Update

[ghstack-poisoned]

.ci/aarch64_linux/README.md

@@ -0,0 +1,19 @@
+# Aarch64 (ARM/Graviton) Support Scripts
+Scripts for building aarch64 PyTorch PIP Wheels. These scripts build the following wheels:
+* torch
+* torchvision
+* torchaudio
+* torchtext
+* torchdata
+## Aarch64_ci_build.sh
+This script is design to support CD operations within PyPi manylinux aarch64 container, and be executed in the container. It prepares the container and then executes __aarch64_wheel_ci_build.py__ to build the wheels. The script "assumes" the PyTorch repo is located at: ```/pytorch``` and will put the wheels into ```/artifacts```.
+### Usage
+```DESIRED_PYTHON=<PythonVersion> aarch64_ci_build.sh```
+
+__NOTE:__ CI build is currently __EXPERMINTAL__
+
+## Build_aarch64_wheel.py
+This app allows a person to build using AWS EC3 resources and requires AWS-CLI and Boto3 with AWS credentials to support building EC2 instances for the wheel builds. Can be used in a codebuild CD or from a local system.
+
+### Usage
+```build_aarch64_wheel.py --key-name <YourPemKey> --use-docker --python 3.8 --branch <RCtag>```

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+set -eux -o pipefail
+
+GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
+
+# Set CUDA architecture lists to match x86 build_cuda.sh
+if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
+fi
+
+# Compress the fatbin with -compress-mode=size for CUDA 13
+if [[ "$DESIRED_CUDA" == *"13"* ]]; then
+    export TORCH_NVCC_FLAGS="-compress-mode=size"
+    # Bundle ptxas into the cu13 wheel, see https://github.com/pytorch/pytorch/issues/163801
+    export BUILD_BUNDLE_PTXAS=1
+fi
+
+SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
+source $SCRIPTPATH/aarch64_ci_setup.sh
+
+###############################################################################
+# Run aarch64 builder python
+###############################################################################
+cd /
+# adding safe directory for git as the permissions will be
+# on the mounted pytorch repo
+git config --global --add safe.directory /pytorch
+pip install -r /pytorch/requirements.txt
+pip install auditwheel==6.2.0 wheel
+if [ "$DESIRED_CUDA" = "cpu" ]; then
+    echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
+    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
+else
+    echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
+    export USE_SYSTEM_NCCL=1
+
+    # Check if we should use NVIDIA libs from PyPI (similar to x86 build_cuda.sh logic)
+    if [[ -z "$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

.ci/aarch64_linux/aarch64_ci_setup.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+set -eux -o pipefail
+
+# This script is used to prepare the Docker container for aarch64_ci_wheel_build.py python script
+# By creating symlinks from desired /opt/python to /usr/local/bin/
+
+NUMPY_VERSION=2.0.2
+if [[ "$DESIRED_PYTHON"  == "3.13" || "$DESIRED_PYTHON" == "3.13t" ]]; then
+    NUMPY_VERSION=2.1.2
+fi
+
+SCRIPTPATH="$( cd "$(dirname "$0")" ; pwd -P )"
+source $SCRIPTPATH/../manywheel/set_desired_python.sh
+
+pip install -q numpy==${NUMPY_VERSION} pyyaml==6.0.2 scons==4.7.0 ninja==1.11.1 patchelf==0.17.2
+
+for tool in python python3 pip pip3 ninja scons patchelf; do
+    ln -sf ${DESIRED_PYTHON_BIN_DIR}/${tool} /usr/local/bin;
+done
+
+python --version

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -0,0 +1,333 @@
+#!/usr/bin/env python3
+# encoding: UTF-8
+
+import os
+import shutil
+from subprocess import check_call, check_output
+
+
+def list_dir(path: str) -> list[str]:
+    """'
+    Helper for getting paths for Python
+    """
+    return check_output(["ls", "-1", path]).decode().split("\n")
+
+
+def replace_tag(filename) -> None:
+    with open(filename) as f:
+        lines = f.readlines()
+    for i, line in enumerate(lines):
+        if line.startswith("Tag:"):
+            lines[i] = line.replace("-linux_", "-manylinux_2_28_")
+            print(f"Updated tag from {line} to {lines[i]}")
+            break
+
+    with open(filename, "w") as f:
+        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)
+    os.mkdir(f"{folder}/tmp")
+    os.system(f"unzip {wheel_path} -d {folder}/tmp")
+    # Delete original wheel since it will be repackaged
+    os.system(f"rm {wheel_path}")
+
+    # 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",
+        ]
+
+        # 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)
+
+        # 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/"):
+        if f.is_dir() and f.name.endswith(".dist-info"):
+            replace_tag(f"{f.path}/WHEEL")
+            break
+
+    os.system(f"wheel pack {folder}/tmp/ -d {folder}")
+    os.system(f"rm -rf {folder}/tmp/")
+
+
+def complete_wheel(folder: str) -> str:
+    """
+    Complete wheel build and put in artifact location
+    """
+    wheel_name = list_dir(f"/{folder}/dist")[0]
+
+    # Please note for cuda we don't run auditwheel since we use custom script to package
+    # the cuda dependencies to the wheel file using update_wheel() method.
+    # However we need to make sure filename reflects the correct Manylinux platform.
+    if "pytorch" in folder and not enable_cuda:
+        print("Repairing Wheel with AuditWheel")
+        check_call(["auditwheel", "repair", f"dist/{wheel_name}"], cwd=folder)
+        repaired_wheel_name = list_dir(f"/{folder}/wheelhouse")[0]
+
+        print(f"Moving {repaired_wheel_name} wheel to /{folder}/dist")
+        os.rename(
+            f"/{folder}/wheelhouse/{repaired_wheel_name}",
+            f"/{folder}/dist/{repaired_wheel_name}",
+        )
+    else:
+        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
+
+    print(f"Copying {repaired_wheel_name} to artifacts")
+    shutil.copy2(
+        f"/{folder}/dist/{repaired_wheel_name}", f"/artifacts/{repaired_wheel_name}"
+    )
+
+    return repaired_wheel_name
+
+
+def parse_arguments():
+    """
+    Parse inline arguments
+    """
+    from argparse import ArgumentParser
+
+    parser = ArgumentParser("AARCH64 wheels python CD")
+    parser.add_argument("--debug", action="store_true")
+    parser.add_argument("--build-only", action="store_true")
+    parser.add_argument("--test-only", type=str)
+    parser.add_argument("--enable-mkldnn", action="store_true")
+    parser.add_argument("--enable-cuda", action="store_true")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    """
+    Entry Point
+    """
+    args = parse_arguments()
+    enable_mkldnn = args.enable_mkldnn
+    enable_cuda = args.enable_cuda
+    branch = check_output(
+        ["git", "rev-parse", "--abbrev-ref", "HEAD"], cwd="/pytorch"
+    ).decode()
+
+    print("Building PyTorch wheel")
+    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:
+        version = override_package_version
+        build_vars += (
+            f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={version} PYTORCH_BUILD_NUMBER=1 "
+        )
+    elif branch in ["nightly", "main"]:
+        build_date = (
+            check_output(["git", "log", "--pretty=format:%cs", "-1"], cwd="/pytorch")
+            .decode()
+            .replace("-", "")
+        )
+        version = (
+            check_output(["cat", "version.txt"], cwd="/pytorch").decode().strip()[:-2]
+        )
+        if enable_cuda:
+            build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={version}.dev{build_date}+{desired_cuda} PYTORCH_BUILD_NUMBER=1 "
+        else:
+            build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={version}.dev{build_date} PYTORCH_BUILD_NUMBER=1 "
+    elif branch.startswith(("v1.", "v2.")):
+        build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={branch[1 : branch.find('-')]} PYTORCH_BUILD_NUMBER=1 "
+
+    if enable_mkldnn:
+        print("build pytorch with mkldnn+acl backend")
+        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=/opt/OpenBLAS "
+    else:
+        print("build pytorch without mkldnn backend")
+
+    os.system(f"cd /pytorch; {build_vars} python3 -m build --wheel --no-isolation")
+    if enable_cuda:
+        print("Updating Cuda Dependency")
+        filename = os.listdir("/pytorch/dist/")
+        wheel_path = f"/pytorch/dist/{filename[0]}"
+        package_cuda_wheel(wheel_path, desired_cuda)
+    pytorch_wheel_name = complete_wheel("/pytorch/")
+    print(f"Build Complete. Created {pytorch_wheel_name}..")

