matrix }}
+    secrets: inherit
+
  linux-jammy-rocm-py3_10-build:
    name: linux-jammy-rocm-py3.10
    uses: ./.github/workflows/_linux-build.yml
--- a/.github/workflows/pull.yml
+++ b/.github/workflows/pull.yml
@ -132,17 +132,17 @@ jobs:
      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
      test-matrix: |
        { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "default", shard: 2, num_shards: 6, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "default", shard: 3, num_shards: 6, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "default", shard: 4, num_shards: 6, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "default", shard: 5, num_shards: 6, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "default", shard: 6, num_shards: 6, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "default", shard: 1, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "default", shard: 2, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "default", shard: 3, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "default", shard: 4, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "default", shard: 5, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "default", shard: 6, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "default", shard: 7, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
        ]}
      sync-tag: asan-build
    secrets: inherit

-
  linux-jammy-py3_10-clang18-asan-test:
    name: linux-jammy-py3.10-clang18-asan
    uses: ./.github/workflows/_linux-test.yml
--- a/.github/workflows/rocm-mi300.yml
+++ b/.github/workflows/rocm-mi300.yml
@ -70,4 +70,5 @@ jobs:
      build-environment: linux-noble-rocm-py3.12-mi300
      docker-image: ${{ needs.linux-noble-rocm-py3_12-build.outputs.docker-image }}
      test-matrix: ${{ needs.linux-noble-rocm-py3_12-build.outputs.test-matrix }}
+      tests-to-include: "inductor/test_ck_backend"
    secrets: inherit
--- a/.github/workflows/trunk.yml
+++ b/.github/workflows/trunk.yml
@ -224,13 +224,12 @@ jobs:
      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

-  # NB: Keep this in sync with inductor-perf-test-nightly.yml
-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-build:
-    name: cuda12.8-py3.10-gcc9-sm80
+  inductor-build:
+    name: inductor-build
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.12-gcc9-sm80
      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
      cuda-arch-list: '8.0'
    secrets: inherit
@ -240,13 +239,16 @@ jobs:
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build-environment: linux-jammy-py3.9-gcc11
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
      test-matrix: |
        { include: [
          { config: "verify_cachebench", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
        ]}
+      build-additional-packages: "vision audio torchao"
    secrets: inherit

  verify-cachebench-cpu-test:
--- 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-binary\", \"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 }}
--- a/.github/workflows/vllm.yml
+++ b/.github/workflows/vllm.yml
@ -2,6 +2,9 @@ name: vllm-test

 on:
  push:
+    branches:
+      - main
+      - release/*
    tags:
      - ciflow/vllm/*
  workflow_dispatch:
@ -45,14 +48,18 @@ jobs:
          { config: "vllm_basic_models_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_entrypoints_test", shard: 1, num_shards: 1,runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_regression_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
-          { config: "vllm_lora_280_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_multi_model_processor_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_pytorch_compilation_unit_tests", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
+          { config: "vllm_lora_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
+          { config: "vllm_multi_model_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
+          { config: "vllm_languagde_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
+          { config: "vllm_distributed_test_2_gpu_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_lora_test", shard: 0, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_lora_test", shard: 1, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_lora_test", shard: 2, num_shards: 4,  runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
          { config: "vllm_lora_test", shard: 3, num_shards: 4,  runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
-          { config: "vllm_lora_tp_test_distributed", shard: 1, num_shards: 1, runner: "linux.aws.h100.4"},
+          { config: "vllm_lora_tp_test_distributed", shard: 1, num_shards: 1, runner: "linux.g6.12xlarge.nvidia.gpu"},
+          { config: "vllm_distributed_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.12xlarge.nvidia.gpu"}
        ]}
    secrets: inherit

--- a/.gitignore
+++ b/.gitignore
@ -259,6 +259,9 @@ gen
 .pytest_cache
 aten/build/*

+# Linker scripts for prioritized text optimization
+cmake/linker_script.ld
+
 # Bram
 plsdontbreak