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -0,0 +1,999 @@
+#!/usr/bin/env python3
+
+# This script is for building  AARCH64 wheels using AWS EC2 instances.
+# To generate binaries for the release follow these steps:
+# 1. Update mappings for each of the Domain Libraries by adding new row to a table like this:
+#         "v1.11.0": ("0.11.0", "rc1"),
+# 2. Run script with following arguments for each of the supported python versions and required tag, for example:
+# build_aarch64_wheel.py --key-name <YourPemKey> --use-docker --python 3.8 --branch v1.11.0-rc3
+
+
+import os
+import subprocess
+import sys
+import time
+from typing import Optional, Union
+
+import boto3
+
+
+# AMI images for us-east-1, change the following based on your ~/.aws/config
+os_amis = {
+    "ubuntu20_04": "ami-052eac90edaa9d08f",  # login_name: ubuntu
+    "ubuntu22_04": "ami-0c6c29c5125214c77",  # login_name: ubuntu
+    "redhat8": "ami-0698b90665a2ddcf1",  # login_name: ec2-user
+}
+
+ubuntu20_04_ami = os_amis["ubuntu20_04"]
+
+
+def compute_keyfile_path(key_name: Optional[str] = None) -> tuple[str, str]:
+    if key_name is None:
+        key_name = os.getenv("AWS_KEY_NAME")
+        if key_name is None:
+            return os.getenv("SSH_KEY_PATH", ""), ""
+
+    homedir_path = os.path.expanduser("~")
+    default_path = os.path.join(homedir_path, ".ssh", f"{key_name}.pem")
+    return os.getenv("SSH_KEY_PATH", default_path), key_name
+
+
+ec2 = boto3.resource("ec2")
+
+
+def ec2_get_instances(filter_name, filter_value):
+    return ec2.instances.filter(
+        Filters=[{"Name": filter_name, "Values": [filter_value]}]
+    )
+
+
+def ec2_instances_of_type(instance_type="t4g.2xlarge"):
+    return ec2_get_instances("instance-type", instance_type)
+
+
+def ec2_instances_by_id(instance_id):
+    rc = list(ec2_get_instances("instance-id", instance_id))
+    return rc[0] if len(rc) > 0 else None
+
+
+def start_instance(
+    key_name, ami=ubuntu20_04_ami, instance_type="t4g.2xlarge", ebs_size: int = 50
+):
+    inst = ec2.create_instances(
+        ImageId=ami,
+        InstanceType=instance_type,
+        SecurityGroups=["ssh-allworld"],
+        KeyName=key_name,
+        MinCount=1,
+        MaxCount=1,
+        BlockDeviceMappings=[
+            {
+                "DeviceName": "/dev/sda1",
+                "Ebs": {
+                    "DeleteOnTermination": True,
+                    "VolumeSize": ebs_size,
+                    "VolumeType": "standard",
+                },
+            }
+        ],
+    )[0]
+    print(f"Create instance {inst.id}")
+    inst.wait_until_running()
+    running_inst = ec2_instances_by_id(inst.id)
+    print(f"Instance started at {running_inst.public_dns_name}")
+    return running_inst
+
+
+class RemoteHost:
+    addr: str
+    keyfile_path: str
+    login_name: str
+    container_id: Optional[str] = None
+    ami: Optional[str] = None
+
+    def __init__(self, addr: str, keyfile_path: str, login_name: str = "ubuntu"):
+        self.addr = addr
+        self.keyfile_path = keyfile_path
+        self.login_name = login_name
+
+    def _gen_ssh_prefix(self) -> list[str]:
+        return [
+            "ssh",
+            "-o",
+            "StrictHostKeyChecking=no",
+            "-i",
+            self.keyfile_path,
+            f"{self.login_name}@{self.addr}",
+            "--",
+        ]
+
+    @staticmethod
+    def _split_cmd(args: Union[str, list[str]]) -> list[str]:
+        return args.split() if isinstance(args, str) else args
+
+    def run_ssh_cmd(self, args: Union[str, list[str]]) -> None:
+        subprocess.check_call(self._gen_ssh_prefix() + self._split_cmd(args))
+
+    def check_ssh_output(self, args: Union[str, list[str]]) -> str:
+        return subprocess.check_output(
+            self._gen_ssh_prefix() + self._split_cmd(args)
+        ).decode("utf-8")
+
+    def scp_upload_file(self, local_file: str, remote_file: str) -> None:
+        subprocess.check_call(
+            [
+                "scp",
+                "-i",
+                self.keyfile_path,
+                local_file,
+                f"{self.login_name}@{self.addr}:{remote_file}",
+            ]
+        )
+
+    def scp_download_file(
+        self, remote_file: str, local_file: Optional[str] = None
+    ) -> None:
+        if local_file is None:
+            local_file = "."
+        subprocess.check_call(
+            [
+                "scp",
+                "-i",
+                self.keyfile_path,
+                f"{self.login_name}@{self.addr}:{remote_file}",
+                local_file,
+            ]
+        )
+
+    def start_docker(self, image="quay.io/pypa/manylinux2014_aarch64:latest") -> None:
+        self.run_ssh_cmd("sudo apt-get install -y docker.io")
+        self.run_ssh_cmd(f"sudo usermod -a -G docker {self.login_name}")
+        self.run_ssh_cmd("sudo service docker start")
+        self.run_ssh_cmd(f"docker pull {image}")
+        self.container_id = self.check_ssh_output(
+            f"docker run -t -d -w /root {image}"
+        ).strip()
+
+    def using_docker(self) -> bool:
+        return self.container_id is not None
+
+    def run_cmd(self, args: Union[str, list[str]]) -> None:
+        if not self.using_docker():
+            return self.run_ssh_cmd(args)
+        assert self.container_id is not None
+        docker_cmd = self._gen_ssh_prefix() + [
+            "docker",
+            "exec",
+            "-i",
+            self.container_id,
+            "bash",
+        ]
+        p = subprocess.Popen(docker_cmd, stdin=subprocess.PIPE)
+        p.communicate(
+            input=" ".join(["source .bashrc && "] + self._split_cmd(args)).encode(
+                "utf-8"
+            )
+        )
+        rc = p.wait()
+        if rc != 0:
+            raise subprocess.CalledProcessError(rc, docker_cmd)
+
+    def check_output(self, args: Union[str, list[str]]) -> str:
+        if not self.using_docker():
+            return self.check_ssh_output(args)
+        assert self.container_id is not None
+        docker_cmd = self._gen_ssh_prefix() + [
+            "docker",
+            "exec",
+            "-i",
+            self.container_id,
+            "bash",
+        ]
+        p = subprocess.Popen(docker_cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+        (out, err) = p.communicate(
+            input=" ".join(["source .bashrc && "] + self._split_cmd(args)).encode(
+                "utf-8"
+            )
+        )
+        rc = p.wait()
+        if rc != 0:
+            raise subprocess.CalledProcessError(rc, docker_cmd, output=out, stderr=err)
+        return out.decode("utf-8")
+
+    def upload_file(self, local_file: str, remote_file: str) -> None:
+        if not self.using_docker():
+            return self.scp_upload_file(local_file, remote_file)
+        tmp_file = os.path.join("/tmp", os.path.basename(local_file))
+        self.scp_upload_file(local_file, tmp_file)
+        self.run_ssh_cmd(
+            ["docker", "cp", tmp_file, f"{self.container_id}:/root/{remote_file}"]
+        )
+        self.run_ssh_cmd(["rm", tmp_file])
+
+    def download_file(self, remote_file: str, local_file: Optional[str] = None) -> None:
+        if not self.using_docker():
+            return self.scp_download_file(remote_file, local_file)
+        tmp_file = os.path.join("/tmp", os.path.basename(remote_file))
+        self.run_ssh_cmd(
+            ["docker", "cp", f"{self.container_id}:/root/{remote_file}", tmp_file]
+        )
+        self.scp_download_file(tmp_file, local_file)
+        self.run_ssh_cmd(["rm", tmp_file])
+
+    def download_wheel(
+        self, remote_file: str, local_file: Optional[str] = None
+    ) -> None:
+        if self.using_docker() and local_file is None:
+            basename = os.path.basename(remote_file)
+            local_file = basename.replace(
+                "-linux_aarch64.whl", "-manylinux2014_aarch64.whl"
+            )
+        self.download_file(remote_file, local_file)
+
+    def list_dir(self, path: str) -> list[str]:
+        return self.check_output(["ls", "-1", path]).split("\n")
+
+
+def wait_for_connection(addr, port, timeout=15, attempt_cnt=5):
+    import socket
+
+    for i in range(attempt_cnt):
+        try:
+            with socket.create_connection((addr, port), timeout=timeout):
+                return
+        except (ConnectionRefusedError, TimeoutError):  # noqa: PERF203
+            if i == attempt_cnt - 1:
+                raise
+            time.sleep(timeout)
+
+
+def update_apt_repo(host: RemoteHost) -> None:
+    time.sleep(5)
+    host.run_cmd("sudo systemctl stop apt-daily.service || true")
+    host.run_cmd("sudo systemctl stop unattended-upgrades.service || true")
+    host.run_cmd(
+        "while systemctl is-active --quiet apt-daily.service; do sleep 1; done"
+    )
+    host.run_cmd(
+        "while systemctl is-active --quiet unattended-upgrades.service; do sleep 1; done"
+    )
+    host.run_cmd("sudo apt-get update")
+    time.sleep(3)
+    host.run_cmd("sudo apt-get update")
+
+
+def install_condaforge(
+    host: RemoteHost, suffix: str = "latest/download/Miniforge3-Linux-aarch64.sh"
+) -> None:
+    print("Install conda-forge")
+    host.run_cmd(f"curl -OL https://github.com/conda-forge/miniforge/releases/{suffix}")
+    host.run_cmd(f"sh -f {os.path.basename(suffix)} -b")
+    host.run_cmd(f"rm -f {os.path.basename(suffix)}")
+    if host.using_docker():
+        host.run_cmd("echo 'PATH=$HOME/miniforge3/bin:$PATH'>>.bashrc")
+    else:
+        host.run_cmd(
+            [
+                "sed",
+                "-i",
+                "'/^# If not running interactively.*/i PATH=$HOME/miniforge3/bin:$PATH'",
+                ".bashrc",
+            ]
+        )
+
+
+def install_condaforge_python(host: RemoteHost, python_version="3.8") -> None:
+    if python_version == "3.6":
+        # Python-3.6 EOLed and not compatible with conda-4.11
+        install_condaforge(
+            host, suffix="download/4.10.3-10/Miniforge3-4.10.3-10-Linux-aarch64.sh"
+        )
+        host.run_cmd(f"conda install -y python={python_version} numpy pyyaml")
+    else:
+        install_condaforge(
+            host, suffix="download/4.11.0-4/Miniforge3-4.11.0-4-Linux-aarch64.sh"
+        )
+        # Pytorch-1.10 or older are not compatible with setuptools=59.6 or newer
+        host.run_cmd(
+            f"conda install -y python={python_version} numpy pyyaml setuptools>=59.5.0"
+        )
+
+
+def embed_libgomp(host: RemoteHost, use_conda, wheel_name) -> None:
+    host.run_cmd("pip3 install auditwheel")
+    host.run_cmd(
+        "conda install -y patchelf" if use_conda else "sudo apt-get install -y patchelf"
+    )
+    from tempfile import NamedTemporaryFile
+
+    with NamedTemporaryFile() as tmp:
+        tmp.write(embed_library_script.encode("utf-8"))
+        tmp.flush()
+        host.upload_file(tmp.name, "embed_library.py")
+
+    print("Embedding libgomp into wheel")
+    if host.using_docker():
+        host.run_cmd(f"python3 embed_library.py {wheel_name} --update-tag")
+    else:
+        host.run_cmd(f"python3 embed_library.py {wheel_name}")
+
+
+def checkout_repo(
+    host: RemoteHost,
+    *,
+    branch: str = "main",
+    url: str,
+    git_clone_flags: str,
+    mapping: dict[str, tuple[str, str]],
+) -> Optional[str]:
+    for prefix in mapping:
+        if not branch.startswith(prefix):
+            continue
+        tag = f"v{mapping[prefix][0]}-{mapping[prefix][1]}"
+        host.run_cmd(f"git clone {url} -b {tag} {git_clone_flags}")
+        return mapping[prefix][0]
+
+    host.run_cmd(f"git clone {url} -b {branch} {git_clone_flags}")
+    return None
+
+
+def build_torchvision(
+    host: RemoteHost,
+    *,
+    branch: str = "main",
+    use_conda: bool = True,
+    git_clone_flags: str,
+    run_smoke_tests: bool = True,
+) -> str:
+    print("Checking out TorchVision repo")
+    build_version = checkout_repo(
+        host,
+        branch=branch,
+        url="https://github.com/pytorch/vision",
+        git_clone_flags=git_clone_flags,
+        mapping={
+            "v1.7.1": ("0.8.2", "rc2"),
+            "v1.8.0": ("0.9.0", "rc3"),
+            "v1.8.1": ("0.9.1", "rc1"),
+            "v1.9.0": ("0.10.0", "rc1"),
+            "v1.10.0": ("0.11.1", "rc1"),
+            "v1.10.1": ("0.11.2", "rc1"),
+            "v1.10.2": ("0.11.3", "rc1"),
+            "v1.11.0": ("0.12.0", "rc1"),
+            "v1.12.0": ("0.13.0", "rc4"),
+            "v1.12.1": ("0.13.1", "rc6"),
+            "v1.13.0": ("0.14.0", "rc4"),
+            "v1.13.1": ("0.14.1", "rc2"),
+            "v2.0.0": ("0.15.1", "rc2"),
+            "v2.0.1": ("0.15.2", "rc2"),
+        },
+    )
+    print("Building TorchVision wheel")
+
+    # Please note libnpg and jpeg are required to build image.so extension
+    if use_conda:
+        host.run_cmd("conda install -y libpng jpeg")
+        # Remove .so files to force static linking
+        host.run_cmd(
+            "rm miniforge3/lib/libpng.so miniforge3/lib/libpng16.so miniforge3/lib/libjpeg.so"
+        )
+        # And patch setup.py to include libz dependency for libpng
+        host.run_cmd(
+            [
+                'sed -i -e \'s/image_link_flags\\.append("png")/image_link_flags += ["png", "z"]/\' vision/setup.py'
+            ]
+        )
+
+    build_vars = ""
+    if branch == "nightly":
+        version = host.check_output(
+            ["if [ -f vision/version.txt ]; then cat vision/version.txt; fi"]
+        ).strip()
+        if len(version) == 0:
+            # In older revisions, version was embedded in setup.py
+            version = (
+                host.check_output(["grep", '"version = \'"', "vision/setup.py"])
+                .strip()
+                .split("'")[1][:-2]
+            )
+        build_date = (
+            host.check_output("cd vision && git log --pretty=format:%s -1")
+            .strip()
+            .split()[0]
+            .replace("-", "")
+        )
+        build_vars += f"BUILD_VERSION={version}.dev{build_date}"
+    elif build_version is not None:
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+    if host.using_docker():
+        build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
+
+    host.run_cmd(f"cd vision && {build_vars} python3 -m build --wheel --no-isolation")
+    vision_wheel_name = host.list_dir("vision/dist")[0]
+    embed_libgomp(host, use_conda, os.path.join("vision", "dist", vision_wheel_name))
+
+    print("Copying TorchVision wheel")
+    host.download_wheel(os.path.join("vision", "dist", vision_wheel_name))
+    if run_smoke_tests:
+        host.run_cmd(
+            f"pip3 install {os.path.join('vision', 'dist', vision_wheel_name)}"
+        )
+        host.run_cmd("python3 vision/test/smoke_test.py")
+    print("Delete vision checkout")
+    host.run_cmd("rm -rf vision")
+
+    return vision_wheel_name
+
+
+def build_torchdata(
+    host: RemoteHost,
+    *,
+    branch: str = "main",
+    use_conda: bool = True,
+    git_clone_flags: str = "",
+) -> str:
+    print("Checking out TorchData repo")
+    git_clone_flags += " --recurse-submodules"
+    build_version = checkout_repo(
+        host,
+        branch=branch,
+        url="https://github.com/pytorch/data",
+        git_clone_flags=git_clone_flags,
+        mapping={
+            "v1.13.1": ("0.5.1", ""),
+            "v2.0.0": ("0.6.0", "rc5"),
+            "v2.0.1": ("0.6.1", "rc1"),
+        },
+    )
+    print("Building TorchData wheel")
+    build_vars = ""
+    if branch == "nightly":
+        version = host.check_output(
+            ["if [ -f data/version.txt ]; then cat data/version.txt; fi"]
+        ).strip()
+        build_date = (
+            host.check_output("cd data && git log --pretty=format:%s -1")
+            .strip()
+            .split()[0]
+            .replace("-", "")
+        )
+        build_vars += f"BUILD_VERSION={version}.dev{build_date}"
+    elif build_version is not None:
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+    if host.using_docker():
+        build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
+
+    host.run_cmd(f"cd data && {build_vars} python3 -m build --wheel --no-isolation")
+    wheel_name = host.list_dir("data/dist")[0]
+    embed_libgomp(host, use_conda, os.path.join("data", "dist", wheel_name))
+
+    print("Copying TorchData wheel")
+    host.download_wheel(os.path.join("data", "dist", wheel_name))
+
+    return wheel_name
+
+
+def build_torchtext(
+    host: RemoteHost,
+    *,
+    branch: str = "main",
+    use_conda: bool = True,
+    git_clone_flags: str = "",
+) -> str:
+    print("Checking out TorchText repo")
+    git_clone_flags += " --recurse-submodules"
+    build_version = checkout_repo(
+        host,
+        branch=branch,
+        url="https://github.com/pytorch/text",
+        git_clone_flags=git_clone_flags,
+        mapping={
+            "v1.9.0": ("0.10.0", "rc1"),
+            "v1.10.0": ("0.11.0", "rc2"),
+            "v1.10.1": ("0.11.1", "rc1"),
+            "v1.10.2": ("0.11.2", "rc1"),
+            "v1.11.0": ("0.12.0", "rc1"),
+            "v1.12.0": ("0.13.0", "rc2"),
+            "v1.12.1": ("0.13.1", "rc5"),
+            "v1.13.0": ("0.14.0", "rc3"),
+            "v1.13.1": ("0.14.1", "rc1"),
+            "v2.0.0": ("0.15.1", "rc2"),
+            "v2.0.1": ("0.15.2", "rc2"),
+        },
+    )
+    print("Building TorchText wheel")
+    build_vars = ""
+    if branch == "nightly":
+        version = host.check_output(
+            ["if [ -f text/version.txt ]; then cat text/version.txt; fi"]
+        ).strip()
+        build_date = (
+            host.check_output("cd text && git log --pretty=format:%s -1")
+            .strip()
+            .split()[0]
+            .replace("-", "")
+        )
+        build_vars += f"BUILD_VERSION={version}.dev{build_date}"
+    elif build_version is not None:
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+    if host.using_docker():
+        build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
+
+    host.run_cmd(f"cd text && {build_vars} python3 -m build --wheel --no-isolation")
+    wheel_name = host.list_dir("text/dist")[0]
+    embed_libgomp(host, use_conda, os.path.join("text", "dist", wheel_name))
+
+    print("Copying TorchText wheel")
+    host.download_wheel(os.path.join("text", "dist", wheel_name))
+
+    return wheel_name
+
+
+def build_torchaudio(
+    host: RemoteHost,
+    *,
+    branch: str = "main",
+    use_conda: bool = True,
+    git_clone_flags: str = "",
+) -> str:
+    print("Checking out TorchAudio repo")
+    git_clone_flags += " --recurse-submodules"
+    build_version = checkout_repo(
+        host,
+        branch=branch,
+        url="https://github.com/pytorch/audio",
+        git_clone_flags=git_clone_flags,
+        mapping={
+            "v1.9.0": ("0.9.0", "rc2"),
+            "v1.10.0": ("0.10.0", "rc5"),
+            "v1.10.1": ("0.10.1", "rc1"),
+            "v1.10.2": ("0.10.2", "rc1"),
+            "v1.11.0": ("0.11.0", "rc1"),
+            "v1.12.0": ("0.12.0", "rc3"),
+            "v1.12.1": ("0.12.1", "rc5"),
+            "v1.13.0": ("0.13.0", "rc4"),
+            "v1.13.1": ("0.13.1", "rc2"),
+            "v2.0.0": ("2.0.1", "rc3"),
+            "v2.0.1": ("2.0.2", "rc2"),
+        },
+    )
+    print("Building TorchAudio wheel")
+    build_vars = ""
+    if branch == "nightly":
+        version = (
+            host.check_output(["grep", '"version = \'"', "audio/setup.py"])
+            .strip()
+            .split("'")[1][:-2]
+        )
+        build_date = (
+            host.check_output("cd audio && git log --pretty=format:%s -1")
+            .strip()
+            .split()[0]
+            .replace("-", "")
+        )
+        build_vars += f"BUILD_VERSION={version}.dev{build_date}"
+    elif build_version is not None:
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+    if host.using_docker():
+        build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
+
+    host.run_cmd(
+        f"cd audio && export FFMPEG_ROOT=$(pwd)/third_party/ffmpeg && export USE_FFMPEG=1 \
+        && ./packaging/ffmpeg/build.sh \
+        && {build_vars} python3 -m build --wheel --no-isolation"
+    )
+
+    wheel_name = host.list_dir("audio/dist")[0]
+    embed_libgomp(host, use_conda, os.path.join("audio", "dist", wheel_name))
+
+    print("Copying TorchAudio wheel")
+    host.download_wheel(os.path.join("audio", "dist", wheel_name))
+
+    return wheel_name
+
+
+def configure_system(
+    host: RemoteHost,
+    *,
+    compiler: str = "gcc-8",
+    use_conda: bool = True,
+    python_version: str = "3.8",
+) -> None:
+    if use_conda:
+        install_condaforge_python(host, python_version)
+
+    print("Configuring the system")
+    if not host.using_docker():
+        update_apt_repo(host)
+        host.run_cmd("sudo apt-get install -y ninja-build g++ git cmake gfortran unzip")
+    else:
+        host.run_cmd("yum install -y sudo")
+        host.run_cmd("conda install -y ninja scons")
+
+    if not use_conda:
+        host.run_cmd(
+            "sudo apt-get install -y python3-dev python3-yaml python3-setuptools python3-wheel python3-pip"
+        )
+    host.run_cmd("pip3 install dataclasses typing-extensions")
+    if not use_conda:
+        print("Installing Cython + numpy from PyPy")
+        host.run_cmd("sudo pip3 install Cython")
+        host.run_cmd("sudo pip3 install numpy")
+
+
+def build_domains(
+    host: RemoteHost,
+    *,
+    branch: str = "main",
+    use_conda: bool = True,
+    git_clone_flags: str = "",
+) -> tuple[str, str, str, str]:
+    vision_wheel_name = build_torchvision(
+        host, branch=branch, use_conda=use_conda, git_clone_flags=git_clone_flags
+    )
+    audio_wheel_name = build_torchaudio(
+        host, branch=branch, use_conda=use_conda, git_clone_flags=git_clone_flags
+    )
+    data_wheel_name = build_torchdata(
+        host, branch=branch, use_conda=use_conda, git_clone_flags=git_clone_flags
+    )
+    text_wheel_name = build_torchtext(
+        host, branch=branch, use_conda=use_conda, git_clone_flags=git_clone_flags
+    )
+    return (vision_wheel_name, audio_wheel_name, data_wheel_name, text_wheel_name)
+
+
+def start_build(
+    host: RemoteHost,
+    *,
+    branch: str = "main",
+    compiler: str = "gcc-8",
+    use_conda: bool = True,
+    python_version: str = "3.8",
+    pytorch_only: bool = False,
+    pytorch_build_number: Optional[str] = None,
+    shallow_clone: bool = True,
+    enable_mkldnn: bool = False,
+) -> tuple[str, str, str, str, str]:
+    git_clone_flags = " --depth 1 --shallow-submodules" if shallow_clone else ""
+    if host.using_docker() and not use_conda:
+        print("Auto-selecting conda option for docker images")
+        use_conda = True
+    if not host.using_docker():
+        print("Disable mkldnn for host builds")
+        enable_mkldnn = False
+
+    configure_system(
+        host, compiler=compiler, use_conda=use_conda, python_version=python_version
+    )
+
+    if host.using_docker():
+        print("Move libgfortant.a into a standard location")
+        # HACK: pypa gforntran.a is compiled without PIC, which leads to the following error
+        # libgfortran.a(error.o)(.text._gfortrani_st_printf+0x34): unresolvable R_AARCH64_ADR_PREL_PG_HI21 relocation against symbol `__stack_chk_guard@@GLIBC_2.17'  # noqa: E501, B950
+        # Workaround by copying gfortran library from the host
+        host.run_ssh_cmd("sudo apt-get install -y gfortran-8")
+        host.run_cmd("mkdir -p /usr/lib/gcc/aarch64-linux-gnu/8")
+        host.run_ssh_cmd(
+            [
+                "docker",
+                "cp",
+                "/usr/lib/gcc/aarch64-linux-gnu/8/libgfortran.a",
+                f"{host.container_id}:/opt/rh/devtoolset-10/root/usr/lib/gcc/aarch64-redhat-linux/10/",
+            ]
+        )
+
+    print("Checking out PyTorch repo")
+    host.run_cmd(
+        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:
+        build_opts += f" -C--build-option=--build-number={pytorch_build_number}"
+    # Breakpad build fails on aarch64
+    build_vars = "USE_BREAKPAD=0 "
+    if branch == "nightly":
+        build_date = (
+            host.check_output("cd pytorch && git log --pretty=format:%s -1")
+            .strip()
+            .split()[0]
+            .replace("-", "")
+        )
+        version = host.check_output("cat pytorch/version.txt").strip()[:-2]
+        build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={version}.dev{build_date} PYTORCH_BUILD_NUMBER=1"
+    if branch.startswith(("v1.", "v2.")):
+        build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={branch[1 : branch.find('-')]} PYTORCH_BUILD_NUMBER=1"
+    if host.using_docker():
+        build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
+    if enable_mkldnn:
+        host.run_cmd("pytorch/.ci/docker/common/install_acl.sh")
+        print("build pytorch with mkldnn+acl backend")
+        build_vars += " USE_MKLDNN=ON USE_MKLDNN_ACL=ON"
+        build_vars += " BLAS=OpenBLAS"
+        build_vars += " OpenBLAS_HOME=/opt/OpenBLAS"
+        build_vars += " ACL_ROOT_DIR=/acl"
+        host.run_cmd(
+            f"cd $HOME/pytorch && {build_vars} python3 -m build --wheel --no-isolation{build_opts}"
+        )
+        print("Repair the wheel")
+        pytorch_wheel_name = host.list_dir("pytorch/dist")[0]
+        ld_library_path = "/acl/build:$HOME/pytorch/build/lib"
+        host.run_cmd(
+            f"export LD_LIBRARY_PATH={ld_library_path} && auditwheel repair $HOME/pytorch/dist/{pytorch_wheel_name}"
+        )
+        print("replace the original wheel with the repaired one")
+        pytorch_repaired_wheel_name = host.list_dir("wheelhouse")[0]
+        host.run_cmd(
+            f"cp $HOME/wheelhouse/{pytorch_repaired_wheel_name} $HOME/pytorch/dist/{pytorch_wheel_name}"
+        )
+    else:
+        print("build pytorch without mkldnn backend")
+        host.run_cmd(
+            f"cd pytorch && {build_vars} python3 -m build --wheel --no-isolation{build_opts}"
+        )
+
+    print("Deleting build folder")
+    host.run_cmd("cd pytorch && rm -rf build")
+    pytorch_wheel_name = host.list_dir("pytorch/dist")[0]
+    embed_libgomp(host, use_conda, os.path.join("pytorch", "dist", pytorch_wheel_name))
+    print("Copying the wheel")
+    host.download_wheel(os.path.join("pytorch", "dist", pytorch_wheel_name))
+
+    print("Installing PyTorch wheel")
+    host.run_cmd(f"pip3 install pytorch/dist/{pytorch_wheel_name}")
+
+    if pytorch_only:
+        return (pytorch_wheel_name, None, None, None, None)
+    domain_wheels = build_domains(
+        host, branch=branch, use_conda=use_conda, git_clone_flags=git_clone_flags
+    )
+
+    return (pytorch_wheel_name, *domain_wheels)
+
+
+embed_library_script = """
+#!/usr/bin/env python3
+
+from auditwheel.patcher import Patchelf
+from auditwheel.wheeltools import InWheelCtx
+from auditwheel.elfutils import elf_file_filter
+from auditwheel.repair import copylib
+from auditwheel.lddtree import lddtree
+from subprocess import check_call
+import os
+import shutil
+import sys
+from tempfile import TemporaryDirectory
+
+
+def replace_tag(filename):
+   with open(filename, 'r') as f:
+     lines = f.read().split("\\n")
+   for i,line in enumerate(lines):
+       if not line.startswith("Tag: "):
+           continue
+       lines[i] = line.replace("-linux_", "-manylinux2014_")
+       print(f'Updated tag from {line} to {lines[i]}')
+
+   with open(filename, 'w') as f:
+       f.write("\\n".join(lines))
+
+
+class AlignedPatchelf(Patchelf):
+    def set_soname(self, file_name: str, new_soname: str) -> None:
+        check_call(['patchelf', '--page-size', '65536', '--set-soname', new_soname, file_name])
+
+    def replace_needed(self, file_name: str, soname: str, new_soname: str) -> None:
+        check_call(['patchelf', '--page-size', '65536', '--replace-needed', soname, new_soname, file_name])
+
+
+def embed_library(whl_path, lib_soname, update_tag=False):
+    patcher = AlignedPatchelf()
+    out_dir = TemporaryDirectory()
+    whl_name = os.path.basename(whl_path)
+    tmp_whl_name = os.path.join(out_dir.name, whl_name)
+    with InWheelCtx(whl_path) as ctx:
+        torchlib_path = os.path.join(ctx._tmpdir.name, 'torch', 'lib')
+        ctx.out_wheel=tmp_whl_name
+        new_lib_path, new_lib_soname = None, None
+        for filename, elf in elf_file_filter(ctx.iter_files()):
+            if not filename.startswith('torch/lib'):
+                continue
+            libtree = lddtree(filename)
+            if lib_soname not in libtree['needed']:
+                continue
+            lib_path = libtree['libs'][lib_soname]['path']
+            if lib_path is None:
+                print(f"Can't embed {lib_soname} as it could not be found")
+                break
+            if lib_path.startswith(torchlib_path):
+                continue
+
+            if new_lib_path is None:
+                new_lib_soname, new_lib_path = copylib(lib_path, torchlib_path, patcher)
+            patcher.replace_needed(filename, lib_soname, new_lib_soname)
+            print(f'Replacing {lib_soname} with {new_lib_soname} for {filename}')
+        if update_tag:
+            # Add manylinux2014 tag
+            for filename in ctx.iter_files():
+                if os.path.basename(filename) != 'WHEEL':
+                    continue
+                replace_tag(filename)
+    shutil.move(tmp_whl_name, whl_path)
+
+
+if __name__ == '__main__':
+    embed_library(sys.argv[1], 'libgomp.so.1', len(sys.argv) > 2 and sys.argv[2] == '--update-tag')
+"""
+
+
+def run_tests(host: RemoteHost, whl: str, branch="main") -> None:
+    print("Configuring the system")
+    update_apt_repo(host)
+    host.run_cmd("sudo apt-get install -y python3-pip git")
+    host.run_cmd("sudo pip3 install Cython")
+    host.run_cmd("sudo pip3 install numpy")
+    host.upload_file(whl, ".")
+    host.run_cmd(f"sudo pip3 install {whl}")
+    host.run_cmd("python3 -c 'import torch;print(torch.rand((3,3))'")
+    host.run_cmd(f"git clone -b {branch} https://github.com/pytorch/pytorch")
+    host.run_cmd("cd pytorch/test; python3 test_torch.py -v")
+
+
+def get_instance_name(instance) -> Optional[str]:
+    if instance.tags is None:
+        return None
+    for tag in instance.tags:
+        if tag["Key"] == "Name":
+            return tag["Value"]
+    return None
+
+
+def list_instances(instance_type: str) -> None:
+    print(f"All instances of type {instance_type}")
+    for instance in ec2_instances_of_type(instance_type):
+        ifaces = instance.network_interfaces
+        az = ifaces[0].subnet.availability_zone if len(ifaces) > 0 else None
+        print(
+            f"{instance.id} {get_instance_name(instance)} {instance.public_dns_name} {instance.state['Name']} {az}"
+        )
+
+
+def terminate_instances(instance_type: str) -> None:
+    print(f"Terminating all instances of type {instance_type}")
+    instances = list(ec2_instances_of_type(instance_type))
+    for instance in instances:
+        print(f"Terminating {instance.id}")
+        instance.terminate()
+    print("Waiting for termination to complete")
+    for instance in instances:
+        instance.wait_until_terminated()
+
+
+def parse_arguments():
+    from argparse import ArgumentParser
+
+    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")
+    parser.add_argument("--test-only", type=str)
+    group = parser.add_mutually_exclusive_group()
+    group.add_argument("--os", type=str, choices=list(os_amis.keys()))
+    group.add_argument("--ami", type=str)
+    parser.add_argument(
+        "--python-version",
+        type=str,
+        choices=[f"3.{d}" for d in range(6, 12)],
+        default=None,
+    )
+    parser.add_argument("--alloc-instance", action="store_true")
+    parser.add_argument("--list-instances", action="store_true")
+    parser.add_argument("--pytorch-only", action="store_true")
+    parser.add_argument("--keep-running", action="store_true")
+    parser.add_argument("--terminate-instances", action="store_true")
+    parser.add_argument("--instance-type", type=str, default="t4g.2xlarge")
+    parser.add_argument("--ebs-size", type=int, default=50)
+    parser.add_argument("--branch", type=str, default="main")
+    parser.add_argument("--use-docker", action="store_true")
+    parser.add_argument(
+        "--compiler",
+        type=str,
+        choices=["gcc-7", "gcc-8", "gcc-9", "clang"],
+        default="gcc-8",
+    )
+    parser.add_argument("--use-torch-from-pypi", action="store_true")
+    parser.add_argument("--pytorch-build-number", type=str, default=None)
+    parser.add_argument("--disable-mkldnn", action="store_true")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_arguments()
+    ami = (
+        args.ami
+        if args.ami is not None
+        else os_amis[args.os]
+        if args.os is not None
+        else ubuntu20_04_ami
+    )
+    keyfile_path, key_name = compute_keyfile_path(args.key_name)
+
+    if args.list_instances:
+        list_instances(args.instance_type)
+        sys.exit(0)
+
+    if args.terminate_instances:
+        terminate_instances(args.instance_type)
+        sys.exit(0)
+
+    if len(key_name) == 0:
+        raise RuntimeError("""
+            Cannot start build without key_name, please specify
+            --key-name argument or AWS_KEY_NAME environment variable.""")
+    if len(keyfile_path) == 0 or not os.path.exists(keyfile_path):
+        raise RuntimeError(f"""
+            Cannot find keyfile with name: [{key_name}] in path: [{keyfile_path}], please
+            check `~/.ssh/` folder or manually set SSH_KEY_PATH environment variable.""")
+
+    # Starting the instance
+    inst = start_instance(
+        key_name, ami=ami, instance_type=args.instance_type, ebs_size=args.ebs_size
+    )
+    instance_name = f"{args.key_name}-{args.os}"
+    if args.python_version is not None:
+        instance_name += f"-py{args.python_version}"
+    inst.create_tags(
+        DryRun=False,
+        Tags=[
+            {
+                "Key": "Name",
+                "Value": instance_name,
+            }
+        ],
+    )
+    addr = inst.public_dns_name
+    wait_for_connection(addr, 22)
+    host = RemoteHost(addr, keyfile_path)
+    host.ami = ami
+    if args.use_docker:
+        update_apt_repo(host)
+        host.start_docker()
+
+    if args.test_only:
+        run_tests(host, args.test_only)
+        sys.exit(0)
+
+    if args.alloc_instance:
+        if args.python_version is None:
+            sys.exit(0)
+        install_condaforge_python(host, args.python_version)
+        sys.exit(0)
+
+    python_version = args.python_version if args.python_version is not None else "3.10"
+
+    if args.use_torch_from_pypi:
+        configure_system(host, compiler=args.compiler, python_version=python_version)
+        print("Installing PyTorch wheel")
+        host.run_cmd("pip3 install torch")
+        build_domains(
+            host, branch=args.branch, git_clone_flags=" --depth 1 --shallow-submodules"
+        )
+    else:
+        start_build(
+            host,
+            branch=args.branch,
+            compiler=args.compiler,
+            python_version=python_version,
+            pytorch_only=args.pytorch_only,
+            pytorch_build_number=args.pytorch_build_number,
+            enable_mkldnn=not args.disable_mkldnn,
+        )
+    if not args.keep_running:
+        print(f"Waiting for instance {inst.id} to terminate")
+        inst.terminate()
+        inst.wait_until_terminated()

.ci/aarch64_linux/embed_library.py

@@ -0,0 +1,87 @@
+#!/usr/bin/env python3
+
+import os
+import shutil
+import sys
+from subprocess import check_call
+from tempfile import TemporaryDirectory
+
+from auditwheel.elfutils import elf_file_filter
+from auditwheel.lddtree import lddtree
+from auditwheel.patcher import Patchelf
+from auditwheel.repair import copylib
+from auditwheel.wheeltools import InWheelCtx
+
+
+def replace_tag(filename):
+    with open(filename) as f:
+        lines = f.read().split("\\n")
+    for i, line in enumerate(lines):
+        if not line.startswith("Tag: "):
+            continue
+        lines[i] = line.replace("-linux_", "-manylinux2014_")
+        print(f"Updated tag from {line} to {lines[i]}")
+
+    with open(filename, "w") as f:
+        f.write("\\n".join(lines))
+
+
+class AlignedPatchelf(Patchelf):
+    def set_soname(self, file_name: str, new_soname: str) -> None:
+        check_call(
+            ["patchelf", "--page-size", "65536", "--set-soname", new_soname, file_name]
+        )
+
+    def replace_needed(self, file_name: str, soname: str, new_soname: str) -> None:
+        check_call(
+            [
+                "patchelf",
+                "--page-size",
+                "65536",
+                "--replace-needed",
+                soname,
+                new_soname,
+                file_name,
+            ]
+        )
+
+
+def embed_library(whl_path, lib_soname, update_tag=False):
+    patcher = AlignedPatchelf()
+    out_dir = TemporaryDirectory()
+    whl_name = os.path.basename(whl_path)
+    tmp_whl_name = os.path.join(out_dir.name, whl_name)
+    with InWheelCtx(whl_path) as ctx:
+        torchlib_path = os.path.join(ctx._tmpdir.name, "torch", "lib")
+        ctx.out_wheel = tmp_whl_name
+        new_lib_path, new_lib_soname = None, None
+        for filename, _ in elf_file_filter(ctx.iter_files()):
+            if not filename.startswith("torch/lib"):
+                continue
+            libtree = lddtree(filename)
+            if lib_soname not in libtree["needed"]:
+                continue
+            lib_path = libtree["libs"][lib_soname]["path"]
+            if lib_path is None:
+                print(f"Can't embed {lib_soname} as it could not be found")
+                break
+            if lib_path.startswith(torchlib_path):
+                continue
+
+            if new_lib_path is None:
+                new_lib_soname, new_lib_path = copylib(lib_path, torchlib_path, patcher)
+            patcher.replace_needed(filename, lib_soname, new_lib_soname)
+            print(f"Replacing {lib_soname} with {new_lib_soname} for {filename}")
+        if update_tag:
+            # Add manylinux2014 tag
+            for filename in ctx.iter_files():
+                if os.path.basename(filename) != "WHEEL":
+                    continue
+                replace_tag(filename)
+    shutil.move(tmp_whl_name, whl_path)
+
+
+if __name__ == "__main__":
+    embed_library(
+        sys.argv[1], "libgomp.so.1", len(sys.argv) > 2 and sys.argv[2] == "--update-tag"
+    )

.ci/docker/build.sh

@@ -188,7 +188,7 @@ case "$tag" in
     fi
     GCC_VERSION=11
     VISION=yes
-    ROCM_VERSION=7.0
+    ROCM_VERSION=7.1
     NINJA_VERSION=1.9.0
     TRITON=yes
     KATEX=yes

.ci/docker/common/install_rocm.sh

@@ -60,14 +60,16 @@ EOF
         DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated rocm-llvm-dev
     fi
 
-    # precompiled miopen kernels added in ROCm 3.5, renamed in ROCm 5.5
-    # search for all unversioned packages
-    # if search fails it will abort this script; use true to avoid case where search fails
-    MIOPENHIPGFX=$(apt-cache search --names-only miopen-hip-gfx | awk '{print $1}' | grep -F -v . || true)
-    if [[ "x${MIOPENHIPGFX}" = x ]]; then
-      echo "miopen-hip-gfx package not available" && exit 1
-    else
-      DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ${MIOPENHIPGFX}
+    if [[ $(ver $ROCM_VERSION) -lt $(ver 7.1) ]]; then
+      # precompiled miopen kernels added in ROCm 3.5, renamed in ROCm 5.5, removed in ROCm 7.1
+      # search for all unversioned packages
+      # if search fails it will abort this script; use true to avoid case where search fails
+      MIOPENHIPGFX=$(apt-cache search --names-only miopen-hip-gfx | awk '{print $1}' | grep -F -v . || true)
+      if [[ "x${MIOPENHIPGFX}" = x ]]; then
+        echo "miopen-hip-gfx package not available" && exit 1
+      else
+        DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ${MIOPENHIPGFX}
+      fi
     fi
 
     # ROCm 6.0 had a regression where journal_mode was enabled on the kdb files resulting in permission errors at runtime

.ci/docker/common/install_rocm_magma.sh

@@ -12,8 +12,8 @@ function do_install() {
 
     rocm_version_nodot=${rocm_version//./}
 
-    # post merge of https://github.com/icl-utk-edu/magma/pull/65
-    MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
+    # https://github.com/icl-utk-edu/magma/pull/65
+    MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
     magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"
 
     rocm_dir="/opt/rocm"

.ci/manywheel/build.sh

@@ -4,17 +4,14 @@ set -ex
 
 SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
-# Source the common build script for architecture-specific configurations (MKLDNN, ACL, etc.)
-source "${SCRIPTPATH}/../pytorch/build.sh" || true
-
 case "${GPU_ARCH_TYPE:-BLANK}" in
-    cuda | cuda-aarch64)
+    cuda)
         bash "${SCRIPTPATH}/build_cuda.sh"
         ;;
     rocm)
         bash "${SCRIPTPATH}/build_rocm.sh"
         ;;
-    cpu | cpu-cxx11-abi | cpu-aarch64 | cpu-s390x)
+    cpu | cpu-cxx11-abi | cpu-s390x)
         bash "${SCRIPTPATH}/build_cpu.sh"
         ;;
     xpu)

.ci/manywheel/build_common.sh

@@ -18,31 +18,12 @@ retry () {
     $*  || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
 }
 
-# Detect architecture first
-ARCH=$(uname -m)
-echo "Detected architecture: $ARCH"
-
 PLATFORM=""
 # TODO move this into the Docker images
 OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
 if [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
     retry yum install -q -y zip openssl
-    # Set platform based on architecture
-    case $ARCH in
-        x86_64)
-            PLATFORM="manylinux_2_28_x86_64"
-            ;;
-        aarch64)
-            PLATFORM="manylinux_2_28_aarch64"
-            ;;
-        s390x)
-            PLATFORM="manylinux_2_28_s390x"
-            ;;
-        *)
-            echo "Unsupported architecture: $ARCH"
-            exit 1
-            ;;
-    esac
+    PLATFORM="manylinux_2_28_x86_64"
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
     retry dnf install -q -y zip openssl
 elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
@@ -57,8 +38,6 @@ else
     exit 1
 fi
 
-echo "Platform set to: $PLATFORM"
-
 # We use the package name to test the package by passing this to 'pip install'
 # This is the env variable that setup.py uses to name the package. Note that
 # pip 'normalizes' the name first by changing all - to _
@@ -320,8 +299,8 @@ for pkg in /$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/torch*linux*.w
             # ROCm workaround for roctracer dlopens
             if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
                 patchedpath=$(fname_without_so_number $destpath)
-            # Keep the so number for XPU dependencies, libgomp.so.1, ACL libraries, and NVPL libraries to avoid twice load
-            elif [[ "$DESIRED_CUDA" == *"xpu"* || "$filename" == "libgomp.so.1" || "$filename" == libarm_compute* || "$filename" == libnvpl* || "$filename" == "libgfortran.so.5" ]]; then
+            # Keep the so number for XPU dependencies and libgomp.so.1 to avoid twice load
+            elif [[ "$DESIRED_CUDA" == *"xpu"* || "$filename" == "libgomp.so.1" ]]; then
                 patchedpath=$destpath
             else
                 patchedpath=$(fname_with_sha256 $destpath)
@@ -367,22 +346,9 @@ for pkg in /$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/torch*linux*.w
     done
 
     # create Manylinux 2_28 tag this needs to happen before regenerate the RECORD
-    # Support all architectures (x86_64, aarch64, s390x)
-    if [[ "$IS_MANYLINUX2_28" == "1" && $GPU_ARCH_TYPE != "xpu" ]]; then
+    if [[ $PLATFORM == "manylinux_2_28_x86_64" && $GPU_ARCH_TYPE != "cpu-s390x" && $GPU_ARCH_TYPE != "xpu" ]]; then
         wheel_file=$(echo $(basename $pkg) | sed -e 's/-cp.*$/.dist-info\/WHEEL/g')
-        echo "Updating wheel tag for $ARCH architecture"
-        # Replace linux_* with manylinux_2_28_* based on architecture
-        case $ARCH in
-            x86_64)
-                sed -i -e 's#linux_x86_64#manylinux_2_28_x86_64#g' $wheel_file
-                ;;
-            aarch64)
-                sed -i -e 's#linux_aarch64#manylinux_2_28_aarch64#g' $wheel_file
-                ;;
-            s390x)
-                sed -i -e 's#linux_s390x#manylinux_2_28_s390x#g' $wheel_file
-                ;;
-        esac
+        sed -i -e s#linux_x86_64#"${PLATFORM}"# $wheel_file;
     fi
 
     # regenerate the RECORD file with new hashes

.ci/manywheel/build_cpu.sh

@@ -15,10 +15,6 @@ if [[ -z "$EXTRA_CAFFE2_CMAKE_FLAGS" ]]; then
     EXTRA_CAFFE2_CMAKE_FLAGS=()
 fi
 
-# Detect architecture
-ARCH=$(uname -m)
-echo "Building CPU wheel for architecture: $ARCH"
-
 WHEELHOUSE_DIR="wheelhousecpu"
 LIBTORCH_HOUSE_DIR="libtorch_housecpu"
 if [[ -z "$PYTORCH_FINAL_PACKAGE_DIR" ]]; then
@@ -38,10 +34,8 @@ elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
 elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
     LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
 elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
-    if [[ "$ARCH" == "s390x" ]]; then
+    if [[ "$(uname -m)" == "s390x" ]]; then
         LIBGOMP_PATH="/usr/lib/s390x-linux-gnu/libgomp.so.1"
-    elif [[ "$ARCH" == "aarch64" ]]; then
-        LIBGOMP_PATH="/usr/lib/aarch64-linux-gnu/libgomp.so.1"
     else
         LIBGOMP_PATH="/usr/lib/x86_64-linux-gnu/libgomp.so.1"
     fi
@@ -55,34 +49,6 @@ DEPS_SONAME=(
     "libgomp.so.1"
 )
 
-# Add ARM-specific library dependencies for CPU builds
-if [[ "$ARCH" == "aarch64" ]]; then
-    echo "Adding ARM-specific CPU library dependencies"
-
-    # ARM Compute Library (if available)
-    if [[ -d "/acl/build" ]]; then
-        echo "Adding ARM Compute Library for CPU"
-        DEPS_LIST+=(
-            "/acl/build/libarm_compute.so"
-            "/acl/build/libarm_compute_graph.so"
-        )
-        DEPS_SONAME+=(
-            "libarm_compute.so"
-            "libarm_compute_graph.so"
-        )
-    fi
-
-    # ARM system libraries
-    DEPS_LIST+=(
-        "/usr/lib64/libgfortran.so.5"
-        "/opt/OpenBLAS/lib/libopenblas.so.0"
-    )
-    DEPS_SONAME+=(
-        "libgfortran.so.5"
-        "libopenblas.so.0"
-    )
-fi
-
 rm -rf /usr/local/cuda*
 
 SOURCE_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null && pwd )"

.ci/manywheel/build_cuda.sh

@@ -29,10 +29,6 @@ if [[ -z "$EXTRA_CAFFE2_CMAKE_FLAGS" ]]; then
     EXTRA_CAFFE2_CMAKE_FLAGS=()
 fi
 
-# Detect architecture
-ARCH=$(uname -m)
-echo "Building for architecture: $ARCH"
-
 # Determine CUDA version and architectures to build for
 #
 # NOTE: We should first check `DESIRED_CUDA` when determining `CUDA_VERSION`,
@@ -57,60 +53,34 @@ fi
 cuda_version_nodot=$(echo $CUDA_VERSION | tr -d '.')
 EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
 
-# Function to remove architectures from a list
-remove_archs() {
-    local result="$1"
-    shift
-    for arch in "$@"; do
-        result="${result//${arch};/}"
-    done
-    echo "$result"
-}
-
-# Function to filter CUDA architectures for aarch64
-# aarch64 ARM GPUs only support certain compute capabilities
-# Keep: 8.0 (A100), 9.0+ (Hopper, Grace Hopper, newer)
-# Remove: < 8.0 (no ARM GPUs), 8.6 (x86_64 RTX 3090/A6000 only)
-filter_aarch64_archs() {
-    local arch_list="$1"
-    # Explicitly remove architectures not needed on aarch64
-    arch_list=$(remove_archs "$arch_list" "5.0" "6.0" "7.0" "7.5" "8.6")
-    echo "$arch_list"
-}
-
-# Base: Common architectures across all modern CUDA versions
-TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0"
-
 case ${CUDA_VERSION} in
-    12.6) TORCH_CUDA_ARCH_LIST="5.0;6.0;${TORCH_CUDA_ARCH_LIST}" ;;  # Only 12.6 includes Legacy Maxwell/Pascal that will be removed in future releases
-    12.8) TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};10.0;12.0" ;;  # +Hopper/Blackwell support
-    12.9) TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};10.0;12.0+PTX" # +Hopper/Blackwell support + PTX for forward compatibility
+    #removing sm_50-sm_60 as these architectures are deprecated in CUDA 12.8/9 and will be removed in future releases
+    #however we would like to keep sm_70 architecture see: https://github.com/pytorch/pytorch/issues/157517
+    12.8)
+        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0"
+        ;;
+    12.9)
+        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0+PTX"
+        # WAR to resolve the ld error in libtorch build with CUDA 12.9
         if [[ "$PACKAGE_TYPE" == "libtorch" ]]; then
-            TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST//7.0;/}"  # Remove 7.0 to resolve the ld error
-            TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST//8.6;/}"  # Remove 8.6 for libtorch
+            TORCH_CUDA_ARCH_LIST="7.5;8.0;9.0;10.0;12.0+PTX"
         fi
         ;;
     13.0)
-        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;$([[ "$ARCH" == "aarch64" ]] && echo "11.0;" || echo "")12.0+PTX"
-        export TORCH_NVCC_FLAGS="-compress-mode=size"
-        export BUILD_BUNDLE_PTXAS=1
+        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;12.0+PTX"
+        ;;
+    12.6)
+        TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6;9.0"
+        ;;
+    *)
+        echo "unknown cuda version $CUDA_VERSION"
+        exit 1
         ;;
-    *) echo "unknown cuda version $CUDA_VERSION"; exit 1 ;;
 esac
 
-# Filter for aarch64: Remove < 8.0 and 8.6
-[[ "$ARCH" == "aarch64" ]] && TORCH_CUDA_ARCH_LIST=$(filter_aarch64_archs "$TORCH_CUDA_ARCH_LIST")
-
-echo "TORCH_CUDA_ARCH_LIST set to: $TORCH_CUDA_ARCH_LIST"
 export TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}
 echo "${TORCH_CUDA_ARCH_LIST}"
 
-# Disable MAGMA for aarch64 as pre-built libraries are x86-64 only
-if [[ "$ARCH" == "aarch64" ]]; then
-    echo "Disabling MAGMA for aarch64 architecture"
-    export USE_MAGMA=0
-fi
-
 # Package directories
 WHEELHOUSE_DIR="wheelhouse$cuda_version_nodot"
 LIBTORCH_HOUSE_DIR="libtorch_house$cuda_version_nodot"
@@ -274,51 +244,6 @@ else
     exit 1
 fi
 
-# Add ARM-specific library dependencies
-if [[ "$ARCH" == "aarch64" ]]; then
-    echo "Adding ARM-specific library dependencies"
-
-    # ARM Compute Library (if available)
-    if [[ -d "/acl/build" ]]; then
-        echo "Adding ARM Compute Library"
-        DEPS_LIST+=(
-            "/acl/build/libarm_compute.so"
-            "/acl/build/libarm_compute_graph.so"
-        )
-        DEPS_SONAME+=(
-            "libarm_compute.so"
-            "libarm_compute_graph.so"
-        )
-    fi
-
-    # ARM system libraries
-    DEPS_LIST+=(
-        "/lib64/libgomp.so.1"
-        "/usr/lib64/libgfortran.so.5"
-    )
-    DEPS_SONAME+=(
-        "libgomp.so.1"
-        "libgfortran.so.5"
-    )
-
-    # NVPL libraries (ARM optimized BLAS/LAPACK)
-    if [[ -d "/usr/local/lib" && -f "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0" ]]; then
-        echo "Adding NVPL libraries for ARM"
-        DEPS_LIST+=(
-            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0"
-            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0"
-            "/usr/local/lib/libnvpl_lapack_core.so.0"
-            "/usr/local/lib/libnvpl_blas_core.so.0"
-        )
-        DEPS_SONAME+=(
-            "libnvpl_lapack_lp64_gomp.so.0"
-            "libnvpl_blas_lp64_gomp.so.0"
-            "libnvpl_lapack_core.so.0"
-            "libnvpl_blas_core.so.0"
-        )
-    fi
-fi
-
 # run_tests.sh requires DESIRED_CUDA to know what tests to exclude
 export DESIRED_CUDA="$cuda_version_nodot"
 
@@ -326,11 +251,9 @@ export DESIRED_CUDA="$cuda_version_nodot"
 rm -rf /usr/local/cuda || true
 ln -s "/usr/local/cuda-${CUDA_VERSION}" /usr/local/cuda
 
-# Switch `/usr/local/magma` to the desired CUDA version (skip for aarch64)
-if [[ "$ARCH" != "aarch64" ]]; then
-    rm -rf /usr/local/magma || true
-    ln -s /usr/local/cuda-${CUDA_VERSION}/magma /usr/local/magma
-fi
+# Switch `/usr/local/magma` to the desired CUDA version
+rm -rf /usr/local/magma || true
+ln -s /usr/local/cuda-${CUDA_VERSION}/magma /usr/local/magma
 
 export CUDA_VERSION=$(ls /usr/local/cuda/lib64/libcudart.so.*|sort|tac | head -1 | rev | cut -d"." -f -3 | rev) # 10.0.130
 export CUDA_VERSION_SHORT=$(ls /usr/local/cuda/lib64/libcudart.so.*|sort|tac | head -1 | rev | cut -d"." -f -3 | rev | cut -f1,2 -d".") # 10.0

.ci/onnx/common.sh

@@ -21,3 +21,87 @@ if [[ "${BUILD_ENVIRONMENT}" == *rocm* ]]; then
 fi
 
 mkdir -p "$pytest_reports_dir" || true
+
+##########################################
+# copied from .ci/pytorch/common_utils.sh
+##########################################
+
+function get_pinned_commit() {
+  cat .github/ci_commit_pins/"${1}".txt
+}
+
+function pip_install_whl() {
+  # This is used to install PyTorch and other build artifacts wheel locally
+  # without using any network connection
+
+  # Convert the input arguments into an array
+  local args=("$@")
+
+  # Check if the first argument contains multiple paths separated by spaces
+  if [[ "${args[0]}" == *" "* ]]; then
+    # Split the string by spaces into an array
+    IFS=' ' read -r -a paths <<< "${args[0]}"
+    # Loop through each path and install individually
+    for path in "${paths[@]}"; do
+      echo "Installing $path"
+      python3 -mpip install --no-index --no-deps "$path"
+    done
+  else
+    # Loop through each argument and install individually
+    for path in "${args[@]}"; do
+      echo "Installing $path"
+      python3 -mpip install --no-index --no-deps "$path"
+    done
+  fi
+}
+
+function pip_build_and_install() {
+  local build_target=$1
+  local wheel_dir=$2
+
+  local found_whl=0
+  for file in "${wheel_dir}"/*.whl
+  do
+    if [[ -f "${file}" ]]; then
+      found_whl=1
+      break
+    fi
+  done
+
+  # Build the wheel if it doesn't exist
+  if [ "${found_whl}" == "0" ]; then
+    python3 -m pip wheel \
+      --no-build-isolation \
+      --no-deps \
+      -w "${wheel_dir}" \
+      "${build_target}"
+  fi
+
+  for file in "${wheel_dir}"/*.whl
+  do
+    pip_install_whl "${file}"
+  done
+}
+
+function install_torchvision() {
+  local orig_preload
+  local commit
+  commit=$(get_pinned_commit vision)
+  orig_preload=${LD_PRELOAD}
+  if [ -n "${LD_PRELOAD}" ]; then
+    # Silence dlerror to work-around glibc ASAN bug, see https://sourceware.org/bugzilla/show_bug.cgi?id=27653#c9
+    echo 'char* dlerror(void) { return "";}'|gcc -fpic -shared -o "${HOME}/dlerror.so" -x c -
+    LD_PRELOAD=${orig_preload}:${HOME}/dlerror.so
+  fi
+
+  if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]]; then
+    # Not sure if both are needed, but why not
+    export FORCE_CUDA=1
+    export WITH_CUDA=1
+  fi
+  pip_build_and_install "git+https://github.com/pytorch/vision.git@${commit}" dist/vision
+
+  if [ -n "${LD_PRELOAD}" ]; then
+    LD_PRELOAD=${orig_preload}
+  fi
+}

.ci/onnx/test.sh

@@ -19,7 +19,7 @@ git config --global --add safe.directory /var/lib/jenkins/workspace
 
 if [[ "$BUILD_ENVIRONMENT" == *onnx* ]]; then
   # TODO: This can be removed later once vision is also part of the Docker image
-  pip install -q --no-use-pep517 "git+https://github.com/pytorch/vision.git@$(cat .github/ci_commit_pins/vision.txt)"
+  install_torchvision
   # JIT C++ extensions require ninja, so put it into PATH.
   export PATH="/var/lib/jenkins/.local/bin:$PATH"
   # NB: ONNX test is fast (~15m) so it's ok to retry it few more times to avoid any flaky issue, we