--- a/CLAUDE.md
+++ b/CLAUDE.md
@ -0,0 +1,15 @@
+# Testing
+
+Use our test class and test runner:
+
+```
+from torch.testing._internal.common_utils import run_tests, TestCase
+
+class TestFeature(TestCase):
+    ...
+
+if __name__ == "__main__":
+    run_tests()
+```
+
+To test Tensor equality, use assertEqual.
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -233,6 +233,7 @@ cmake_dependent_option(INSTALL_TEST "Install test binaries if BUILD_TEST is on"
 option(USE_CPP_CODE_COVERAGE "Compile C/C++ with code coverage flags" OFF)
 option(USE_COLORIZE_OUTPUT "Colorize output during compilation" ON)
 option(USE_ASAN "Use Address+Undefined Sanitizers" OFF)
+option(USE_LSAN "Use Leak Sanitizer" OFF)
 option(USE_TSAN "Use Thread Sanitizer" OFF)
 option(USE_CUDA "Use CUDA" ON)
 option(USE_XPU "Use XPU" ON)
@ -378,6 +379,13 @@ cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
                       OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                        "CPU_AARCH64" OFF)
+# prioritized text linker, ON by default for AArch64+Linux, option visible to all AArch64, x86 and ppc64le.
+set(USE_PRIORITIZED_TEXT_DEFAULT OFF)
+if(LINUX AND CPU_AARCH64)
+  set(USE_PRIORITIZED_TEXT_DEFAULT ON)
+endif()
+cmake_dependent_option(USE_PRIORITIZED_TEXT_FOR_LD "Use prioritized text linker for ld."
+  "${USE_PRIORITIZED_TEXT_DEFAULT}" "CPU_INTEL OR CPU_AARCH64 OR CPU_POWER" OFF)

 option(USE_MIMALLOC "Use mimalloc" OFF)
 # Enable third party mimalloc library to improve memory allocation performance
@ -656,6 +664,11 @@ endif(MSVC)

 string(APPEND CMAKE_CUDA_FLAGS " -Xfatbin -compress-all")

+# Set linker max-page-size to 64KiB on AArch64 Linux
+if(LINUX AND CPU_AARCH64)
+  add_link_options_if_supported("-z,max-page-size=0x10000")
+endif()
+
 # Set INTERN_BUILD_MOBILE for all mobile builds. Components that are not
 # applicable to mobile are disabled by this variable. Setting
 # `BUILD_PYTORCH_MOBILE_WITH_HOST_TOOLCHAIN` environment variable can force it
@ -880,10 +893,21 @@ cmake_dependent_option(
  USE_FBGEMM_GENAI
  "Whether to build FBGEMM GenAI quantized GEMM kernels.\
  Will be disabled if not supported by the platform"
-  OFF
-  "USE_CUDA OR USE_ROCM"
+  ON
+  "USE_ROCM"
  OFF)

+IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
+  message(WARNING "Unsupported ROCM arch for FBGEMM GenAI, will set USE_FBGEMM_GENAI to OFF")
+  set(USE_FBGEMM_GENAI off)
+endif()
+
+# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100.
+if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0" AND CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
+  message(STATUS "Setting USE_FBGEMM_GENAI to ON, doing CUDA build for SM100a")
+  set(USE_FBGEMM_GENAI ON)
+endif()
+
 # CAVEAT: Again, Flash Attention2 will error while building for sm52 while Mem
 # Eff Attention won't
 cmake_dependent_option(
@ -1409,3 +1433,57 @@ if(BUILD_BUNDLE_PTXAS AND USE_CUDA)
  install(PROGRAMS "${PROJECT_BINARY_DIR}/ptxas"
          DESTINATION "${CMAKE_INSTALL_BINDIR}")
 endif()
+
+if(USE_PRIORITIZED_TEXT_FOR_LD)
+  add_compile_options(
+    $<$<COMPILE_LANGUAGE:C,CXX>:-ffunction-sections>
+    $<$<COMPILE_LANGUAGE:C,CXX>:-fdata-sections>
+  )
+  set(LINKER_SCRIPT_FILE_OUT "${CMAKE_SOURCE_DIR}/cmake/linker_script.ld")
+  set(LINKER_SCRIPT_FILE_IN "${CMAKE_SOURCE_DIR}/cmake/prioritized_text.txt")
+
+  add_custom_command(
+    OUTPUT "${LINKER_SCRIPT_FILE_OUT}"
+    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py --filein "${LINKER_SCRIPT_FILE_IN}" --fout "${LINKER_SCRIPT_FILE_OUT}"
+    DEPENDS ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py "${LINKER_SCRIPT_FILE_IN}"
+    COMMENT "Generating prioritized text linker files"
+    VERBATIM
+  )
+
+  add_custom_target(generate_linker_script DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
+
+  if(BUILD_PYTHON)
+    set(LINKER_OPT_TARGETS torch_python)
+  endif()
+
+  if(NOT BUILD_LIBTORCHLESS)
+    list(APPEND LINKER_OPT_TARGETS torch_cpu c10)
+    if(USE_CUDA)
+      list(APPEND LINKER_OPT_TARGETS torch_cuda c10_cuda)
+    endif()
+    if(USE_XPU)
+      list(APPEND LINKER_OPT_TARGETS torch_xpu c10_xpu)
+    endif()
+    if(USE_ROCM)
+      list(APPEND LINKER_OPT_TARGETS torch_hip c10_hip)
+    endif()
+  endif()
+
+  foreach(tgt IN LISTS LINKER_OPT_TARGETS)
+    if(TARGET ${tgt})
+      add_dependencies("${tgt}" generate_linker_script)
+      target_link_options_if_supported(${tgt} "-T,${LINKER_SCRIPT_FILE_OUT}")
+      set_property(TARGET ${tgt} APPEND PROPERTY LINK_DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
+    else()
+       message(WARNING "Requested target '${tgt}' for linker script optimization was not found.")
+    endif()
+  endforeach()
+
+else()
+  if(LINUX AND CPU_AARCH64)
+    message(WARNING [[
+    It is strongly recommend to enable linker script optimization for all AArch64 Linux builds.
+    To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
+    ]])
+  endif()
+endif()
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@ -88,13 +88,13 @@ source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows

 * If you want to have no-op incremental rebuilds (which are fast), see [Make no-op build fast](#make-no-op-build-fast) below.

-* When installing with `python -m pip install -e .` (in contrast to `python -m pip install .`) Python runtime will use
+* When installing with `python -m pip install -e . -v --no-build-isolation` (in contrast to `python -m pip install . -v --no-build-isolation`) Python runtime will use
  the current local source-tree when importing `torch` package. (This is done by creating [`.egg-link`](https://wiki.python.org/moin/PythonPackagingTerminology#egg-link) file in `site-packages` folder)
  This way you do not need to repeatedly install after modifying Python files (`.py`).
  However, you would need to reinstall if you modify Python interface (`.pyi`, `.pyi.in`) or non-Python files (`.cpp`, `.cc`, `.cu`, `.h`, ...).


-  One way to avoid running `python -m pip install -e .` every time one makes a change to C++/CUDA/ObjectiveC files on Linux/Mac,
+  One way to avoid running `python -m pip install -e . -v --no-build-isolation` every time one makes a change to C++/CUDA/ObjectiveC files on Linux/Mac,
  is to create a symbolic link from `build` folder to `torch/lib`, for example, by issuing following:
  ```bash
  pushd torch/lib; sh -c "ln -sf ../../build/lib/libtorch_cpu.* ."; popd
@ -116,7 +116,7 @@ source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows

  Next run `python setup.py clean`. After that, you can install in editable mode again.

-* If you run into errors when running `python -m pip install -e .`, here are some debugging steps:
+* If you run into errors when running `python -m pip install -e . -v --no-build-isolation`, here are some debugging steps:
  1. Run `printf '#include <stdio.h>\nint main() { printf("Hello World");}'|clang -x c -; ./a.out` to make sure
  your CMake works and can compile this simple Hello World program without errors.
  2. Nuke your `build` directory. The `setup.py` script compiles binaries into the `build` folder and caches many
@ -129,10 +129,10 @@ source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
      git clean -xdf
      python setup.py clean
      git submodule update --init --recursive
-      python -m pip install -r requirements.txt
+      python -m pip install --group dev
      python -m pip install --no-build-isolation -v -e .
      ```
-  4. The main step within `python -m pip install -e .` is running `cmake --build build` from the `build` directory. If you want to
+  4. The main step within `python -m pip install -e . -v --no-build-isolation` is running `make` from the `build` directory. If you want to
    experiment with some environment variables, you can pass them into the command:
      ```bash
      ENV_KEY1=ENV_VAL1[, ENV_KEY2=ENV_VAL2]* CMAKE_FRESH=1 python -m pip install --no-build-isolation -v -e .
@ -259,6 +259,7 @@ dependencies as well as the nightly binaries into the repo directory.
      support for PyTorch.
 * [tools](tools) - Code generation scripts for the PyTorch library.
  See [README](tools/README.md) of this directory for more details.
+* [torchgen](torchgen) - contains the logic and tooling for generating PyTorch's low-level C++ and Python bindings from operator definitions, typically specified in native_functions.yaml
 * [test](test) - Python unit tests for PyTorch Python frontend.
  * [test_torch.py](test/test_torch.py) - Basic tests for PyTorch
    functionality.
@ -294,7 +295,7 @@ The following packages should be installed with `pip`:
 - `pytest` - recommended to run tests more selectively
 Running
 ```
-pip install -r requirements.txt
+pip install --group dev
 ```
 will install these dependencies for you.

@ -645,9 +646,9 @@ can be selected interactively with your mouse to zoom in on a particular part of
 the program execution timeline. The `--native` command-line option tells
 `py-spy` to record stack frame entries for PyTorch C++ code. To get line numbers
 for C++ code it may be necessary to compile PyTorch in debug mode by prepending
-your `python -m pip install -e .` call to compile PyTorch with `DEBUG=1`.
-Depending on your operating system it may also be necessary to run `py-spy` with
-root privileges.
+your `python -m pip install -e . -v --no-build-isolation` call to compile
+PyTorch with `DEBUG=1`. Depending on your operating system it may also be
+necessary to run `py-spy` with root privileges.

 `py-spy` can also work in an `htop`-like "live profiling" mode and can be
 tweaked to adjust the stack sampling rate, see the `py-spy` readme for more
@ -655,10 +656,10 @@ details.

 ## Managing multiple build trees

-One downside to using `python -m pip install -e .` is that your development
-version of PyTorch will be installed globally on your account (e.g., if
-you run `import torch` anywhere else, the development version will be
-used).
+One downside to using `python -m pip install -e . -v --no-build-isolation` is
+that your development version of PyTorch will be installed globally on your
+account (e.g., if you run `import torch` anywhere else, the development version
+will be used).

 If you want to manage multiple builds of PyTorch, you can make use of
 [venv environments](https://docs.python.org/3/library/venv.html) to maintain
@ -719,7 +720,7 @@ options.

 ### Code completion and IDE support

-When using `python -m pip install -e .`, PyTorch will generate
+When using `python -m pip install -e . -v --no-build-isolation`, PyTorch will generate
 a `compile_commands.json` file that can be used by many editors
 to provide command completion and error highlighting for PyTorch's
 C++ code. You need to `pip install ninja` to generate accurate
--- a/README.md
+++ b/README.md
@ -243,7 +243,7 @@ git submodule update --init --recursive

 ```bash
 # Run this command from the PyTorch directory after cloning the source code using the “Get the PyTorch Source“ section above
-pip install -r requirements.txt
+pip install --group dev
 ```

 **On Linux**
@ -394,7 +394,7 @@ On macOS