.ci/pytorch/build.sh

@@ -86,20 +86,10 @@ else
   fi
 fi
 
-# Enable MKLDNN with ARM Compute Library for ARM builds
 if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
   export USE_MKLDNN=1
-
-  # ACL is required for aarch64 builds
-  if [[ ! -d "/acl" ]]; then
-    echo "ERROR: ARM Compute Library not found at /acl"
-    echo "ACL is required for aarch64 builds. Check Docker image setup."
-    exit 1
-  fi
-
   export USE_MKLDNN_ACL=1
   export ACL_ROOT_DIR=/acl
-  echo "ARM Compute Library enabled for MKLDNN: ACL_ROOT_DIR=/acl"
 fi
 
 if [[ "$BUILD_ENVIRONMENT" == *riscv64* ]]; then

.ci/pytorch/smoke_test/check_binary_symbols.py

@@ -100,6 +100,347 @@ def check_lib_statically_linked_libstdc_cxx_abi_symbols(lib: str) -> None:
         )
 
 
+def _compile_and_extract_symbols(
+    cpp_content: str, compile_flags: list[str], exclude_list: list[str] | None = None
+) -> list[str]:
+    """
+    Helper to compile a C++ file and extract all symbols.
+
+    Args:
+        cpp_content: C++ source code to compile
+        compile_flags: Compilation flags
+        exclude_list: List of symbol names to exclude. Defaults to ["main"].
+
+    Returns:
+        List of all symbols found in the object file (excluding those in exclude_list).
+    """
+    import subprocess
+    import tempfile
+
+    if exclude_list is None:
+        exclude_list = ["main"]
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        tmppath = Path(tmpdir)
+        cpp_file = tmppath / "test.cpp"
+        obj_file = tmppath / "test.o"
+
+        cpp_file.write_text(cpp_content)
+
+        result = subprocess.run(
+            compile_flags + [str(cpp_file), "-o", str(obj_file)],
+            capture_output=True,
+            text=True,
+            timeout=60,
+        )
+
+        if result.returncode != 0:
+            raise RuntimeError(f"Compilation failed: {result.stderr}")
+
+        symbols = get_symbols(str(obj_file))
+
+        # Return all symbol names, excluding those in the exclude list
+        return [name for _addr, _stype, name in symbols if name not in exclude_list]
+
+
+def check_stable_only_symbols(install_root: Path) -> None:
+    """
+    Test TORCH_STABLE_ONLY and TORCH_TARGET_VERSION by compiling test code.
+
+    This approach tests:
+    1. WITHOUT macros -> many torch symbols exposed (compilation succeeds)
+    2. WITH TORCH_STABLE_ONLY -> compilation fails with #error directive
+    3. WITH TORCH_TARGET_VERSION -> compilation fails with #error directive
+    4. WITH both macros -> compilation fails with #error directive
+    """
+    import subprocess
+    import tempfile
+
+    include_dir = install_root / "include"
+    assert include_dir.exists(), f"Expected {include_dir} to be present"
+
+    test_cpp_content = """
+// Main torch C++ API headers
+#include <torch/torch.h>
+#include <torch/all.h>
+
+// ATen tensor library
+#include <ATen/ATen.h>
+
+// Core c10 headers (commonly used)
+#include <c10/core/Device.h>
+#include <c10/core/DeviceType.h>
+#include <c10/core/ScalarType.h>
+#include <c10/core/TensorOptions.h>
+#include <c10/util/Optional.h>
+
+int main() { return 0; }
+"""
+
+    base_compile_flags = [
+        "g++",
+        "-std=c++17",
+        f"-I{include_dir}",
+        f"-I{include_dir}/torch/csrc/api/include",
+        "-c",  # Compile only, don't link
+    ]
+
+    # Compile WITHOUT any macros - should succeed
+    symbols_without = _compile_and_extract_symbols(
+        cpp_content=test_cpp_content,
+        compile_flags=base_compile_flags,
+    )
+
+    # We expect constexpr symbols, inline functions used by other headers etc.
+    # to produce symbols
+    num_symbols_without = len(symbols_without)
+    print(f"Found {num_symbols_without} symbols without any macros defined")
+    assert num_symbols_without != 0, (
+        "Expected a non-zero number of symbols without any macros"
+    )
+
+    # Helper to verify compilation fails with expected error
+    def _expect_compilation_failure(compile_flags: list[str], macro_name: str) -> None:
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tmppath = Path(tmpdir)
+            cpp_file = tmppath / "test.cpp"
+            obj_file = tmppath / "test.o"
+
+            cpp_file.write_text(test_cpp_content)
+
+            result = subprocess.run(
+                compile_flags + [str(cpp_file), "-o", str(obj_file)],
+                capture_output=True,
+                text=True,
+                timeout=60,
+            )
+
+            if result.returncode == 0:
+                raise RuntimeError(
+                    f"Expected compilation to fail with {macro_name} defined, but it succeeded"
+                )
+
+            stderr = result.stderr
+            expected_error_msg = (
+                "This file should not be included when either TORCH_STABLE_ONLY "
+                "or TORCH_TARGET_VERSION is defined."
+            )
+
+            if expected_error_msg not in stderr:
+                raise RuntimeError(
+                    f"Expected error message to contain:\n  '{expected_error_msg}'\n"
+                    f"but got:\n{stderr[:1000]}"
+                )
+
+            print(f"Compilation correctly failed with {macro_name} defined")
+
+    compile_flags_with_stable_only = base_compile_flags + ["-DTORCH_STABLE_ONLY"]
+    _expect_compilation_failure(compile_flags_with_stable_only, "TORCH_STABLE_ONLY")
+
+    compile_flags_with_target_version = base_compile_flags + [
+        "-DTORCH_TARGET_VERSION=1"
+    ]
+    _expect_compilation_failure(
+        compile_flags_with_target_version, "TORCH_TARGET_VERSION"
+    )
+
+    compile_flags_with_both = base_compile_flags + [
+        "-DTORCH_STABLE_ONLY",
+        "-DTORCH_TARGET_VERSION=1",
+    ]
+    _expect_compilation_failure(compile_flags_with_both, "both macros")
+
+
+def check_stable_api_symbols(install_root: Path) -> None:
+    """
+    Test that stable API headers still expose symbols with TORCH_STABLE_ONLY.
+    The torch/csrc/stable/c/shim.h header is tested in check_stable_c_shim_symbols
+    """
+    include_dir = install_root / "include"
+    assert include_dir.exists(), f"Expected {include_dir} to be present"
+
+    stable_dir = include_dir / "torch" / "csrc" / "stable"
+    assert stable_dir.exists(), f"Expected {stable_dir} to be present"
+
+    stable_headers = list(stable_dir.rglob("*.h"))
+    if not stable_headers:
+        raise RuntimeError("Could not find any stable headers")
+
+    includes = []
+    for header in stable_headers:
+        rel_path = header.relative_to(include_dir)
+        includes.append(f"#include <{rel_path.as_posix()}>")
+
+    includes_str = "\n".join(includes)
+    test_stable_content = f"""
+{includes_str}
+int main() {{ return 0; }}
+"""
+
+    compile_flags = [
+        "g++",
+        "-std=c++17",
+        f"-I{include_dir}",
+        f"-I{include_dir}/torch/csrc/api/include",
+        "-c",
+        "-DTORCH_STABLE_ONLY",
+    ]
+
+    symbols_stable = _compile_and_extract_symbols(
+        cpp_content=test_stable_content,
+        compile_flags=compile_flags,
+    )
+    num_symbols_stable = len(symbols_stable)
+    print(f"Found {num_symbols_stable} symbols in torch/csrc/stable")
+    assert num_symbols_stable > 0, (
+        f"Expected stable headers to expose symbols with TORCH_STABLE_ONLY, "
+        f"but found {num_symbols_stable} symbols"
+    )
+
+
+def check_headeronly_symbols(install_root: Path) -> None:
+    """
+    Test that header-only utility headers still expose symbols with TORCH_STABLE_ONLY.
+    """
+    include_dir = install_root / "include"
+    assert include_dir.exists(), f"Expected {include_dir} to be present"
+
+    # Find all headers in torch/headeronly
+    headeronly_dir = include_dir / "torch" / "headeronly"
+    assert headeronly_dir.exists(), f"Expected {headeronly_dir} to be present"
+    headeronly_headers = list(headeronly_dir.rglob("*.h"))
+    if not headeronly_headers:
+        raise RuntimeError("Could not find any headeronly headers")
+
+    # Filter out platform-specific headers that may not compile everywhere
+    platform_specific_keywords = [
+        "cpu/vec",
+    ]
+
+    filtered_headers = []
+    for header in headeronly_headers:
+        rel_path = header.relative_to(include_dir).as_posix()
+        if not any(
+            keyword in rel_path.lower() for keyword in platform_specific_keywords
+        ):
+            filtered_headers.append(header)
+
+    includes = []
+    for header in filtered_headers:
+        rel_path = header.relative_to(include_dir)
+        includes.append(f"#include <{rel_path.as_posix()}>")
+
+    includes_str = "\n".join(includes)
+    test_headeronly_content = f"""
+{includes_str}
+int main() {{ return 0; }}
+"""
+
+    compile_flags = [
+        "g++",
+        "-std=c++17",
+        f"-I{include_dir}",
+        f"-I{include_dir}/torch/csrc/api/include",
+        "-c",
+        "-DTORCH_STABLE_ONLY",
+    ]
+
+    symbols_headeronly = _compile_and_extract_symbols(
+        cpp_content=test_headeronly_content,
+        compile_flags=compile_flags,
+    )
+    num_symbols_headeronly = len(symbols_headeronly)
+    print(f"Found {num_symbols_headeronly} symbols in torch/headeronly")
+    assert num_symbols_headeronly > 0, (
+        f"Expected headeronly headers to expose symbols with TORCH_STABLE_ONLY, "
+        f"but found {num_symbols_headeronly} symbols"
+    )
+
+
+def check_aoti_shim_symbols(install_root: Path) -> None:
+    """
+    Test that AOTI shim headers still expose symbols with TORCH_STABLE_ONLY.
+    """
+    include_dir = install_root / "include"
+    assert include_dir.exists(), f"Expected {include_dir} to be present"
+
+    # There are no constexpr symbols etc., so we need to actually use functions
+    # so that some symbols are found.
+    test_shim_content = """
+#include <torch/csrc/inductor/aoti_torch/c/shim.h>
+int main() {
+    int32_t (*fp1)() = &aoti_torch_device_type_cpu;
+    int32_t (*fp2)() = &aoti_torch_dtype_float32;
+    (void)fp1; (void)fp2;
+    return 0;
+}
+"""
+
+    compile_flags = [
+        "g++",
+        "-std=c++17",
+        f"-I{include_dir}",
+        f"-I{include_dir}/torch/csrc/api/include",
+        "-c",
+        "-DTORCH_STABLE_ONLY",
+    ]
+
+    symbols_shim = _compile_and_extract_symbols(
+        cpp_content=test_shim_content,
+        compile_flags=compile_flags,
+    )
+    num_symbols_shim = len(symbols_shim)
+    assert num_symbols_shim > 0, (
+        f"Expected shim headers to expose symbols with TORCH_STABLE_ONLY, "
+        f"but found {num_symbols_shim} symbols"
+    )
+
+
+def check_stable_c_shim_symbols(install_root: Path) -> None:
+    """
+    Test that stable C shim headers still expose symbols with TORCH_STABLE_ONLY.
+    """
+    include_dir = install_root / "include"
+    assert include_dir.exists(), f"Expected {include_dir} to be present"
+
+    # Check if the stable C shim exists
+    stable_shim = include_dir / "torch" / "csrc" / "stable" / "c" / "shim.h"
+    if not stable_shim.exists():
+        raise RuntimeError("Could not find stable c shim")
+
+    # There are no constexpr symbols etc., so we need to actually use functions
+    # so that some symbols are found.
+    test_stable_shim_content = """
+#include <torch/csrc/stable/c/shim.h>
+int main() {
+    // Reference stable C API functions to create undefined symbols
+    AOTITorchError (*fp1)(const char*, uint32_t*, int32_t*) = &torch_parse_device_string;
+    AOTITorchError (*fp2)(uint32_t*) = &torch_get_num_threads;
+    (void)fp1; (void)fp2;
+    return 0;
+}
+"""
+
+    compile_flags = [
+        "g++",
+        "-std=c++17",
+        f"-I{include_dir}",
+        f"-I{include_dir}/torch/csrc/api/include",
+        "-c",
+        "-DTORCH_STABLE_ONLY",
+    ]
+
+    symbols_stable_shim = _compile_and_extract_symbols(
+        cpp_content=test_stable_shim_content,
+        compile_flags=compile_flags,
+    )
+    num_symbols_stable_shim = len(symbols_stable_shim)
+    assert num_symbols_stable_shim > 0, (
+        f"Expected stable C shim headers to expose symbols with TORCH_STABLE_ONLY, "
+        f"but found {num_symbols_stable_shim} symbols"
+    )
+
+
 def check_lib_symbols_for_abi_correctness(lib: str) -> None:
     print(f"lib: {lib}")
     cxx11_symbols = grep_symbols(lib, LIBTORCH_CXX11_PATTERNS)
@@ -129,6 +470,13 @@ def main() -> None:
     check_lib_symbols_for_abi_correctness(libtorch_cpu_path)
     check_lib_statically_linked_libstdc_cxx_abi_symbols(libtorch_cpu_path)
 
+    # Check symbols when TORCH_STABLE_ONLY is defined
+    check_stable_only_symbols(install_root)
+    check_stable_api_symbols(install_root)
+    check_headeronly_symbols(install_root)
+    check_aoti_shim_symbols(install_root)
+    check_stable_c_shim_symbols(install_root)
+
 
 if __name__ == "__main__":
     main()

.ci/pytorch/test.sh

@@ -1250,6 +1250,97 @@ test_custom_script_ops() {
   assert_git_not_dirty
 }
 
+test_libtorch_agnostic_targetting() {
+    echo "Testing libtorch_agnostic runs correctly on TORCH_TARGET_VERSION"
+
+    REPO_DIR=$(pwd)
+    WHEEL_DIR="${REPO_DIR}/test/cpp_extensions/.wheels"
+
+    # Build wheel with current PyTorch (this has TORCH_TARGET_VERSION 2_9_0)
+    echo "Building 2.9 extension wheel with current PyTorch..."
+    pushd test/cpp_extensions/libtorch_agnostic_2_9_extension
+    time python setup.py bdist_wheel
+
+    # Save the wheel
+    mkdir -p "$WHEEL_DIR"
+    cp dist/*.whl "$WHEEL_DIR/"
+    WHEEL_FILE=$(find "$WHEEL_DIR" -maxdepth 1 -name "*.whl" -type f | head -1)
+    echo "Built wheel: $(basename "$WHEEL_FILE")"
+    popd
+
+    # Create venv and install PyTorch 2.9
+    python -m venv venv_pytorch_2_9
+    # shellcheck disable=SC1091
+    . venv_pytorch_2_9/bin/activate
+
+    # Clear PYTHONPATH to avoid using the development PyTorch
+    echo "Clearing PYTHONPATH to use only venv packages..."
+    unset PYTHONPATH
+
+    # Upgrade pip to latest version
+    echo "Upgrading pip to latest version..."
+    pip install --upgrade pip
+    pip --version
+
+    echo "Installing PyTorch 2.9..."
+
+    # Install from release channel only
+    PYTORCH_VERSION="2.9.0"
+
+    # Extract CUDA version from BUILD_ENVIRONMENT (e.g., "cuda12.1" -> "cu121")
+    if [[ "$BUILD_ENVIRONMENT" =~ cuda([0-9]+)\.([0-9]+) ]]; then
+        CUDA_MAJOR="${BASH_REMATCH[1]}"
+        CUDA_MINOR="${BASH_REMATCH[2]}"
+        CUDA_VERSION="cu${CUDA_MAJOR}${CUDA_MINOR}"
+        echo "  Detected CUDA ${CUDA_MAJOR}.${CUDA_MINOR} from BUILD_ENVIRONMENT, using ${CUDA_VERSION}"
+    else
+        # Default to CPU build
+        CUDA_VERSION="cpu"
+        echo "  No CUDA detected in BUILD_ENVIRONMENT, using CPU build"
+    fi
+
+    if pip install torch=="${PYTORCH_VERSION}" --index-url https://download.pytorch.org/whl/${CUDA_VERSION}/; then
+        echo "Installed PyTorch ${PYTORCH_VERSION} from release channel (${CUDA_VERSION})"
+    else
+        echo "  FAILED to install PyTorch 2.9.0 from release channel"
+        echo "  URL: https://download.pytorch.org/whl/${CUDA_VERSION}/"
+        deactivate
+        rm -rf venv_pytorch_2_9
+        return 1
+    fi
+
+    INSTALLED_VERSION=$(python -c "import torch; print(torch.__version__)" 2>/dev/null || echo "unknown")
+    echo "  Installed version: $INSTALLED_VERSION"
+
+    # Install test dependencies
+    echo "Installing test dependencies..."
+    pip install expecttest numpy unittest-xml-reporting
+
+    # Install the pre-built wheel
+    echo ""
+    echo "Installing pre-built 2.9 extension wheel (built with PyTorch 2.10)..."
+    pip install "$WHEEL_FILE"
+    echo "Installed $(basename "$WHEEL_FILE") into PyTorch 2.9 environment"
+
+    # Run tests with PyTorch 2.9 runtime (2.10 tests will be skipped automatically)
+    echo ""
+    echo "Running tests with PyTorch 2.9 runtime (using wheel built on PyTorch 2.10)..."
+    if time python test/cpp_extensions/test_libtorch_agnostic.py -v; then
+        echo ""
+        echo "  Wheel built with current torch and TORCH_TARGET_VERSION 2_9_0 works with PyTorch 2.9 runtime!"
+    else
+        echo "targeting test failed"
+        deactivate
+        rm -rf venv_pytorch_2_9 "$WHEEL_DIR"
+        return 1
+    fi
+
+    deactivate
+    rm -rf venv_pytorch_2_9 "$WHEEL_DIR"
+
+    assert_git_not_dirty
+}
+
 test_jit_hooks() {
   echo "Testing jit hooks in cpp"
   HOOK_BUILD="${CUSTOM_TEST_ARTIFACT_BUILD_DIR}/jit-hook-build"
@@ -1722,6 +1813,8 @@ elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]];
 elif [[ "${TEST_CONFIG}" == *backward* ]]; then
   test_forward_backward_compatibility
   # Do NOT add tests after bc check tests, see its comment.
+elif [[ "${TEST_CONFIG}" == *libtorch_agnostic_targetting* ]]; then
+  test_libtorch_agnostic_targetting
 elif [[ "${TEST_CONFIG}" == *xla* ]]; then
   install_torchvision
   build_xla

.github/labeler.yml

@@ -91,13 +91,6 @@
 "ciflow/trunk":
 - .ci/docker/ci_commit_pins/triton.txt
 
-"oncall: distributed":
-- torch/csrc/distributed/**
-- torch/distributed/**
-- torch/nn/parallel/**
-- test/distributed/**
-- torch/testing/_internal/distributed/**
-
 "release notes: distributed (checkpoint)":
 - torch/distributed/checkpoint/**
 - test/distributed/checkpoint/**

.github/workflows/_binary-build-linux.yml

@@ -260,8 +260,11 @@ jobs:
             "${DOCKER_IMAGE}"
           )
           docker exec -t -w "${PYTORCH_ROOT}" "${container_name}" bash -c "bash .circleci/scripts/binary_populate_env.sh"
-          # Unified build script for all architectures (x86_64, aarch64, s390x)
-          docker exec -t "${container_name}" bash -c "source ${BINARY_ENV_FILE} && bash /pytorch/.ci/${{ inputs.PACKAGE_TYPE }}/build.sh"
+          if [[ ${BUILD_ENVIRONMENT} == *"aarch64"* ]]; then
+            docker exec -t "${container_name}" bash -c "source ${BINARY_ENV_FILE} && bash /pytorch/.ci/aarch64_linux/aarch64_ci_build.sh"
+          else
+            docker exec -t "${container_name}" bash -c "source ${BINARY_ENV_FILE} && bash /pytorch/.ci/${{ inputs.PACKAGE_TYPE }}/build.sh"
+          fi
 
       - name: Chown artifacts
         if: ${{ steps.filter.outputs.is-test-matrix-empty == 'False' && inputs.build_environment != 'linux-s390x-binary-manywheel' }}

.github/workflows/pull.yml

@@ -70,6 +70,7 @@ jobs:
           { config: "distributed", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
           { config: "distributed", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
           { config: "numpy_2_x", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "libtorch_agnostic_targetting", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
         ]}
     secrets: inherit
 

.github/workflows/trunk.yml

@@ -83,6 +83,7 @@ jobs:
           { config: "distributed", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.12xlarge.nvidia.gpu" },
           { config: "distributed", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.12xlarge.nvidia.gpu" },
           { config: "pr_time_benchmarks", shard: 1, num_shards: 1, runner: "linux.g4dn.metal.nvidia.gpu" },
+          { config: "libtorch_agnostic_targetting", shard: 1, num_shards: 1, runner: "linux.g4dn.metal.nvidia.gpu" },
         ]}
     secrets: inherit
 

aten/src/ATen/LegacyBatchedTensorImpl.h

@@ -144,7 +144,7 @@ inline std::bitset<kVmapNumLevels> createVmapLevelsBitset(BatchDimsRef bdims) {
 }
 
 inline std::ostream& operator<<(std::ostream& out, const BatchDim& bdim) {
-  out << "(lvl=" << bdim.level() << ", dim=" << bdim.dim() << ")";
+  out << "(lvl=" << bdim.level() << ", dim=" << bdim.dim() << ')';
   return out;
 }
 

aten/src/ATen/TensorIndexing.cpp

@@ -9,7 +9,7 @@ namespace indexing {
 const EllipsisIndexType Ellipsis = EllipsisIndexType();
 
 std::ostream& operator<<(std::ostream& stream, const Slice& slice) {
-  stream << slice.start() << ":" << slice.stop() << ":" << slice.step();
+  stream << slice.start() << ':' << slice.stop() << ':' << slice.step();
   return stream;
 }
 
@@ -31,12 +31,12 @@ std::ostream& operator<<(std::ostream& stream, const TensorIndex& tensor_index)
 }
 
 std::ostream& operator<<(std::ostream& stream, const std::vector<TensorIndex>& tensor_indices) {
-  stream << "(";
+  stream << '(';
   for (const auto i : c10::irange(tensor_indices.size())) {
     stream << tensor_indices[i];
     if (i < tensor_indices.size() - 1) stream << ", ";
   }
-  stream << ")";
+  stream << ')';
   return stream;
 }
 

aten/src/ATen/TensorNames.cpp

@@ -113,7 +113,7 @@ void TensorNames::checkUnique(const char* op_name) const {
 std::ostream& operator<<(std::ostream& out, const TensorName& tensorname) {
   out << tensorname.name_ << " (index ";
   out << tensorname.origin_idx_ << " of ";
-  out << tensorname.origin_ << ")";
+  out << tensorname.origin_ << ')';
   return out;
 }
 

aten/src/ATen/TensorUtils.cpp

@@ -13,9 +13,9 @@ std::ostream& operator<<(std::ostream & out, const TensorGeometryArg& t) {
   if (t.pos == 0) {
     // 0 is distinguished; it usually indicates 'self' or the return
     // tensor
-    out << "'" << t.name << "'";
+    out << '\'' << t.name << '\'';
   } else {
-    out << "argument #" << t.pos << " '" << t.name << "'";
+    out << "argument #" << t.pos << " '" << t.name << '\'';
   }
   return out;
 }
@@ -154,7 +154,7 @@ void checkSameGPU(CheckedFrom c, const TensorArg& t1, const TensorArg& t2) {
       oss << "Tensor for " << t2 << " is on CPU, ";
     }
     oss << "but expected " << ((!t1->is_cpu() && !t2->is_cpu()) ? "them" : "it")
-        << " to be on GPU (while checking arguments for " << c << ")";
+        << " to be on GPU (while checking arguments for " << c << ')';
     TORCH_CHECK(false, oss.str());
   }
   TORCH_CHECK(
@@ -199,7 +199,7 @@ void checkScalarTypes(CheckedFrom c, const TensorArg& t,
         i++;
       }
       oss << "; but got " << t->toString()
-          << " instead (while checking arguments for " << c << ")";
+          << " instead (while checking arguments for " << c << ')';
       TORCH_CHECK(false, oss.str());
     }
 }

aten/src/ATen/Version.cpp

@@ -43,8 +43,8 @@ std::string get_mkldnn_version() {
     // https://github.com/intel/ideep/issues/29
     {
       const dnnl_version_t* ver = dnnl_version();
-      ss << "Intel(R) MKL-DNN v" << ver->major << "." << ver->minor << "." << ver->patch
-         << " (Git Hash " << ver->hash << ")";
+      ss << "Intel(R) MKL-DNN v" << ver->major << '.' << ver->minor << '.' << ver->patch
+         << " (Git Hash " << ver->hash << ')';
     }
   #else
     ss << "MKLDNN not found";
@@ -81,7 +81,7 @@ std::string get_openmp_version() {
           break;
       }
       if (ver_str) {
-        ss << " (a.k.a. OpenMP " << ver_str << ")";
+        ss << " (a.k.a. OpenMP " << ver_str << ')';
       }
     }
   #else
@@ -135,38 +135,38 @@ std::string show_config() {
 
 #if defined(__GNUC__)
   {
-    ss << "  - GCC " << __GNUC__ << "." << __GNUC_MINOR__ << "\n";
+    ss << "  - GCC " << __GNUC__ << '.' << __GNUC_MINOR__ << '\n';
   }
 #endif
 
 #if defined(__cplusplus)
   {
-    ss << "  - C++ Version: " << __cplusplus << "\n";
+    ss << "  - C++ Version: " << __cplusplus << '\n';
   }
 #endif
 
 #if defined(__clang_major__)
   {
-    ss << "  - clang " << __clang_major__ << "." << __clang_minor__ << "." << __clang_patchlevel__ << "\n";
+    ss << "  - clang " << __clang_major__ << '.' << __clang_minor__ << '.' << __clang_patchlevel__ << '\n';
   }
 #endif
 
 #if defined(_MSC_VER)
   {
-    ss << "  - MSVC " << _MSC_FULL_VER << "\n";
+    ss << "  - MSVC " << _MSC_FULL_VER << '\n';
   }
 #endif
 
 #if AT_MKL_ENABLED()
-  ss << "  - " << get_mkl_version() << "\n";
+  ss << "  - " << get_mkl_version() << '\n';
 #endif
 
 #if AT_MKLDNN_ENABLED()
-  ss << "  - " << get_mkldnn_version() << "\n";
+  ss << "  - " << get_mkldnn_version() << '\n';
 #endif
 
 #ifdef _OPENMP
-  ss << "  - " << get_openmp_version() << "\n";
+  ss << "  - " << get_openmp_version() << '\n';
 #endif
 
 #if AT_BUILD_WITH_LAPACK()
@@ -183,7 +183,7 @@ std::string show_config() {
   ss << "  - Cross compiling on MacOSX\n";
 #endif
 
-  ss << "  - "<< used_cpu_capability() << "\n";
+  ss << "  - "<< used_cpu_capability() << '\n';
 
   if (hasCUDA()) {
     ss << detail::getCUDAHooks().showConfig();
@@ -200,10 +200,10 @@ std::string show_config() {
   ss << "  - Build settings: ";
   for (const auto& pair : caffe2::GetBuildOptions()) {
     if (!pair.second.empty()) {
-      ss << pair.first << "=" << pair.second << ", ";
+      ss << pair.first << '=' << pair.second << ", ";
     }
   }
-  ss << "\n";
+  ss << '\n';
 
   // TODO: do HIP
   // TODO: do XLA

aten/src/ATen/code_template.h

@@ -209,7 +209,7 @@ struct CodeTemplate {
   // to indent correctly in the context.
   void emitIndent(std::ostream& out, size_t indent) const {
     for ([[maybe_unused]] const auto i : c10::irange(indent)) {
-      out << " ";
+      out << ' ';
     }
   }
   void emitStringWithIndents(

aten/src/ATen/core/Dimname.cpp

@@ -10,7 +10,7 @@ std::ostream& operator<<(std::ostream& out, const Dimname& dimname) {
   if (dimname.type() == NameType::WILDCARD) {
     out << "None";
   } else {
-    out << "'" << dimname.symbol().toUnqualString() << "'";
+    out << '\'' << dimname.symbol().toUnqualString() << '\'';
   }
   return out;
 }

aten/src/ATen/core/Range.cpp

@@ -5,7 +5,7 @@
 namespace at {
 
 std::ostream& operator<<(std::ostream& out, const Range& range) {
-  out << "Range[" << range.begin << ", " << range.end << "]";
+  out << "Range[" << range.begin << ", " << range.end << ']';
   return out;
 }
 

aten/src/ATen/core/Tensor.cpp

@@ -71,7 +71,7 @@ void TensorBase::enforce_invariants() {
 
 void TensorBase::print() const {
   if (defined()) {
-    std::cerr << "[" << toString() << " " << sizes() << "]" << '\n';
+    std::cerr << '[' << toString() << ' ' << sizes() << ']' << '\n';
   } else {
     std::cerr << "[UndefinedTensor]" << '\n';
   }

aten/src/ATen/core/Vitals.cpp

@@ -9,8 +9,8 @@ APIVitals VitalsAPI;
 
 std::ostream& operator<<(std::ostream& os, TorchVital const& tv) {
   for (const auto& m : tv.attrs) {
-    os << "[TORCH_VITAL] " << tv.name << "." << m.first << "\t\t "
-       << m.second.value << "\n";
+    os << "[TORCH_VITAL] " << tv.name << '.' << m.first << "\t\t "
+       << m.second.value << '\n';
   }
   return os;
 }

aten/src/ATen/core/alias_info.h

@@ -100,18 +100,18 @@ inline bool operator==(const AliasInfo& lhs, const AliasInfo& rhs) {
 
 // this does match the way things are represented in the schema
 inline std::ostream& operator<<(std::ostream& out, const AliasInfo& aliasInfo) {
-  out << "(";
+  out << '(';
   bool first = true;
   for (const auto& set : aliasInfo.beforeSets()) {
     if (first) {
       first = false;
     } else {
-      out << "|";
+      out << '|';
     }
     out << set.toUnqualString();
   }
   if (aliasInfo.isWrite()) {
-    out << "!";
+    out << '!';
   }
   if (aliasInfo.beforeSets() != aliasInfo.afterSets()) {
     out << " -> ";
@@ -120,12 +120,12 @@ inline std::ostream& operator<<(std::ostream& out, const AliasInfo& aliasInfo) {
       if (first) {
         first = false;
       } else {
-        out << "|";
+        out << '|';
       }
       out << set.toUnqualString();
     }
   }
-  out << ")";
+  out << ')';
   return out;
 }
 } // namespace c10

aten/src/ATen/core/blob.h

@@ -198,7 +198,7 @@ inline void swap(Blob& lhs, Blob& rhs)  noexcept {
 }
 
 inline std::ostream& operator<<(std::ostream& out, const Blob& v) {
-  return out << "Blob[" << v.TypeName() << "]";
+  return out << "Blob[" << v.TypeName() << ']';
 }
 
 } // namespace caffe2

aten/src/ATen/core/class_type.cpp

@@ -456,8 +456,8 @@ bool ClassType::isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const {
           *why_not << "Method on class '" << repr_str()
                    << "' (1) is not compatible with interface '"
                    << rhs.repr_str() << "' (2)\n"
-                   << "  (1) " << self_method->getSchema() << "\n"
-                   << "  (2) " << schema << "\n";
+                   << "  (1) " << self_method->getSchema() << '\n'
+                   << "  (2) " << schema << '\n';
         }
         return false;
       }

aten/src/ATen/core/class_type.h

@@ -100,7 +100,7 @@ struct TORCH_API ClassType : public NamedType {
   std::string repr_str() const override {
     std::stringstream ss;
     ss << str()
-       << " (of Python compilation unit at: " << compilation_unit().get() << ")";
+       << " (of Python compilation unit at: " << compilation_unit().get() << ')';
     return ss.str();
   }
 

aten/src/ATen/core/dispatch/DispatchKeyExtractor.cpp

@@ -58,12 +58,12 @@ std::string DispatchKeyExtractor::dumpState() const {
   std::ostringstream oss;
   for (const auto i : c10::irange(c10::utils::bitset::NUM_BITS())) {
     if (dispatch_arg_indices_reverse_.get(i)) {
-      oss << "1";
+      oss << '1';
     } else {
-      oss << "0";
+      oss << '0';
     }
   }
-  oss << " " << nonFallthroughKeys_ << "\n";
+  oss << ' ' << nonFallthroughKeys_ << '\n';
   return oss.str();
 }
 

aten/src/ATen/core/dispatch/Dispatcher.cpp

@@ -69,8 +69,8 @@ private:
 
 void _print_dispatch_trace(const std::string& label, const std::string& op_name, const DispatchKeySet& dispatchKeySet) {
   auto nesting_value = dispatch_trace_nesting_value();
-  for (int64_t i = 0; i < nesting_value; ++i) std::cerr << " ";
-  std::cerr << label << " op=[" << op_name << "], key=[" << toString(dispatchKeySet.highestPriorityTypeId()) << "]" << std::endl;
+  for (int64_t i = 0; i < nesting_value; ++i) std::cerr << ' ';
+  std::cerr << label << " op=[" << op_name << "], key=[" << toString(dispatchKeySet.highestPriorityTypeId()) << ']' << std::endl;
 }
 } // namespace detail
 

aten/src/ATen/core/dispatch/OperatorEntry.cpp

@@ -570,7 +570,7 @@ void OperatorEntry::checkInvariants() const {
 
 std::string OperatorEntry::listAllDispatchKeys() const {
   std::ostringstream str;
-  str << "[";
+  str << '[';
 
   bool has_kernels = false;
   for (auto k : allDispatchKeysInFullSet()) {
@@ -584,7 +584,7 @@ std::string OperatorEntry::listAllDispatchKeys() const {
     str << k;
     has_kernels = true;
   }
-  str << "]";
+  str << ']';
   return str.str();
 }
 
@@ -683,12 +683,12 @@ void OperatorEntry::setReportErrorCallback_(std::unique_ptr<c10::SafePyObject> c
 // This WON'T report backend fallbacks.
 std::string OperatorEntry::dumpState() const {
   std::ostringstream oss;
-  oss << "name: " << name_ << "\n";
+  oss << "name: " << name_ << '\n';
   if (schema_) {
-    oss << "schema: " << schema_->schema << "\n";
-    oss << "debug: " << schema_->debug << "\n";
+    oss << "schema: " << schema_->schema << '\n';
+    oss << "debug: " << schema_->debug << '\n';
     oss << "alias analysis kind: " << toString(schema_->schema.aliasAnalysis())
-        << (schema_->schema.isDefaultAliasAnalysisKind() ? " (default)" : "") << "\n";
+        << (schema_->schema.isDefaultAliasAnalysisKind() ? " (default)" : "") << '\n';
   } else {
     oss << "schema: (none)\n";
   }

aten/src/ATen/core/function_schema.cpp

@@ -7,7 +7,7 @@
 namespace c10 {
 
 void FunctionSchema::dump() const {
-  std::cout << *this << "\n";
+  std::cout << *this << '\n';
 }
 
 const std::vector<Argument>& FunctionSchema::getCorrectList(SchemaArgType type) const {
@@ -210,9 +210,9 @@ std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema) {
 
   out << schema.name();
   if (!schema.overload_name().empty()) {
-    out << "." << schema.overload_name();
+    out << '.' << schema.overload_name();
   }
-  out << "(";
+  out << '(';
 
   bool seen_kwarg_only = false;
   for (const auto i : c10::irange(schema.arguments().size())) {
@@ -273,7 +273,7 @@ std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema) {
   }
 
   if (need_paren) {
-    out << "(";
+    out << '(';
   }
   for (const auto i : c10::irange(returns.size())) {
     if (i > 0) {
@@ -288,7 +288,7 @@ std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema) {
     out << "...";
   }
   if (need_paren) {
-    out << ")";
+    out << ')';
   }
   return out;
 }
@@ -471,7 +471,7 @@ bool FunctionSchema::isForwardCompatibleWith(
     if (!arguments().at(i).isForwardCompatibleWith(old.arguments().at(i))) {
       if (why_not) {
         why_not
-            << "'" << arguments().at(i).name() << "'"
+            << '\'' << arguments().at(i).name() << '\''
             << " is not forward compatible with the older version of the schema";
       }
       return false;
@@ -511,7 +511,7 @@ bool FunctionSchema::isForwardCompatibleWith(
              .isForwardCompatibleWith(old.arguments().at(i))) {
       if (why_not) {
         why_not << "Out argument '"
-                << "'" << arguments().at(i).name()
+                << '\'' << arguments().at(i).name()
                 << " is not FC with the older version of the schema";
       }
       return false;

aten/src/ATen/core/function_schema.h

@@ -571,7 +571,7 @@ inline std::ostream& operator<<(std::ostream& out, const Argument& arg) {
     if (arg.N()) {
         N = std::to_string(*arg.N());
     }
-    out << "[" << N << "]";
+    out << '[' << N << ']';
   } else {
     out << unopt_type->str();
   }
@@ -582,15 +582,15 @@ inline std::ostream& operator<<(std::ostream& out, const Argument& arg) {
   }
 
   if (is_opt) {
-    out << "?";
+    out << '?';
   }
 
   if (!arg.name().empty()) {
-    out << " " << arg.name();
+    out << ' ' << arg.name();
   }
 
   if (arg.default_value()) {
-    out << "=";
+    out << '=';
     if ((type->kind() == c10::TypeKind::StringType ||
         unopt_type->kind() == c10::TypeKind::StringType) &&
         arg.default_value().value().isString()) {

aten/src/ATen/core/ivalue.cpp

@@ -66,7 +66,7 @@ bool operator==(const ivalue::Tuple& lhs, const ivalue::Tuple& rhs) {
 }
 
 std::ostream& operator<<(std::ostream& out, const ivalue::EnumHolder& v) {
-  out << v.qualifiedClassName() << "." << v.name();
+  out << v.qualifiedClassName() << '.' << v.name();
   return out;
 }
 
@@ -526,7 +526,7 @@ std::ostream& printMaybeAnnotatedList(
       !elementTypeCanBeInferredFromMembers(list_elem_type)) {
     out << "annotate(" << the_list.type<c10::Type>()->annotation_str() << ", ";
     printList(out, the_list.toListRef(), "[", "]", formatter);
-    out << ")";
+    out << ')';
     return out;
   } else {
     return printList(out, the_list.toListRef(), "[", "]", formatter);
@@ -538,7 +538,7 @@ std::ostream& printDict(
     std::ostream& out,
     const Dict& v,
     const IValueFormatter& formatter) {
-  out << "{";
+  out << '{';
 
   bool first = true;
   for (const auto& pair : v) {
@@ -552,7 +552,7 @@ std::ostream& printDict(
     first = false;
   }
 
-  out << "}";
+  out << '}';
   return out;
 }
 }
@@ -565,8 +565,8 @@ static std::ostream& printMaybeAnnotatedDict(
   auto value_type = the_dict.type()->castRaw<DictType>()->getValueType();
   if (the_dict.toGenericDict().empty() ||
       !elementTypeCanBeInferredFromMembers(value_type)) {
-    out << "annotate(" << the_dict.type<c10::Type>()->annotation_str() << ",";
-    printDict(out, the_dict.toGenericDict(), formatter) << ")";
+    out << "annotate(" << the_dict.type<c10::Type>()->annotation_str() << ',';
+    printDict(out, the_dict.toGenericDict(), formatter) << ')';
   } else {
     return printDict(out, the_dict.toGenericDict(), formatter);
   }
@@ -577,7 +577,7 @@ static std::ostream& printComplex(std::ostream & out, const IValue & v) {
   c10::complex<double> d = v.toComplexDouble();
   IValue real(d.real()), imag(std::abs(d.imag()));
   auto sign = d.imag() >= 0 ? '+' : '-';
-  return out << real << sign << imag << "j";
+  return out << real << sign << imag << 'j';
 }
 
 std::ostream& IValue::repr(
@@ -605,9 +605,9 @@ std::ostream& IValue::repr(
         if (static_cast<double>(i) == d) {
           // -0.0 (signed zero) needs to be parsed as -0.
           if (i == 0 && std::signbit(d)) {
-            return out << "-" << i << ".";
+            return out << '-' << i << '.';
           }
-          return out << i << ".";
+          return out << i << '.';
         }
       }
       auto orig_prec = out.precision();
@@ -643,20 +643,20 @@ std::ostream& IValue::repr(
       device_stream << v.toDevice();
       out << "torch.device(";
       c10::printQuotedString(out, device_stream.str());
-      return out << ")";
+      return out << ')';
     }
     case IValue::Tag::Generator: {
       auto generator = v.toGenerator();
       out << "torch.Generator(device=";
       c10::printQuotedString(out, generator.device().str());
-      out << ", seed=" << generator.current_seed() << ")";
+      out << ", seed=" << generator.current_seed() << ')';
       return out;
     }
     case IValue::Tag::GenericDict:
       return printMaybeAnnotatedDict(out, v, formatter);
     case IValue::Tag::Enum: {
       auto enum_holder = v.toEnumHolder();
-      return out << enum_holder->qualifiedClassName() << "." <<
+      return out << enum_holder->qualifiedClassName() << '.' <<
           enum_holder->name();
     }
     case IValue::Tag::Object: {
@@ -801,7 +801,7 @@ std::ostream& operator<<(std::ostream & out, const IValue & v) {
       if (c == FP_NORMAL || c == FP_ZERO) {
         int64_t i = static_cast<int64_t>(d);
         if (static_cast<double>(i) == d) {
-          return out << i << ".";
+          return out << i << '.';
         }
       }
       auto orig_prec = out.precision();
@@ -852,7 +852,7 @@ std::ostream& operator<<(std::ostream & out, const IValue & v) {
       return printDict(out, v.toGenericDict(), formatter);
     case IValue::Tag::PyObject: {
       auto py_obj = v.toPyObject();
-      return out << "<PyObject at" << py_obj << ">";
+      return out << "<PyObject at" << py_obj << '>';
     }
     case IValue::Tag::Generator:
       return out << "Generator";
@@ -862,22 +862,22 @@ std::ostream& operator<<(std::ostream & out, const IValue & v) {
       // TODO we should attempt to call __str__ if the object defines it.
       auto obj = v.toObject();
       // print this out the way python would do it
-      return out << "<" << obj->name() << " object at " << obj.get() << ">";
+      return out << '<' << obj->name() << " object at " << obj.get() << '>';
     }
     case IValue::Tag::Enum: {
       auto enum_holder = v.toEnumHolder();
-      return out << "Enum<" << enum_holder->unqualifiedClassName() << "." <<
-          enum_holder->name() << ">";
+      return out << "Enum<" << enum_holder->unqualifiedClassName() << '.' <<
+          enum_holder->name() << '>';
     }
 
   }
-  return out << "<Invalid IValue tag=" << std::to_string(static_cast<uint32_t>(v.tag)) << ">";
+  return out << "<Invalid IValue tag=" << std::to_string(static_cast<uint32_t>(v.tag)) << '>';
 }
 
 #undef TORCH_FORALL_TAGS
 
 void IValue::dump() const {
-  std::cout << *this << "\n";
+  std::cout << *this << '\n';
 }
 
 std::shared_ptr<ClassType> ivalue::Object::type() const {
@@ -1050,7 +1050,7 @@ c10::intrusive_ptr<ivalue::Object> ivalue::Object::deepcopy(
       std::stringstream err;
       err << "Cannot serialize custom bound C++ class";
       if (auto qualname = type()->name()) {
-        err << " " << qualname->qualifiedName();
+        err << ' ' << qualname->qualifiedName();
       }
       err << ". Please define serialization methods via def_pickle() for "
             "this class.";

aten/src/ATen/core/jit_type.h

@@ -211,7 +211,7 @@ struct TORCH_API OptionalType : public UnionType {
 
   std::string str() const override {
     std::stringstream ss;
-    ss << getElementType()->str() << "?";
+    ss << getElementType()->str() << '?';
     return ss.str();
   }
 
@@ -240,7 +240,7 @@ struct TORCH_API OptionalType : public UnionType {
 
   std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override {
     std::stringstream ss;
-    ss << "Optional[" << getElementType()->annotation_str(printer) << "]";
+    ss << "Optional[" << getElementType()->annotation_str(printer) << ']';
     return ss.str();
   }
 };
@@ -906,7 +906,7 @@ struct TORCH_API ListType
 
   std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override {
     std::stringstream ss;
-    ss << "List[" << getElementType()->annotation_str(printer) << "]";
+    ss << "List[" << getElementType()->annotation_str(printer) << ']';
     return ss.str();
   }
 };
@@ -946,7 +946,7 @@ struct TORCH_API DictType : public SharedType {
   std::string str() const override {
     std::stringstream ss;
     ss << "Dict(" << getKeyType()->str() << ", " << getValueType()->str()
-       << ")";
+       << ')';
     return ss.str();
   }
 