 ```bash
 export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
-MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ CMAKE_ONLY=1 python setup.py build
+MACOSX_DEPLOYMENT_TARGET=11.0 CMAKE_ONLY=1 python setup.py build
 ccmake build  # or cmake-gui build
 ```

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

 | PyTorch version | Python | C++ | Stable CUDA | Experimental CUDA | Stable ROCm |
 | --- | --- | --- | --- | --- | --- |
+| 2.9 | >=3.10, <=(3.14, 3.14t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 13.0 (CUDNN 9.13.0.50) | ROCm 6.4 |
 | 2.8 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 12.9 (CUDNN 9.10.2.21) | ROCm 6.4 |
 | 2.7 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8 (CUDNN 9.1.0.70), CUDA 12.6 (CUDNN 9.5.1.17) | CUDA 12.8 (CUDNN 9.7.1.26) | ROCm 6.3 |
 | 2.6 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8, CUDA 12.4 (CUDNN 9.1.0.70) | CUDA 12.6 (CUDNN 9.5.1.17) | ROCm 6.2.4 |
--- a/aten/src/ATen/CMakeLists.txt
+++ b/aten/src/ATen/CMakeLists.txt
@ -216,7 +216,7 @@ file(GLOB mem_eff_attention_cuda_cpp "native/transformers/cuda/mem_eff_attention
 if(USE_CUDA AND (USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION))
  add_library(flash_attention OBJECT EXCLUDE_FROM_ALL ${flash_attention_cuda_kernels_cu} ${flash_attention_cuda_cpp})

-  target_include_directories(flash_attention PUBLIC
+  target_include_directories(flash_attention SYSTEM PUBLIC
    ${PROJECT_SOURCE_DIR}/third_party/flash-attention/csrc
    ${PROJECT_SOURCE_DIR}/third_party/flash-attention/include
    ${PROJECT_SOURCE_DIR}/third_party/cutlass/include
@ -252,47 +252,88 @@ if(USE_MEM_EFF_ATTENTION)
  list(APPEND ATen_ATTENTION_KERNEL_SRCS ${mem_eff_attention_cuda_kernels_cu})
 endif()

-IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-  message(WARNING "Unsupported ROCM arch for FBGEMM GenAI, will set USE_FBGEMM_GENAI to OFF")
-  set(USE_FBGEMM_GENAI off)
-endif()
-
 # FBGEMM GenAI
 IF(USE_FBGEMM_GENAI)
  set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
-  set(FBGEMM_GENAI_DIR ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
+  set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
+  if(USE_CUDA)
+    # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
+    # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
+    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
+    file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
+      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
+      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
+    list(FILTER fbgemm_genai_native_cuda_cu INCLUDE REGEX ${FBGEMM_CUTLASS_KERNELS_REGEX})

-  if(USE_ROCM)
-    # Only include the kernels we want to build to avoid increasing binary size.
-    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-      "${FBGEMM_GENAI_DIR}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-      "${FBGEMM_GENAI_DIR}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
+    # PyTorch is not built for 10.0a in CI, due to lack of portability,
+    # so we need to explicitly build these files for 10.0a.
+    foreach(cu_file ${fbgemm_genai_native_cuda_cu})
+      _BUILD_FOR_ADDITIONAL_ARCHS(
+        "${cu_file}"
+        "100a")
+    endforeach()

-    # Add additional HIPCC compiler flags for performance
-    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-      -mllvm
-      -amdgpu-coerce-illegal-types=1
-      -mllvm
-      -enable-post-misched=0
-      -mllvm
-      -greedy-reverse-local-assignment=1
-      -fhip-new-launch-api)
+    file(GLOB_RECURSE fbgemm_genai_native_cuda_cpp
+      "${FBGEMM_GENAI_SRCS}/common/*.cpp"
+    )
+
+    # Combine all source files into a single list
+    list(APPEND fbgemm_genai_all_sources
+      ${fbgemm_genai_native_cuda_cu}
+      ${fbgemm_genai_native_cuda_cpp}
+    )
+
+    # Now, create the library and provide the sources at the same time
+    add_library(fbgemm_genai OBJECT ${fbgemm_genai_all_sources})

-    hip_add_library(
-      fbgemm_genai STATIC
-      ${fbgemm_genai_native_rocm_hip}
-      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
    set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
+
+    set(fbgemm_genai_mx8mx8bf16_grouped
+      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
+    )

    target_include_directories(fbgemm_genai PUBLIC
-      # FBGEMM version of Composable Kernel is used due to some customizations
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-      ${FBGEMM_GENAI_DIR}/include/
-      ${FBGEMM_GENAI_DIR}/common/include/
+      ${FBGEMM_THIRD_PARTY}/cutlass/include
+      ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
+      ${fbgemm_genai_mx8mx8bf16_grouped}
+      ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
+      ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
    )
+  else()
+    if(USE_ROCM)
+      # Only include the kernels we want to build to avoid increasing binary size.
+      file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
+        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
+        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
+      set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
+
+      # Add additional HIPCC compiler flags for performance
+      set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+        -mllvm
+        -amdgpu-coerce-illegal-types=1
+        -mllvm
+        -enable-post-misched=0
+        -mllvm
+        -greedy-reverse-local-assignment=1
+        -fhip-new-launch-api)
+
+      hip_add_library(
+        fbgemm_genai STATIC
+        ${fbgemm_genai_native_rocm_hip}
+        HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
+      set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
+      target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
+
+      target_include_directories(fbgemm_genai PUBLIC
+        # FBGEMM version of Composable Kernel is used due to some customizations
+        ${FBGEMM_THIRD_PARTY}/composable_kernel/include
+        ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
+        ${FBGEMM_THIRD_PARTY}/cutlass/include
+        ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
+        ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
+        ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
+      )
+    endif()
  endif()
 endif()

@ -635,12 +676,26 @@ if(USE_CUDA AND NOT USE_ROCM)
  add_definitions(-DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED)
  list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
  list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
+
+  # Add FBGEMM_GENAI include directories for torch_ops.h
+  if(USE_FBGEMM_GENAI)
+    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
+  endif()
+
  if($ENV{ATEN_STATIC_CUDA})
-    list(APPEND ATen_CUDA_DEPENDENCY_LIBS
-      ${CUDA_LIBRARIES}
-      CUDA::cusparse_static
-      CUDA::cufft_static_nocallback
-    )
+    if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
+      list(APPEND ATen_CUDA_DEPENDENCY_LIBS
+          ${CUDA_LIBRARIES}
+          CUDA::cusparse_static
+          CUDA::cufft_static_nocallback)
+    else()
+      list(APPEND ATen_CUDA_DEPENDENCY_LIBS
+          ${CUDA_LIBRARIES}
+          CUDA::cusparse_static
+          CUDA::cufft_static)
+    endif()
+
   if(NOT BUILD_LAZY_CUDA_LINALG)
     list(APPEND ATen_CUDA_DEPENDENCY_LIBS
       CUDA::cusolver_static
--- a/aten/src/ATen/Context.cpp
+++ b/aten/src/ATen/Context.cpp
@ -279,10 +279,45 @@ 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"};
 #ifdef USE_ROCM
 static constexpr const auto hipblaslt_allow_tf32 = "HIPBLASLT_ALLOW_TF32";
 #endif

+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;
 }
--- a/aten/src/ATen/Context.h