@@ -1018,7 +1018,7 @@ struct TORCH_API FutureType
 
   std::string str() const override {
     std::stringstream ss;
-    ss << "Future(" << getElementType()->str() << ")";
+    ss << "Future(" << getElementType()->str() << ')';
     return ss.str();
   }
   TypePtr createWithContained(
@@ -1041,7 +1041,7 @@ struct TORCH_API FutureType
 
   std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override {
     std::stringstream ss;
-    ss << "Future[" << getElementType()->annotation_str(printer) << "]";
+    ss << "Future[" << getElementType()->annotation_str(printer) << ']';
     return ss.str();
   }
 };
@@ -1060,7 +1060,7 @@ struct TORCH_API AwaitType
 
   std::string str() const override {
     std::stringstream ss;
-    ss << "Await(" << getElementType()->str() << ")";
+    ss << "Await(" << getElementType()->str() << ')';
     return ss.str();
   }
   TypePtr createWithContained(
@@ -1083,7 +1083,7 @@ struct TORCH_API AwaitType
 
   std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override {
     std::stringstream ss;
-    ss << "Await[" << getElementType()->annotation_str(printer) << "]";
+    ss << "Await[" << getElementType()->annotation_str(printer) << ']';
     return ss.str();
   }
 };
@@ -1102,7 +1102,7 @@ struct TORCH_API RRefType
 
   std::string str() const override {
     std::stringstream ss;
-    ss << "RRef(" << getElementType()->str() << ")";
+    ss << "RRef(" << getElementType()->str() << ')';
     return ss.str();
   }
   TypePtr createWithContained(
@@ -1115,7 +1115,7 @@ struct TORCH_API RRefType
 
   std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override {
     std::stringstream ss;
-    ss << "RRef[" << getElementType()->annotation_str(printer) << "]";
+    ss << "RRef[" << getElementType()->annotation_str(printer) << ']';
     return ss.str();
   }
 };

aten/src/ATen/core/operator_name.cpp

@@ -11,7 +11,7 @@ std::string toString(const OperatorName& opName) {
 std::ostream& operator<<(std::ostream& os, const OperatorName& opName) {
   os << opName.name;
   if (!opName.overload_name.empty()) {
-    os << "." << opName.overload_name;
+    os << '.' << opName.overload_name;
   }
   return os;
 }

aten/src/ATen/core/tensor_type.cpp

@@ -65,7 +65,7 @@ VaryingShape<T> VaryingShape<T>::merge(const VaryingShape<T>& other) const {
 
 template <typename T>
 std::ostream& operator<<(std::ostream& out, const VaryingShape<T>& vs) {
-  out << "(";
+  out << '(';
   if (!vs.size()) {
     out << "*)";
     return out;
@@ -79,10 +79,10 @@ std::ostream& operator<<(std::ostream& out, const VaryingShape<T>& vs) {
     if (v.has_value()) {
       out << v.value();
     } else {
-      out << "*";
+      out << '*';
     }
   }
-  out << ")";
+  out << ')';
   return out;
 }
 
@@ -105,7 +105,7 @@ std::ostream& operator<<(
   }
   auto sizes_opt = ss.sizes();
 
-  os << "(";
+  os << '(';
   for (size_t i = 0; i < rank_opt.value(); i++) {
     if (i > 0) {
       os << ", ";
@@ -113,10 +113,10 @@ std::ostream& operator<<(
     if(sizes_opt.has_value() && sizes_opt.value()[i].is_static()) {
       os << sizes_opt.value()[i];
     } else {
-      os << "*";
+      os << '*';
     }
   }
-  os << ")";
+  os << ')';
 
   return os;
 }
@@ -131,17 +131,17 @@ std::ostream& operator<<(std::ostream& os, const ShapeSymbol& s) {
 }
 
 std::ostream& operator<<(std::ostream& os, const Stride& s) {
-  os << "{";
+  os << '{';
   if (s.stride_index_.has_value()) {
     os << *s.stride_index_;
   } else {
-    os << "*";
+    os << '*';
   }
-  os << ":";
+  os << ':';
   if (s.stride_.has_value()) {
     os << *s.stride_;
   } else {
-    os << "*";
+    os << '*';
   }
   os << '}';
   return os;

aten/src/ATen/core/type.cpp

@@ -67,7 +67,7 @@ std::ostream& operator<<(std::ostream & out, const Type & t) {
       bool has_valid_strides_info = ndim > 0 &&
           value->strides().isComplete() && value->strides().size() == ndim;
 
-      out << "(";
+      out << '(';
       size_t i = 0;
       bool symbolic = type_verbosity() == TypeVerbosity::Symbolic;
       for (i = 0; i < *ndim; ++i) {
@@ -79,7 +79,7 @@ std::ostream& operator<<(std::ostream & out, const Type & t) {
         } else if (symbolic) {
           out << value->symbolic_sizes().at(i);
         } else {
-          out << "*";
+          out << '*';
         }
       }
       if (has_valid_strides_info &&
@@ -91,7 +91,7 @@ std::ostream& operator<<(std::ostream & out, const Type & t) {
           }
           out << value->strides()[i].value();
         }
-        out << "]";
+        out << ']';
       }
       if (type_verbosity() >= TypeVerbosity::Full) {
         if (value->requiresGrad()) {
@@ -107,12 +107,12 @@ std::ostream& operator<<(std::ostream & out, const Type & t) {
           out << "device=" << *value->device();
         }
       }
-      out << ")";
+      out << ')';
     } else {
       if (type_verbosity() >= TypeVerbosity::Full) {
         size_t i = 0;
         if (value->requiresGrad()) {
-          out << "("
+          out << '('
               << "requires_grad=" << *value->requiresGrad();
           i++;
         }
@@ -120,7 +120,7 @@ std::ostream& operator<<(std::ostream & out, const Type & t) {
           out << ((i++ > 0) ? ", " : "(") << "device=" << *value->device();
         }
         if (i > 0) {
-          out << ")";
+          out << ')';
         }
       }
     }
@@ -133,18 +133,18 @@ std::ostream& operator<<(std::ostream & out, const Type & t) {
     out << *prim << "[]";
   } else if (t.kind() == TypeKind::OptionalType) {
     auto prim = t.castRaw<OptionalType>()->getElementType();
-    out << *prim << "?";
+    out << *prim << '?';
   } else if(t.kind() == TypeKind::FutureType) {
     auto elem = t.castRaw<FutureType>()->getElementType();
-    out << "Future[" << *elem << "]";
+    out << "Future[" << *elem << ']';
   } else if(t.kind() == TypeKind::RRefType) {
     auto elem = t.castRaw<RRefType>()->getElementType();
-    out << "RRef[" << *elem << "]";
+    out << "RRef[" << *elem << ']';
   } else if(auto tup = t.cast<TupleType>()) {
     if (tup->schema()) {
       out << "NamedTuple";
     }
-    out << "(";
+    out << '(';
     for(size_t i = 0; i < tup->elements().size(); ++i) {
       if(i > 0)
         out << ", ";
@@ -160,7 +160,7 @@ std::ostream& operator<<(std::ostream & out, const Type & t) {
         out << *(tup->elements()[i]);
       }
     }
-    out << ")";
+    out << ')';
   } else if (t.kind() == TypeKind::FunctionType) {
     out << "Function";
   } else {
@@ -475,7 +475,7 @@ std::optional<TypePtr> unifyTypeList(
       why_not << "Could not unify type list since element " << i << " of type "
               << elements.at(i)->repr_str()
               << " did not match the types before it ("
-              << ret_type->repr_str() << ")";
+              << ret_type->repr_str() << ')';
       return std::nullopt;
     }
     ret_type = *maybe_unified;
@@ -907,13 +907,13 @@ std::string TupleType::str() const {
     // NOLINTNEXTLINE(bugprone-unchecked-optional-access)
     ss << name()->qualifiedName();
   } else {
-    ss << "(";
+    ss << '(';
     for(size_t i = 0; i < elements().size(); ++i) {
       if(i > 0)
         ss << ", ";
       ss << elements()[i]->str();
     }
-    ss << ")";
+    ss << ')';
   }
   return ss.str();
 }
@@ -1003,8 +1003,8 @@ bool InterfaceType::isSubTypeImpl(
           *why_not << "Method on interface '" << lhs.repr_str()
                    << "' (1) is not compatible with interface '"
                    << rhs.repr_str() << "' (2)\n"
-                   << "  (1) " << *self_schema << "\n"
-                   << "  (2) " << schema << "\n";
+                   << "  (1) " << *self_schema << '\n'
+                   << "  (2) " << schema << '\n';
           return false;
         }
         return false;
@@ -1078,7 +1078,7 @@ SymbolicShape SymbolicShape::merge(const SymbolicShape& other) const {
 }
 
 void SymbolicShape::dump() const {
-  std::cout << *this << "\n";
+  std::cout << *this << '\n';
 }
 
 bool EnumType::isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const {

aten/src/ATen/core/union_type.cpp

@@ -205,9 +205,9 @@ UnionType::UnionType(std::vector<TypePtr> reference, TypeKind kind) : SharedType
     for (const auto i : c10::irange(reference.size())) {
       msg << reference[i]->repr_str();
       if (i > 0) {
-        msg << ",";
+        msg << ',';
       }
-      msg << " ";
+      msg << ' ';
     }
     msg << "} has the single type " << types_[0]->repr_str()
          << ". Use the common supertype instead of creating a Union"

aten/src/ATen/cpu/vec/vec256/vec256.h

@@ -80,7 +80,7 @@ std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
     }
     stream << buf[i];
   }
-  stream << "]";
+  stream << ']';
   return stream;
 }
 

aten/src/ATen/cpu/vec/vec512/vec512.h

@@ -55,7 +55,7 @@ std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
     }
     stream << buf[i];
   }
-  stream << "]";
+  stream << ']';
   return stream;
 }
 

aten/src/ATen/cuda/detail/CUDAHooks.cpp

@@ -411,16 +411,16 @@ std::string CUDAHooks::showConfig() const {
     // HIP_VERSION value format was changed after ROCm v4.2 to include the patch number
     if(v < 500) {
       // If major=xx, minor=yy then format -> xxyy
-      oss << (v / 100) << "." << (v % 10);
+      oss << (v / 100) << '.' << (v % 10);
     }
     else {
       // If major=xx, minor=yy & patch=zzzzz then format -> xxyyzzzzz
-      oss << (v / 10000000) << "." << (v / 100000 % 100) << "." << (v % 100000);
+      oss << (v / 10000000) << '.' << (v / 100000 % 100) << '.' << (v % 100000);
     }
 #else
-    oss << (v / 1000) << "." << (v / 10 % 100);
+    oss << (v / 1000) << '.' << (v / 10 % 100);
     if (v % 10 != 0) {
-      oss << "." << (v % 10);
+      oss << '.' << (v % 10);
     }
 #endif
   };
@@ -431,16 +431,16 @@ std::string CUDAHooks::showConfig() const {
   oss << "  - HIP Runtime ";
 #endif
   printCudaStyleVersion(runtimeVersion);
-  oss << "\n";
+  oss << '\n';
 
   // TODO: Make HIPIFY understand CUDART_VERSION macro
 #if !defined(USE_ROCM)
   if (runtimeVersion != CUDART_VERSION) {
     oss << "  - Built with CUDA Runtime ";
     printCudaStyleVersion(CUDART_VERSION);
-    oss << "\n";
+    oss << '\n';
   }
-  oss << "  - NVCC architecture flags: " << NVCC_FLAGS_EXTRA << "\n";
+  oss << "  - NVCC architecture flags: " << NVCC_FLAGS_EXTRA << '\n';
 #endif
 
 #if !defined(USE_ROCM)
@@ -448,9 +448,9 @@ std::string CUDAHooks::showConfig() const {
 
 
   auto printCudnnStyleVersion = [&](size_t v) {
-    oss << (v / 1000) << "." << (v / 100 % 10);
+    oss << (v / 1000) << '.' << (v / 100 % 10);
     if (v % 100 != 0) {
-      oss << "." << (v % 100);
+      oss << '.' << (v % 100);
     }
   };
 
@@ -461,22 +461,22 @@ std::string CUDAHooks::showConfig() const {
   if (cudnnCudartVersion != CUDART_VERSION) {
     oss << "  (built against CUDA ";
     printCudaStyleVersion(cudnnCudartVersion);
-    oss << ")";
+    oss << ')';
   }
-  oss << "\n";
+  oss << '\n';
   if (cudnnVersion != CUDNN_VERSION) {
     oss << "    - Built with CuDNN ";
     printCudnnStyleVersion(CUDNN_VERSION);
-    oss << "\n";
+    oss << '\n';
   }
 #endif
 #else
   // TODO: Check if miopen has the functions above and unify
-  oss << "  - MIOpen " << MIOPEN_VERSION_MAJOR << "." << MIOPEN_VERSION_MINOR << "." << MIOPEN_VERSION_PATCH << "\n";
+  oss << "  - MIOpen " << MIOPEN_VERSION_MAJOR << '.' << MIOPEN_VERSION_MINOR << '.' << MIOPEN_VERSION_PATCH << '\n';
 #endif
 
 #if AT_MAGMA_ENABLED()
-  oss << "  - Magma " << MAGMA_VERSION_MAJOR << "." << MAGMA_VERSION_MINOR << "." << MAGMA_VERSION_MICRO << "\n";
+  oss << "  - Magma " << MAGMA_VERSION_MAJOR << '.' << MAGMA_VERSION_MINOR << '.' << MAGMA_VERSION_MICRO << '\n';
 #endif
 
   return oss.str();

aten/src/ATen/cuda/jiterator.cu

@@ -42,7 +42,7 @@ static inline void launch_jitted_vectorized_kernel_dynamic(
 
   // The cache key includes all the parameters to generate_code + vec_size + dev_idx
   std::stringstream ss;
-  ss << nInputs << "_" << nOutputs << f;
+  ss << nInputs << '_' << nOutputs << f;
   ss << f_inputs_type_str << compute_type_str << result_type_str;
   ss << static_cast<int>(at::cuda::jit::BinaryFuncVariant::NoScalar);
   ss << extra_args_types;
@@ -144,7 +144,7 @@ static inline void launch_jitted_unrolled_kernel_dynamic(
 
   // The cache key includes all the parameters to generate_code + dev_idx
   std::stringstream ss;
-  ss << nInputs << "_" << nOutputs << f;
+  ss << nInputs << '_' << nOutputs << f;
   ss << f_inputs_type_str << compute_type_str << result_type_str;
   ss << contiguous << dynamic_casting;
   ss << static_cast<int>(at::cuda::jit::BinaryFuncVariant::NoScalar);

aten/src/ATen/cuda/tunable/Tunable.cpp

@@ -52,10 +52,10 @@ TuningContext* getTuningContext() {
 std::ostream& operator<<(std::ostream& stream, const ResultEntry& entry) {
   static const bool blaslog = c10::utils::get_env("PYTORCH_TUNABLEOP_BLAS_LOG") == "1";
   if (!blaslog) {
-    return stream << entry.key_ << "," << entry.time_;
+    return stream << entry.key_ << ',' << entry.time_;
   }
   else {
-    return stream << entry.key_ << "," << entry.time_ << ",BLAS_PARAMS: " << entry.blas_sig_;
+    return stream << entry.key_ << ',' << entry.time_ << ",BLAS_PARAMS: " << entry.blas_sig_;
   }
 }
 
@@ -156,10 +156,10 @@ void TuningResultsManager::RecordUntuned( std::ofstream& untuned_file, const std
     if (isNew) {
       static const bool blaslog = c10::utils::get_env("PYTORCH_TUNABLEOP_BLAS_LOG") == "1";
       if (!blaslog) {
-        untuned_file << op_signature << "," << params_signature << std::endl;
+        untuned_file << op_signature << ',' << params_signature << std::endl;
       }
       else {
-        untuned_file << op_signature << "," << params_signature << ",BLAS_PARAMS: " << blas_signature << std::endl;
+        untuned_file << op_signature << ',' << params_signature << ",BLAS_PARAMS: " << blas_signature << std::endl;
       }
       TUNABLE_LOG3("Untuned,", op_signature, ",", params_signature);
     }
@@ -201,7 +201,7 @@ void TuningResultsManager::InitRealtimeAppend(const std::string& filename, const
 
   if(!file_exists || file_empty) {
     for(const auto& [key, val] : validators) {
-      (*realtime_out_) << "Validator," << key << "," << val << std::endl;
+      (*realtime_out_) << "Validator," << key << ',' << val << std::endl;
       realtime_out_->flush();
     }
     validators_written_ = true;
@@ -219,7 +219,7 @@ void TuningResultsManager::AppendResultLine(const std::string& op_sig, const std
     return;
   }
 
-  (*realtime_out_) << op_sig << "," << param_sig << "," << result << std::endl;
+  (*realtime_out_) << op_sig << ',' << param_sig << ',' << result << std::endl;
   realtime_out_->flush(); //ensure immediate write to disk
 
   TUNABLE_LOG3("Realtime append: ", op_sig, "(", param_sig, ") -> ", result);

aten/src/ATen/cudnn/Descriptors.cpp

@@ -93,31 +93,31 @@ std::string cudnnTypeToString(cudnnDataType_t dtype) {
       return "CUDNN_DATA_UINT8x4";
     default:
       std::ostringstream oss;
-      oss << "(unknown data-type " << static_cast<int>(dtype) << ")";
+      oss << "(unknown data-type " << static_cast<int>(dtype) << ')';
       return oss.str();
   }
 }
 
 std::ostream& operator<<(std::ostream & out, const TensorDescriptor& d) {
-  out << "TensorDescriptor " << static_cast<void*>(d.desc()) << "\n";
+  out << "TensorDescriptor " << static_cast<void*>(d.desc()) << '\n';
   int nbDims = 0;
   int dimA[CUDNN_DIM_MAX];
   int strideA[CUDNN_DIM_MAX];
   cudnnDataType_t dtype{};
   cudnnGetTensorNdDescriptor(d.desc(), CUDNN_DIM_MAX, &dtype, &nbDims, dimA, strideA);
-  out << "    type = " << cudnnTypeToString(dtype) << "\n";
-  out << "    nbDims = " << nbDims << "\n";
+  out << "    type = " << cudnnTypeToString(dtype) << '\n';
+  out << "    nbDims = " << nbDims << '\n';
   // Read out only nbDims of the arrays!
   out << "    dimA = ";
   for (auto i : ArrayRef<int>{dimA, static_cast<size_t>(nbDims)}) {
     out << i << ", ";
   }
-  out << "\n";
+  out << '\n';
   out << "    strideA = ";
   for (auto i : ArrayRef<int>{strideA, static_cast<size_t>(nbDims)}) {
     out << i << ", ";
   }
-  out << "\n";
+  out << '\n';
   return out;
 }
 
@@ -168,27 +168,27 @@ std::string cudnnMemoryFormatToString(cudnnTensorFormat_t tformat) {
       return "CUDNN_TENSOR_NHWC";
     default:
       std::ostringstream oss;
-      oss << "(unknown cudnn tensor format " << static_cast<int>(tformat) << ")";
+      oss << "(unknown cudnn tensor format " << static_cast<int>(tformat) << ')';
       return oss.str();
   }
 }
 
 std::ostream& operator<<(std::ostream & out, const FilterDescriptor& d) {
-  out << "FilterDescriptor " << static_cast<void*>(d.desc()) << "\n";
+  out << "FilterDescriptor " << static_cast<void*>(d.desc()) << '\n';
   int nbDims = 0;
   int dimA[CUDNN_DIM_MAX];
   cudnnDataType_t dtype{};
   cudnnTensorFormat_t tformat{};
   cudnnGetFilterNdDescriptor(d.desc(), CUDNN_DIM_MAX, &dtype, &tformat, &nbDims, dimA);
-  out << "    type = " << cudnnTypeToString(dtype) << "\n";
-  out << "    tensor_format = " << cudnnMemoryFormatToString(tformat) << "\n";
-  out << "    nbDims = " << nbDims << "\n";
+  out << "    type = " << cudnnTypeToString(dtype) << '\n';
+  out << "    tensor_format = " << cudnnMemoryFormatToString(tformat) << '\n';
+  out << "    nbDims = " << nbDims << '\n';
   // Read out only nbDims of the arrays!
   out << "    dimA = ";
   for (auto i : ArrayRef<int>{dimA, static_cast<size_t>(nbDims)}) {
     out << i << ", ";
   }
-  out << "\n";
+  out << '\n';
   return out;
 }
 

aten/src/ATen/functorch/DynamicLayer.cpp

@@ -346,15 +346,15 @@ void foreachTensorInplaceWithFlag(std::vector<IValue>& args, int64_t begin, int6
 }
 
 std::ostream& operator<< (std::ostream& os, const DynamicLayer& layer) {
-  os << layer.layerId() << ":" << layer.key();
+  os << layer.layerId() << ':' << layer.key();
   return os;
 }
 std::ostream& operator<< (std::ostream& os, const std::vector<DynamicLayer>& dls) {
   os << "DynamicLayerStack[ ";
   for (const auto& layer : dls) {
-    os << layer << " ";
+    os << layer << ' ';
   }
-  os << "]";
+  os << ']';
   return os;
 }
 

aten/src/ATen/functorch/TensorWrapper.cpp

@@ -22,7 +22,7 @@ void dumpTensor(std::ostream& ss, const Tensor& tensor) {
     if (batched) {
       ss << "Batched[lvl=" << batched->level() << " dim=" << batched->bdim() << ", ";
       dumpTensor(ss, batched->value());
-      ss << "]";
+      ss << ']';
       return;
     }
     ss << "Tensor" << tensor.sizes();
@@ -36,7 +36,7 @@ void dumpTensor(std::ostream& ss, const Tensor& tensor) {
     ss << "dead, ";
   }
   dumpTensor(ss, wrapped->value());
-  ss << "]";
+  ss << ']';
 }
 
 void TensorWrapper::refreshMetadata() {

aten/src/ATen/miopen/Descriptors.cpp

@@ -73,32 +73,32 @@ std::string miopenTypeToString(miopenDataType_t dtype) {
       return "miopenBFloat16";
     default:
       std::ostringstream oss;
-      oss << "(unknown data-type " << static_cast<int>(dtype) << ")";
+      oss << "(unknown data-type " << static_cast<int>(dtype) << ')';
       return oss.str();
   }
 }
 
 std::ostream& operator<<(std::ostream & out, const TensorDescriptor& d) {
-  out << "TensorDescriptor " << static_cast<void*>(d.desc()) << "\n";
+  out << "TensorDescriptor " << static_cast<void*>(d.desc()) << '\n';
   int nbDims = 0;
   int dimA[MIOPEN_DIM_MAX];
   int strideA[MIOPEN_DIM_MAX];
   miopenDataType_t dtype;
   miopenGetTensorDescriptorSize(d.desc(), &nbDims);
   miopenGetTensorDescriptor(d.desc(), &dtype, dimA, strideA);
-  out << "    type = " << miopenTypeToString(dtype) << "\n";
-  out << "    nbDims = " << nbDims << "\n";
+  out << "    type = " << miopenTypeToString(dtype) << '\n';
+  out << "    nbDims = " << nbDims << '\n';
   // Read out only nbDims of the arrays!
   out << "    dimA = ";
   for (auto i : ArrayRef<int>{dimA, static_cast<size_t>(nbDims)}) {
     out << i << ", ";
   }
-  out << "\n";
+  out << '\n';
   out << "    strideA = ";
   for (auto i : ArrayRef<int>{strideA, static_cast<size_t>(nbDims)}) {
     out << i << ", ";
   }
-  out << "\n";
+  out << '\n';
   return out;
 }
 

aten/src/ATen/mps/MPSDevice.h

@@ -22,7 +22,6 @@ enum class MacOSVersion : uint32_t {
   MACOS_VER_15_0_PLUS,
   MACOS_VER_15_1_PLUS,
   MACOS_VER_15_2_PLUS,
-  MACOS_VER_26_0_PLUS,
 };
 
 //-----------------------------------------------------------------

aten/src/ATen/mps/MPSDevice.mm

@@ -65,7 +65,6 @@ bool MPSDevice::isMacOS13Plus(MacOSVersion version) const {
   static bool _macos_15_0_plus = is_os_version_at_least(15, 0);
   static bool _macos_15_1_plus = is_os_version_at_least(15, 1);
   static bool _macos_15_2_plus = is_os_version_at_least(15, 2);
-  static bool _macos_26_0_plus = is_os_version_at_least(26, 0);
 
   switch (version) {
     case MacOSVersion::MACOS_VER_14_4_PLUS:
@@ -76,8 +75,6 @@ bool MPSDevice::isMacOS13Plus(MacOSVersion version) const {
       return _macos_15_1_plus;
     case MacOSVersion::MACOS_VER_15_2_PLUS:
       return _macos_15_2_plus;
-    case MacOSVersion::MACOS_VER_26_0_PLUS:
-      return _macos_26_0_plus;
     default:
       return false;
   }

aten/src/ATen/mps/MPSProfiler.h

@@ -91,7 +91,7 @@ struct OperationInfo : BaseInfo {
     std::stringstream kernelStr;
     kernelStr << kernelName;
     for (const Tensor& tensor : tensors) {
-      kernelStr << ":" << BaseInfo::buildTensorString(tensor, includeBufferId);
+      kernelStr << ':' << BaseInfo::buildTensorString(tensor, includeBufferId);
     }
     return kernelStr.str();
   }

aten/src/ATen/mps/MPSProfiler.mm

@@ -39,9 +39,9 @@ std::string BaseInfo::buildTensorString(const Tensor& tensor, bool includeBuffer
     // see comments for INCLUDE_BUFFER_ID
     if (includeBufferId && deviceType == at::kMPS) {
       id<MTLBuffer> buffer = __builtin_bit_cast(id<MTLBuffer>, tensor.storage().data());
-      tensorStr << "(buf#" << (getIMPSAllocator()->getBufferId(buffer)) << ":" << buffer.retainCount << ")";
+      tensorStr << "(buf#" << (getIMPSAllocator()->getBufferId(buffer)) << ':' << buffer.retainCount << ')';
     }
-    tensorStr << ":" << tensor.scalar_type() << tensor.sizes();
+    tensorStr << ':' << tensor.scalar_type() << tensor.sizes();
     return tensorStr.str();
   } else {
     return "undefined";

aten/src/ATen/native/ConvUtils.h

@@ -167,7 +167,7 @@ static void check_args(CheckedFrom c, IntArrayRef args, size_t expected_size, co
     std::stringstream ss;
     ss << arg_name << " should be greater than zero but got (";
     std::copy(args.begin(), args.end() - 1, std::ostream_iterator<int>(ss,", "));
-    ss << args.back() <<  ")" << " (while checking arguments for " << c << ")";
+    ss << args.back() <<  ")" << " (while checking arguments for " << c << ')';
     TORCH_CHECK(false, ss.str());
   }
 }

aten/src/ATen/native/Convolution.cpp

@@ -639,7 +639,7 @@ static std::ostream& operator<<(std::ostream & out, const ConvParams<T>& params)
       << "  deterministic = " << params.deterministic
       << "  cudnn_enabled = " << params.cudnn_enabled
       << "  allow_tf32 = " << params.allow_tf32
-      << "}";
+      << '}';
   return out;
 }
 

aten/src/ATen/native/SpectralOps.cpp

@@ -847,7 +847,7 @@ Tensor stft(const Tensor& self, const int64_t n_fft, const std::optional<int64_t
        << ", hop_length=" << hop_length << ", win_length=" << win_length \
        << ", window="; \
     if (window.defined()) { \
-      SS << window.toString() << "{" << window.sizes() << "}"; \
+      SS << window.toString() << '{' << window.sizes() << '}'; \
     } else { \
       SS << "None"; \
     } \
@@ -1046,7 +1046,7 @@ Tensor istft(const Tensor& self, const int64_t n_fft, const std::optional<int64_
        << ", hop_length=" << hop_length << ", win_length=" << win_length \
        << ", window="; \
     if (window.defined()) { \
-      SS << window.toString() << "{" << window.sizes() << "}"; \
+      SS << window.toString() << '{' << window.sizes() << '}'; \
     } else { \
       SS << "None"; \
     } \

aten/src/ATen/native/TensorCompare.cpp

@@ -523,7 +523,7 @@ Tensor _functional_assert_async_msg_cpu(
 }
 
 void _print(std::string_view s) {
-  std::cout << s << "\n";
+  std::cout << s << '\n';
 }
 
 // Sorting-based algorithm for isin(); used when the number of test elements is

aten/src/ATen/native/cuda/GroupedBlas.cpp

@@ -607,6 +607,8 @@ _scaled_grouped_mm_cuda_v2(
       // scale shape checks
       _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
       _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      // swizze checks
+      TORCH_CHECK_VALUE(swizzle_a_enum.size() == 1 && swizzle_b_enum.size() == 1, "Expected single swizzle argument");
       return _mx8_mx8_bf16_grouped_mm_fbgemm(
           mat_a,
           mat_b,

aten/src/ATen/native/cuda/KernelUtils.cuh

@@ -5,11 +5,69 @@
 #include <cuda_bf16.h>
 #endif
 
+// ROCm 6.3 is planned to have these functions, but until then here they are.
 #if defined(USE_ROCM)
 #include <device_functions.h>
 #include <hip/hip_fp16.h>
 #include <hip/hip_bf16.h>
-#define ATOMICADD unsafeAtomicAdd
+
+__device__ inline __hip_bfloat162 preview_unsafeAtomicAdd(__hip_bfloat162* address, __hip_bfloat162 value) {
+#if (defined(__gfx942__)) && \
+  __has_builtin(__builtin_amdgcn_flat_atomic_fadd_v2bf16)
+  typedef unsigned short __attribute__((ext_vector_type(2))) vec_short2;
+  static_assert(sizeof(vec_short2) == sizeof(__hip_bfloat162_raw));
+  union {
+    __hip_bfloat162_raw bf162_raw;
+    vec_short2 vs2;
+  } u{static_cast<__hip_bfloat162_raw>(value)};
+  u.vs2 = __builtin_amdgcn_flat_atomic_fadd_v2bf16((vec_short2*)address, u.vs2);
+  return static_cast<__hip_bfloat162>(u.bf162_raw);
+#else
+  static_assert(sizeof(unsigned int) == sizeof(__hip_bfloat162_raw));
+  union u_hold {
+    __hip_bfloat162_raw h2r;
+    unsigned int u32;
+  };
+  u_hold old_val, new_val;
+  old_val.u32 = __hip_atomic_load((unsigned int*)address, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);
+  do {
+    new_val.h2r = __hadd2(old_val.h2r, value);
+  } while (!__hip_atomic_compare_exchange_strong(
+        (unsigned int*)address, &old_val.u32, new_val.u32,
+        __ATOMIC_RELAXED, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT));
+  return old_val.h2r;
+#endif
+}
+
+__device__ inline __half2 preview_unsafeAtomicAdd(__half2* address, __half2 value) {
+#if (defined(__gfx942__)) && \
+  __has_builtin(__builtin_amdgcn_flat_atomic_fadd_v2f16)
+  // The api expects an ext_vector_type of half
+  typedef _Float16 __attribute__((ext_vector_type(2))) vec_fp162;
+  static_assert(sizeof(vec_fp162) == sizeof(__half2_raw));
+  union {
+    __half2_raw h2r;
+    vec_fp162 fp16;
+  } u {static_cast<__half2_raw>(value)};
+  u.fp16 = __builtin_amdgcn_flat_atomic_fadd_v2f16((vec_fp162*)address, u.fp16);
+  return static_cast<__half2>(u.h2r);
+#else
+  static_assert(sizeof(__half2_raw) == sizeof(unsigned int));
+  union u_hold {
+    __half2_raw h2r;
+    unsigned int u32;
+  };
+  u_hold old_val, new_val;
+  old_val.u32 = __hip_atomic_load((unsigned int*)address, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);
+  do {
+    new_val.h2r = __hadd2(old_val.h2r, value);
+  } while (!__hip_atomic_compare_exchange_strong(
+        (unsigned int*)address, &old_val.u32, new_val.u32,
+        __ATOMIC_RELAXED, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT));
+  return old_val.h2r;
+#endif
+}
+#define ATOMICADD preview_unsafeAtomicAdd
 #define NATIVE_ZERO_BF16 __float2bfloat16(0.0f)
 #else
 #define ATOMICADD atomicAdd

aten/src/ATen/native/cuda/Reduce.cu

@@ -11,7 +11,7 @@ static inline std::ostream& operator<<(std::ostream& out, dim3 dim) {
   if (dim.y == 1 && dim.z == 1) {
     out << dim.x;
   } else {
-    out << "[" << dim.x << "," << dim.y << "," << dim.z << "]";
+    out << '[' << dim.x << ',' << dim.y << ',' << dim.z << ']';
   }
   return out;
 }
@@ -27,7 +27,7 @@ std::ostream& operator<<(std::ostream& out, const ReduceConfig& config) {
   out << "input_mult=[";
   for (int i = 0; i < 3; i++) {
     if (i != 0) {
-      out << ",";
+      out << ',';
     }
     out << config.input_mult[i];
   }
@@ -35,7 +35,7 @@ std::ostream& operator<<(std::ostream& out, const ReduceConfig& config) {
   out << "output_mult=[";
   for (int i = 0; i < 2; i++) {
     if (i != 0) {
-      out << ",";
+      out << ',';
     }
     out << config.output_mult[i];
   }
@@ -49,7 +49,7 @@ std::ostream& operator<<(std::ostream& out, const ReduceConfig& config) {
   out << "block=" << config.block() << ", ";
   out << "grid=" << config.grid() << ", ";
   out << "global_memory_size=" << config.global_memory_size();
-  out << ")";
+  out << ')';
   return out;
 }
 

aten/src/ATen/native/cuda/ScaledGroupMM.cu

@@ -364,9 +364,9 @@ void f8f8bf16_grouped_gemm_impl_sm90(
   //       reinterpret_cast<ProblemShape::UnderlyingProblemShape*>(
   //           stride_output_h + group_count);
 
-  //   std::cout << "PTRS " << mat_a.data_ptr() << " " << mat_b.data_ptr() << "
+  //   std::cout << "PTRS " << mat_a.data_ptr() << ' ' << mat_b.data_ptr() << "
   //   "
-  //             << out.data_ptr() << " " << scale_a.data_ptr() << " "
+  //             << out.data_ptr() << ' ' << scale_a.data_ptr() << ' '
   //             << scale_b.data_ptr() << "\n";
   //   for (int i = 0; i < group_count; i++) {
   //     std::cout << "A " << (void*)inputA_ptrs_h[i] << "\n";

aten/src/ATen/native/cuda/jit_utils.cpp

@@ -1057,14 +1057,14 @@ std::string generate_code(
     // TODO these arrays are potentially of the different types, use function
     // traits to determine the types
     declare_load_arrays << f_inputs_type << " arg" << std::to_string(i)
-                        << "[" << std::to_string(thread_work_size) << "];\n";
+                        << '[' << std::to_string(thread_work_size) << "];\n";
   }
   env.s("declare_load_arrays", declare_load_arrays.str());
 
   std::stringstream declare_store_arrays;
   for (int i = 0; i < nOutputs; i++) {
     declare_store_arrays << result_type << " out" << std::to_string(i)
-                        << "[" << std::to_string(thread_work_size) << "];\n";
+                        << '[' << std::to_string(thread_work_size) << "];\n";
   }
   env.s("declare_store_arrays", declare_store_arrays.str());
 
@@ -1217,7 +1217,7 @@ std::string generate_code(
   for (const auto i : c10::irange(nInputs)){
     auto i_string = std::to_string(i);
     vector_inputs << "auto * input" << i_string <<
-        " = reinterpret_cast<const scalar_t*>(data[" << i_string << "+" << nOutputs << "])" <<
+        " = reinterpret_cast<const scalar_t*>(data[" << i_string << '+' << nOutputs << "])" <<
         " + block_work_size * idx;\n";
   }
   env.s("vector_inputs", vector_inputs.str());
@@ -1543,17 +1543,17 @@ NvrtcFunction jit_pwise_function(
 
     // Constructs file path by appending constructed cubin name to cache path
     std::stringstream ss;
-    ss << *cache_dir << "/";
+    ss << *cache_dir << '/';
     ss << kernel_name;
 #ifdef USE_ROCM
     ss << "_arch" << prop->gcnArchName;
 #else
-    ss << "_arch" << cuda_major << "." << cuda_minor;
+    ss << "_arch" << cuda_major << '.' << cuda_minor;
 #endif
-    ss << "_nvrtc" << nvrtc_major << "." << nvrtc_minor;
+    ss << "_nvrtc" << nvrtc_major << '.' << nvrtc_minor;
     ss << (compile_to_sass ? "_sass" : "_ptx");
-    ss << "_" << code.length();
-    ss << "_" << hash_code;
+    ss << '_' << code.length();
+    ss << '_' << hash_code;
     file_path = ss.str();
 
     std::ifstream readin{file_path, std::ios::in | std::ifstream::binary};

aten/src/ATen/native/cudnn/ConvShared.cpp

@@ -82,15 +82,15 @@ namespace native {
 
 std::ostream& operator<<(std::ostream& out, const ConvolutionParams& params) {
   out << "ConvolutionParams \n"
-      << "    memory_format = " << params.memory_format << "\n"
-      << "    data_type = " << cudnnTypeToString(params.dataType) << "\n"
-      << "    padding = " << ArrayRef<int>{params.padding} << "\n"
-      << "    stride = " << ArrayRef<int>{params.stride} << "\n"
-      << "    dilation = " << ArrayRef<int>{params.dilation} << "\n"
-      << "    groups = " << params.groups << "\n"
+      << "    memory_format = " << params.memory_format << '\n'
+      << "    data_type = " << cudnnTypeToString(params.dataType) << '\n'
+      << "    padding = " << ArrayRef<int>{params.padding} << '\n'
+      << "    stride = " << ArrayRef<int>{params.stride} << '\n'
+      << "    dilation = " << ArrayRef<int>{params.dilation} << '\n'
+      << "    groups = " << params.groups << '\n'
       << "    deterministic = " << (params.deterministic ? "true" : "false")
-      << "\n"
-      << "    allow_tf32 = " << (params.allow_tf32 ? "true" : "false") << "\n";
+      << '\n'
+      << "    allow_tf32 = " << (params.allow_tf32 ? "true" : "false") << '\n';
 
   return out;
 }
@@ -173,16 +173,16 @@ std::string repro_from_args(const ConvolutionParams& params) {
             at::globalContext().float32Precision(
                 at::Float32Backend::CUDA, at::Float32Op::MATMUL) ==
             at::Float32Precision::TF32)
-     << "\n";
+     << '\n';
   ss << "torch.backends.cudnn.benchmark = "
-     << pybool(at::globalContext().benchmarkCuDNN()) << "\n";
+     << pybool(at::globalContext().benchmarkCuDNN()) << '\n';
   ss << "torch.backends.cudnn.deterministic = " << pybool(params.deterministic)
-     << "\n";
+     << '\n';
   ss << "torch.backends.cudnn.allow_tf32 = " << pybool(params.allow_tf32)
-     << "\n";
+     << '\n';
   ss << "data = torch.randn(" << ArrayRef<int>(params.input_size, dim)
      << ", dtype=" << full_dtype << ", ";
-  ss << "device='cuda', requires_grad=True)" << to_channels_last << "\n";
+  ss << "device='cuda', requires_grad=True)" << to_channels_last << '\n';
   ss << "net = torch.nn.Conv" << dim - 2 << "d(" << in_channels << ", "
      << out_channels << ", ";
   ss << "kernel_size=" << ArrayRef<int>(&params.weight_size[2], dim - 2)
@@ -192,7 +192,7 @@ std::string repro_from_args(const ConvolutionParams& params) {
   ss << "dilation=" << ArrayRef<int>(params.dilation, dim - 2) << ", ";
   ss << "groups=" << params.groups << ")\n";
   ss << "net = net.cuda()." << partial_dtype << "()" << to_channels_last
-     << "\n";
+     << '\n';
   ss << "out = net(data)\n";
   ss << "out.backward(torch.randn_like(out))\n";
   ss << "torch.cuda.synchronize()\n\n";

aten/src/ATen/native/cudnn/Conv_v7.cpp

@@ -93,11 +93,10 @@ std::ostream& operator<<(std::ostream& out, const ConvolutionArgs& args) {
       << "input: " << args.idesc // already has a trailing newline
       << "output: " << args.odesc // already has a trailing newline
       << "weight: " << args.wdesc // already has a trailing newline
-      << "Pointer addresses: "
-      << "\n"
-      << "    input: " << args.input.const_data_ptr() << "\n"
-      << "    output: " << args.output.const_data_ptr() << "\n"
-      << "    weight: " << args.weight.const_data_ptr() << "\n";
+      << "Pointer addresses: " << '\n'
+      << "    input: " << args.input.const_data_ptr() << '\n'
+      << "    output: " << args.output.const_data_ptr() << '\n'
+      << "    weight: " << args.weight.const_data_ptr() << '\n';
 
   return out;
 }

aten/src/ATen/native/metal/MetalTensorImplStorage.mm

@@ -115,7 +115,7 @@ std::ostream& operator<<(
   std::copy(
       strides.begin(), strides.end() - 1, std::ostream_iterator<int>(oss, ","));
   oss << sizes.back();
-  output << oss.str() << "}";
+  output << oss.str() << '}';
   return output;
 }
 

aten/src/ATen/native/mkldnn/xpu/Conv.cpp

@@ -53,7 +53,7 @@ std::ostream& operator<<(std::ostream& out, const ConvParams& params) {
       << "  transposed = " << params.transposed
       << "  output_padding = " << IntArrayRef{params.output_padding}
       << "  groups = " << params.groups << "  benchmark = " << params.benchmark
-      << "  deterministic = " << params.deterministic << "}";
+      << "  deterministic = " << params.deterministic << '}';
   return out;
 }
 

aten/src/ATen/native/mkldnn/xpu/ScaledBlas.cpp

@@ -5,6 +5,7 @@
 #include <ATen/native/Resize.h>
 #include <ATen/native/mkldnn/xpu/detail/oneDNN.h>
 #include <ATen/native/xpu/Blas.h>
+#include <ATen/xpu/XPUScaledBlas.h>
 #include <torch/library.h>
 
 #ifndef AT_PER_OPERATOR_HEADERS
@@ -339,4 +340,399 @@ Tensor _scaled_mm_xpu(
       out);
 }
 