+++ b/aten/src/ATen/Context.h
@ -310,7 +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
@ -334,6 +340,12 @@ 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(
      const std::string& backend,
@ -417,6 +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;
--- a/aten/src/ATen/DLConvertor.cpp
+++ b/aten/src/ATen/DLConvertor.cpp
@ -308,17 +308,44 @@ void fillVersion<DLManagedTensorVersioned>(
 // constructed out of ATen tensor
 template <class T>
 T* toDLPackImpl(const Tensor& src) {
-  // create a new tensor with possibly normalized strides
-  // gh-83069
-  auto shape = src.sizes();
-  auto strides = src.strides().vec();
-  for (int i = 0; i < src.dim(); i++) {
-    if (shape[i] < 2) {
-      strides[i] = 1;
+  auto view = src;
+
+  // Detect whether there is need to normalize the strides
+  // Background: gh-83069
+  //
+  // However, normalizing strides can come at a high-cost
+  // to slow down toDLPack conversion 3x, so we
+  // only normalize if needed.
+  //
+  // The following code detects whether the src follows
+  // a continuous pattern. If the src follows such pattern (common-case)
+  // then we do not need to normalize the strides.
+  bool need_normalize_strides = false;
+  int64_t expected_stride = 1;
+  for (int i = src.dim() - 1; i >= 0; i--) {
+    // detect if we do not meet continuous pattern
+    // and the size is 1, so there is opportunity to normalize
+    if (src.stride(i) != expected_stride && src.size(i) == 1) {
+      need_normalize_strides = true;
+      break;
    }
+    expected_stride *= src.size(i);
+  }
+
+  // less common case, try normalizing the strides
+  if (need_normalize_strides) {
+    // create a new tensor with possibly normalized strides
+    // gh-83069
+    auto shape = src.sizes();
+    auto strides = src.strides().vec();
+    for (int i = 0; i < src.dim(); i++) {
+      if (shape[i] < 2) {
+        strides[i] = 1;
+      }
+    }
+    view = src.as_strided(shape, strides, src.storage_offset());
  }

-  auto view = src.as_strided(shape, strides, src.storage_offset());
  ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);
  atDLMTensor->handle = view;
  atDLMTensor->tensor.manager_ctx = atDLMTensor;
--- a/aten/src/ATen/DTensorState.cpp
+++ b/aten/src/ATen/DTensorState.cpp
@ -0,0 +1,17 @@
+#include <ATen/DTensorState.h>
+
+namespace at {
+
+namespace {
+thread_local bool kDTensorAllowImplicitReplication = false;
+}
+
+bool get_dtensor_allow_implicit_replication() {
+  return kDTensorAllowImplicitReplication;
+}
+
+void set_dtensor_allow_implicit_replication(bool enabled) {
+  kDTensorAllowImplicitReplication = enabled;
+}
+
+} // namespace at
--- a/aten/src/ATen/DTensorState.h
+++ b/aten/src/ATen/DTensorState.h
@ -0,0 +1,34 @@
+#pragma once
+
+#include <c10/macros/Macros.h>
+
+namespace at {
+
+TORCH_API bool get_dtensor_allow_implicit_replication();
+TORCH_API void set_dtensor_allow_implicit_replication(bool enabled);
+
+struct DTensorAllowImplicitReplication {
+  DTensorAllowImplicitReplication()
+      : prev_dtensor_allow_implicit_replication_(
+            get_dtensor_allow_implicit_replication()) {
+    set_dtensor_allow_implicit_replication(true);
+  }
+
+  DTensorAllowImplicitReplication(const DTensorAllowImplicitReplication&) =
+      delete;
+  DTensorAllowImplicitReplication& operator=(
+      const DTensorAllowImplicitReplication&) = delete;
+  DTensorAllowImplicitReplication(DTensorAllowImplicitReplication&&) = delete;
+  DTensorAllowImplicitReplication& operator=(
+      DTensorAllowImplicitReplication&&) = delete;
+
+  ~DTensorAllowImplicitReplication() {
+    set_dtensor_allow_implicit_replication(
+        prev_dtensor_allow_implicit_replication_);
+  }
+
+ private:
+  bool prev_dtensor_allow_implicit_replication_;
+};
+
+} // namespace at
--- a/Show More
+++ b/Show More
 @ -1 +1 @@
 .4.0
 .5.0