+using acceptance_fn = std::function<bool(
+    c10::ScalarType,
+    std::vector<ScalingType>&,
+    ArrayRef<Tensor>&,
+    c10::ScalarType,
+    std::vector<ScalingType>&,
+    ArrayRef<Tensor>&)>;
+using namespace std::placeholders;
+
+namespace scaled_blas = at::native::onednn::scaled;
+using scaled_blas::convert_int_to_enum;
+using scaled_blas::ScaledGemmImplementation;
+
+std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 2>
+    scale_kernel_dispatch = {{
+        {"tensorwise_tensorwise",
+         scaled_blas::check_tensorwise_recipe,
+         ScaledGemmImplementation::TENSORWISE_TENSORWISE},
+        {"rowwise_rowwise",
+         scaled_blas::check_rowwise_recipe,
+         ScaledGemmImplementation::ROWWISE_ROWWISE},
+
+    }};
+
+Tensor& _scaled_tensorwise_tensorwise(
+    const Tensor& mat_a,
+    const Tensor& mat_b,
+    const Tensor& scale_a,
+    const Tensor& scale_b,
+    const std::optional<Tensor>& bias,
+    const c10::ScalarType out_dtype,
+    bool use_fast_accum,
+    Tensor& out) {
+  // Restrictions:
+  // A, B are FP8, scales are fp32
+
+  TORCH_CHECK_VALUE(
+      isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()),
+      "mat_a and mat_b must be fp8 types, got: ",
+      mat_a.scalar_type(),
+      mat_b.scalar_type());
+  TORCH_CHECK_VALUE(
+      scale_a.numel() == 1 && scale_a.scalar_type() == kFloat,
+      "scale_a must have 1 Float element")
+  TORCH_CHECK_VALUE(
+      scale_b.numel() == 1 && scale_b.scalar_type() == kFloat,
+      "scale_b must have 1 Float element")
+
+  auto scaling_choice_a = ScalingType::TensorWise;
+  auto scaling_choice_b = ScalingType::TensorWise;
+
+  _scaled_gemm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      scaling_choice_a,
+      scaling_choice_b,
+      bias,
+      use_fast_accum,
+      out);
+
+  return out;
+}
+
+Tensor& _scaled_rowwise_rowwise(
+    const Tensor& mat_a,
+    const Tensor& mat_b,
+    const Tensor& scale_a,
+    const Tensor& scale_b,
+    const std::optional<Tensor>& bias,
+    const c10::ScalarType out_dtype,
+    bool use_fast_accum,
+    Tensor& out) {
+  // Restrictions:
+  // A, B are FP8, scales are fp32, shape M/N for A/B
+  TORCH_CHECK_VALUE(
+      isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()),
+      "mat_a and mat_b must be fp8 types, got: ",
+      mat_a.scalar_type(),
+      mat_b.scalar_type());
+  TORCH_CHECK_VALUE(
+      scale_a.size(0) == mat_a.size(0) && scale_a.size(1) == 1,
+      "scale_a must have shape [",
+      mat_a.size(0),
+      ", 1], got [",
+      scale_a.sizes(),
+      "]");
+  TORCH_CHECK_VALUE(
+      scale_a.numel() == mat_a.size(0) && scale_a.scalar_type() == kFloat,
+      "scale_a must have ",
+      mat_a.size(0),
+      " Float elements, got ",
+      scale_a.numel())
+  TORCH_CHECK_VALUE(
+      scale_b.numel() == mat_b.size(1) && scale_b.scalar_type() == kFloat,
+      "scale_b must have ",
+      mat_b.size(1),
+      " Float elements, got ",
+      scale_b.numel())
+
+  TORCH_CHECK_VALUE(
+      scale_a.stride(1) == 1,
+      "expected scale_a.stride(1) to be 1, but got ",
+      scale_a.stride(1));
+  TORCH_CHECK_VALUE(
+      scale_b.stride(1) == 1,
+      "expected scale_b.stride(1) to be 1, but got ",
+      scale_b.stride(1));
+
+  auto scaling_choice_a = ScalingType::RowWise;
+  auto scaling_choice_b = ScalingType::RowWise;
+
+  _scaled_gemm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      scaling_choice_a,
+      scaling_choice_b,
+      bias,
+      use_fast_accum,
+      out);
+
+  return out;
+}
+
+// V2: Computes matrix multiply + bias while applying scaling to input and
+// output matrices Scales are only applicable when matrices are of Float8 type
+// and assumed to be equal to 1.0 by default. If output matrix type is 16 or
+// 32-bit type, scale_result is not applied. Known limitations:
+//  - Only works if mat1 is row-major and mat2 is column-major
+//  - Only works if matrices sizes are divisible by 32
+//  - If 1-dimensional tensors are used then scale_a should be size =
+//  mat1.size(0)
+//    and scale_b should have size = to mat2.size(1)
+//  Arguments:
+//    - `mat_a`: the first operand of the matrix multiply, can be type
+//    `torch.float8_e4m3fn` or `torch.float8_e5m2`
+//    - `mat_b`: the second operand of the matrix multiply, can be type
+//    `torch.float8_e4m3fn` or `torch.float8_e5m2`
+//    - `scale_a`: a tensor with the inverse scale of `mat1`, whose
+//    shape/strides/dtype depend on the scaling scheme
+//    - `scale_recipe_a`: An integer corresponding to an enum describing the
+//    scaling scheme used for `scale_a`
+//    - `swizzle_a`: An integer corresponding to a `SwizzleType` enum describing
+//    the swizzling scheme for `scale_a`.
+//        Not supported for XPU for now.
+//    - `scale_b`: a tensor with the inverse scale of `mat2`, whose
+//    shape/strides/dtype depend on the scaling scheme
+//    - `scale_recipe_b`: An integer corresponding to an enum describing the
+//    scaling scheme used for `scale_b`
+//    - `swizzle_b`: An integer corresponding to a `SwizzleType` enum describing
+//    the swizzling scheme for `scale_b`.
+//        Not supported for XPU for now.
+//    - `bias`: the bias, can be type `torch.float16` or `torch.bfloat16`
+//    - `out_dtype`: the output dtype, can either be a float8 or a higher
+//    precision floating point type
+//    - `contraction_dim`: describe which dimensions are `K` in the matmul.
+//       Not supported for XPU. Should always be empty.
+//    - `use_fast_accum`: Not supported for XPU, should always be false.
+//    - `out`: a reference to the output tensor
+Tensor& _scaled_mm_xpu_v2_out(
+    const Tensor& mat_a,
+    const Tensor& mat_b,
+    ArrayRef<Tensor> scale_a,
+    IntArrayRef scale_recipe_a,
+    IntArrayRef swizzle_a,
+    ArrayRef<Tensor> scale_b,
+    IntArrayRef scale_recipe_b,
+    IntArrayRef swizzle_b,
+    const std::optional<Tensor>& bias,
+    const std::optional<c10::ScalarType> out_dtype,
+    IntArrayRef contraction_dim,
+    bool use_fast_accum,
+    Tensor& out) {
+  TORCH_CHECK_VALUE(mat_a.dim() == 2, "mat_a must be a matrix");
+  TORCH_CHECK_VALUE(mat_b.dim() == 2, "mat_b must be a matrix");
+
+  // If any of M, K, N is 0 - return early (the tensorwise/rowwise float8 gemm
+  // kernels do not support this case).
+  if (mat_a.size(0) == 0 || mat_a.size(1) == 0 || mat_b.size(1) == 0) {
+    // `out` was created with `at::empty`. In the case where we are multiplying
+    // MxK by KxN and K is the zero dim, we need to initialize here to properly
+    // return a tensor of zeros.
+    at::native::resize_output(out, {mat_a.size(0), mat_b.size(1)});
+    if (mat_a.size(1) == 0) {
+      out.zero_();
+    }
+
+    return out;
+  }
+
+  // Note: The `contraction_dim` is not actually used for now. We will need to
+  // align this code when upstreamed CUDA code is done. Currently, only keeps
+  // the code here for check.
+
+  // Check if the input matrix sizes can be multiplied
+  // - if optional contraction dims are provided, use those
+  //   -- mostly for < 1B formats (i.e. nvfp4x2) where cheap .t() is not
+  //   available.
+  if (contraction_dim.size() > 0) {
+    TORCH_CHECK_VALUE(
+        contraction_dim.size() == 2,
+        "contraction_dim must have exactly 2 elements");
+    auto mat_a_dim = contraction_dim[0];
+    auto mat_b_dim = contraction_dim[1];
+    TORCH_CHECK_VALUE(
+        mat_a.size(mat_a_dim) == mat_b.size(mat_b_dim),
+        "mat_a and mat_b shapes cannot be multiplied (",
+        mat_a.size(0),
+        "x",
+        mat_a.size(1),
+        " and ",
+        mat_b.size(0),
+        "x",
+        mat_b.size(1),
+        ") ",
+        "with contraction dims mat_a: ",
+        mat_a_dim,
+        ", mat_b: ",
+        mat_b_dim);
+  } else {
+    TORCH_CHECK_VALUE(
+        mat_a.size(1) == mat_b.size(0),
+        "mat_a and mat_b shapes cannot be multiplied (",
+        mat_a.size(0),
+        "x",
+        mat_a.size(1),
+        " and ",
+        mat_b.size(0),
+        "x",
+        mat_b.size(1),
+        ")");
+  }
+
+  TORCH_CHECK_VALUE(
+      !bias || bias->numel() == mat_b.sizes()[1],
+      "Bias must be size ",
+      mat_b.sizes()[1],
+      " but got ",
+      bias->numel());
+
+  TORCH_CHECK_VALUE(
+      !out_dtype || *out_dtype == out.scalar_type(),
+      "out_dtype must match output matrix type");
+
+  if (bias) {
+    TORCH_CHECK_VALUE(
+        bias->scalar_type() == kFloat ||
+            bias->scalar_type() == c10::ScalarType::BFloat16 ||
+            bias->scalar_type() == c10::ScalarType::Half,
+        "Bias must be Float32 or BFloat16 or Half, but got ",
+        bias->scalar_type());
+  }
+  {
+    auto bias_ = bias.value_or(Tensor());
+    // NOLINTNEXTLINE(*c-array*)
+    TensorArg targs[]{
+        {out, "out", 0},
+        {mat_a, "mat_a", 1},
+        {mat_b, "mat_b", 2},
+        {bias_, "bias", 3},
+        {scale_a[0], "scale_a", 4},
+        {scale_b[0], "scale_b", 5}};
+    checkAllSameGPU(__func__, targs);
+  }
+  // Align with CUDA's default out to be bf16
+  auto out_dtype_ = out_dtype.value_or(c10::ScalarType::BFloat16);
+
+  // Conversion of implicitly-defined enums to explicit
+  auto scale_recipe_a_enum = convert_int_to_enum<ScalingType>(scale_recipe_a);
+  auto swizzle_a_enum = convert_int_to_enum<SwizzleType>(swizzle_a);
+  auto scale_recipe_b_enum = convert_int_to_enum<ScalingType>(scale_recipe_b);
+  auto swizzle_b_enum = convert_int_to_enum<SwizzleType>(swizzle_b);
+
+  // XPU does not support swizzle for now. So directly return false.
+  TORCH_CHECK_VALUE(
+      swizzle_a_enum[0] == at::blas::SwizzleType::NO_SWIZZLE &&
+          swizzle_b_enum[0] == at::blas::SwizzleType::NO_SWIZZLE,
+      "XPU does not support swizzle yet.");
+
+  // at this point we can start working out what we want to be doing
+  // Try to do as few steps as possible.
+  // NOTE: support is deliberately sparse, can explicitly enumerate all
+  // combinations allowed. Do this via a list of defined (name, acceptance,
+  // concrete_impl) tuples.
+  bool found_impl = false;
+  ScaledGemmImplementation gemm_impl = ScaledGemmImplementation::NONE;
+
+  for (const auto& fn_entry : scale_kernel_dispatch) {
+    const auto [name, accept_fn, scaled_gemm_impl] = fn_entry;
+    bool ok = accept_fn(
+        mat_a.scalar_type(),
+        scale_recipe_a_enum,
+        scale_a,
+        mat_b.scalar_type(),
+        scale_recipe_b_enum,
+        scale_b);
+    if (ok) {
+      gemm_impl = scaled_gemm_impl;
+      found_impl = true;
+      break;
+    }
+  }
+  TORCH_CHECK_VALUE(
+      found_impl,
+      "Invalid scaling configuration.\n"
+      "- For TensorWise scaling, a and b should be float8, scales should be float and singletons.\n"
+      "- For RowWise scaling, a and b should be float8, scales should be float, scale_a should be (",
+      mat_a.size(0),
+      ", 1) and scale_b should be (1, ",
+      mat_b.size(1),
+      "), and both should be contiguous.\n"
+      "Got mat_a.dtype()=",
+      mat_a.scalar_type(),
+      ", scale_a[0].dtype()=",
+      scale_a[0].scalar_type(),
+      ", scale_a[0].size()=",
+      scale_a[0].sizes(),
+      ", scale_a[0].stride()=",
+      scale_a[0].strides(),
+      ", ",
+      "mat_b.dtype()=",
+      mat_b.scalar_type(),
+      ", scale_b[0].dtype()=",
+      scale_b[0].scalar_type(),
+      ", scale_b[0].size()=",
+      scale_b[0].sizes(),
+      " and scale_b[0].stride()=",
+      scale_b[0].strides());
+
+  at::native::resize_output(out, {mat_a.size(0), mat_b.size(1)});
+
+  auto bias_ = bias.value_or(Tensor());
+
+  // dispatch to appropriate lower-level calls for error checking & execution
+  if (gemm_impl == ScaledGemmImplementation::TENSORWISE_TENSORWISE) {
+    return _scaled_tensorwise_tensorwise(
+        mat_a,
+        mat_b,
+        scale_a[0],
+        scale_b[0],
+        bias,
+        out_dtype_,
+        use_fast_accum,
+        out);
+  } else if (gemm_impl == ScaledGemmImplementation::ROWWISE_ROWWISE) {
+    return _scaled_rowwise_rowwise(
+        mat_a,
+        mat_b,
+        scale_a[0],
+        scale_b[0],
+        bias,
+        out_dtype_,
+        use_fast_accum,
+        out);
+  } else {
+    TORCH_CHECK_VALUE(
+        false, "Invalid state - found an implementation, but not really");
+  }
+}
+
+Tensor _scaled_mm_xpu_v2(
+    const Tensor& mat_a,
+    const Tensor& mat_b,
+    ArrayRef<Tensor> scale_a,
+    IntArrayRef scale_recipe_a,
+    IntArrayRef swizzle_a,
+    ArrayRef<Tensor> scale_b,
+    IntArrayRef scale_recipe_b,
+    IntArrayRef swizzle_b,
+    const std::optional<Tensor>& bias,
+    const std::optional<c10::ScalarType> out_dtype,
+    IntArrayRef contraction_dim,
+    bool use_fast_accum) {
+  const auto out_dtype_ = out_dtype.value_or(mat_a.scalar_type());
+  Tensor out = at::empty({0}, mat_a.options().dtype(out_dtype_));
+
+  return _scaled_mm_xpu_v2_out(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_recipe_a,
+      swizzle_a,
+      scale_b,
+      scale_recipe_b,
+      swizzle_b,
+      bias,
+      out_dtype,
+      contraction_dim,
+      use_fast_accum,
+      out);
+}
+
 } // namespace at::native

aten/src/ATen/native/mps/operations/Attention.mm

@@ -69,139 +69,75 @@ static std::tuple<Tensor, Tensor> sdpa_general_mps(const Tensor& query,
   auto out = at::empty({batchSize, num_head, qSize, headSize}, query.options());
   auto attn = at::empty({batchSize, num_head, qSize, maxSeqLength}, query.options());
   auto scale_factor = sdp::calculate_scale(query, scale).expect_float();
-  static const bool is_macOS_26_0_or_newer = is_macos_13_or_newer(MacOSVersion::MACOS_VER_26_0_PLUS);
   @autoreleasepool {
     auto mkey = __func__ + getTensorsStringKey({query, key, value}) + ":" + std::to_string(is_causal) + ":" +
         std::to_string(attn_mask.has_value());
+    auto cachedGraph =
+        LookUpOrCreateCachedGraph<CachedGraph>(mkey, [&, q_ = query, k_ = key, v_ = value](auto mpsGraph, auto graph) {
+          auto qTensor = mpsGraphRankedPlaceHolder(mpsGraph, q_);
+          auto kTensor = mpsGraphRankedPlaceHolder(mpsGraph, k_);
+          auto vTensor = mpsGraphRankedPlaceHolder(mpsGraph, v_);
+          auto kT = [mpsGraph transposeTensor:kTensor dimension:2 withDimension:3 name:nil];
+          auto scaleTensor = [mpsGraph constantWithScalar:scale_factor
+                                                    shape:getMPSShape({1})
+                                                 dataType:MPSDataTypeFloat32];
 
-    CachedGraph* cachedGraph;
-    //if(is_macOS_26_0_or_newer) {
-    if(true) {
-        cachedGraph =
-            LookUpOrCreateCachedGraph<CachedGraph>(mkey, [&, q_ = query, k_ = key, v_ = value](auto mpsGraph, auto graph) {
-              auto qTensor = mpsGraphRankedPlaceHolder(mpsGraph, q_);
-              auto kTensor = mpsGraphRankedPlaceHolder(mpsGraph, k_);
-              auto vTensor = mpsGraphRankedPlaceHolder(mpsGraph, v_);
+          auto maskedMM = [mpsGraph matrixMultiplicationWithPrimaryTensor:qTensor secondaryTensor:kT name:nil];
 
-              if (is_causal) {
-                MPSShape* maskShape = @[@(qSize), @(maxSeqLength)];
-                auto x = [mpsGraph coordinateAlongAxis:-1 withShape:@[@(qSize), @1] name:nil];
-                auto y = [mpsGraph coordinateAlongAxis:-2 withShape:@[@1, @(maxSeqLength)] name:nil];
-                auto isLess = [mpsGraph lessThanOrEqualToWithPrimaryTensor:x secondaryTensor:y name:nil];
-                auto causalMask = [mpsGraph selectWithPredicateTensor:isLess 
-                                            truePredicateTensor:[mpsGraph constantWithScalar:0 dataType:qTensor.dataType] 
-                                            falsePredicateTensor:[mpsGraph constantWithScalar:-INFINITY dataType:qTensor.dataType] 
-                                            name:nil];
-                graph->maskTensor = causalMask;
-              } else if (attn_mask) {
-                graph->maskTensor = mpsGraphRankedPlaceHolder(mpsGraph, *attn_mask);
-              }
+          if (macOS15_0_plus && [maskedMM dataType] == MPSDataTypeFloat32) {
+            // bug in MacOS15, without this trick SDPA leaks memory, adding 0.0f gets ignored(still takes SDPA sequence
+            // path which leaks)
+            auto oneTensor = [mpsGraph constantWithScalar:1e-20f shape:getMPSShape({1}) dataType:MPSDataTypeFloat32];
+            maskedMM = [mpsGraph additionWithPrimaryTensor:maskedMM secondaryTensor:oneTensor name:nil];
+          }
 
-              // Account for case where all values were masked causing division by 0 in softmax (issue:#156707)
-              // Overwrites expected NANs in sm with zeros.
-//              auto negInfTensor = [mpsGraph constantWithScalar:-INFINITY shape:maskedMM.shape dataType:maskedMM.dataType];
-//              auto elem_neg_inf = [mpsGraph equalWithPrimaryTensor:maskedMM secondaryTensor:negInfTensor name:nil];
-//              auto all_neg_infs_along_axis = [mpsGraph reductionAndWithTensor:elem_neg_inf axis:3 name:nil];
-//              auto zero_mask = [mpsGraph broadcastTensor:all_neg_infs_along_axis toShape:maskedMM.shape name:nil];
-//              auto zeroTensor = [mpsGraph constantWithScalar:0.0 shape:maskedMM.shape dataType:maskedMM.dataType];
-//
-//              auto sm = [mpsGraph softMaxWithTensor:maskedMM axis:3 name:nil];
-//              MPSGraphTensor* correctedSM = [mpsGraph selectWithPredicateTensor:zero_mask
-//                                                            truePredicateTensor:zeroTensor
-//                                                           falsePredicateTensor:sm
-//                                                                           name:nil];
-//
-//              auto output = [mpsGraph matrixMultiplicationWithPrimaryTensor:correctedSM secondaryTensor:vTensor name:nil];
+          // upcasting to float32 if needed to improve precision when multiplying by the scale factor
+          maskedMM = castMPSTensor(mpsGraph, maskedMM, MPSDataTypeFloat32);
+          maskedMM = [mpsGraph multiplicationWithPrimaryTensor:maskedMM secondaryTensor:scaleTensor name:nil];
 
-              MPSGraphTensor* output;
-              if(graph->maskTensor != nil) {
-                output = [mpsGraph scaledDotProductAttentionWithQueryTensor:qTensor 
-                                                          keyTensor:kTensor 
-                                                        valueTensor:vTensor
-                                                         maskTensor:graph->maskTensor
-                                                              scale:scale_factor
-                                                               name:@"MPSGraph SDPA"];
-              } else {
-                output = [mpsGraph scaledDotProductAttentionWithQueryTensor:qTensor 
-                                                          keyTensor:kTensor 
-                                                        valueTensor:vTensor
-                                                              scale:scale_factor
-                                                               name:@"MPSGraph SDPA"];
-              }
-              graph->qTensor = qTensor;
-              graph->kTensor = kTensor;
-              graph->vTensor = vTensor;
-              graph->outputTensor = castMPSTensor(mpsGraph, output, qTensor.dataType);
-//              graph->attnTensor = castMPSTensor(mpsGraph, sm, qTensor.dataType);
-            });
-    } else {
-        cachedGraph =
-            LookUpOrCreateCachedGraph<CachedGraph>(mkey, [&, q_ = query, k_ = key, v_ = value](auto mpsGraph, auto graph) {
-              auto qTensor = mpsGraphRankedPlaceHolder(mpsGraph, q_);
-              auto kTensor = mpsGraphRankedPlaceHolder(mpsGraph, k_);
-              auto vTensor = mpsGraphRankedPlaceHolder(mpsGraph, v_);
-              auto kT = [mpsGraph transposeTensor:kTensor dimension:2 withDimension:3 name:nil];
-              auto scaleTensor = [mpsGraph constantWithScalar:scale_factor
-                                                        shape:getMPSShape({1})
-                                                     dataType:MPSDataTypeFloat32];
+          if (is_causal) {
+            auto causalMask = [mpsGraph constantWithScalar:1.0f
+                                                     shape:getMPSShape({qSize, maxSeqLength})
+                                                  dataType:MPSDataTypeBool];
+            causalMask = [mpsGraph bandPartWithTensor:causalMask numLower:-1 numUpper:0 name:nil];
+            auto minusInf = [mpsGraph constantWithScalar:-1e20 shape:maskedMM.shape dataType:maskedMM.dataType];
+            maskedMM = [mpsGraph selectWithPredicateTensor:causalMask
+                                       truePredicateTensor:maskedMM
+                                      falsePredicateTensor:minusInf
+                                                      name:nil];
+          } else if (attn_mask) {
+            graph->maskTensor = mpsGraphRankedPlaceHolder(mpsGraph, *attn_mask);
+            maskedMM = [mpsGraph additionWithPrimaryTensor:maskedMM
+                                           secondaryTensor:castMPSTensor(mpsGraph, graph->maskTensor, maskedMM.dataType)
+                                                      name:nil];
+          }
 
-              auto maskedMM = [mpsGraph matrixMultiplicationWithPrimaryTensor:qTensor secondaryTensor:kT name:nil];
+          // Account for case where all values were masked causing division by 0 in softmax (issue:#156707)
+          // Overwrites expected NANs in sm with zeros.
+          auto negInfTensor = [mpsGraph constantWithScalar:-INFINITY shape:maskedMM.shape dataType:maskedMM.dataType];
+          auto elem_neg_inf = [mpsGraph equalWithPrimaryTensor:maskedMM secondaryTensor:negInfTensor name:nil];
+          auto all_neg_infs_along_axis = [mpsGraph reductionAndWithTensor:elem_neg_inf axis:3 name:nil];
+          auto zero_mask = [mpsGraph broadcastTensor:all_neg_infs_along_axis toShape:maskedMM.shape name:nil];
+          auto zeroTensor = [mpsGraph constantWithScalar:0.0 shape:maskedMM.shape dataType:maskedMM.dataType];
 
-              if (macOS15_0_plus && [maskedMM dataType] == MPSDataTypeFloat32) {
-                // bug in MacOS15, without this trick SDPA leaks memory, adding 0.0f gets ignored(still takes SDPA sequence
-                // path which leaks)
-                auto oneTensor = [mpsGraph constantWithScalar:1e-20f shape:getMPSShape({1}) dataType:MPSDataTypeFloat32];
-                maskedMM = [mpsGraph additionWithPrimaryTensor:maskedMM secondaryTensor:oneTensor name:nil];
-              }
+          auto sm = [mpsGraph softMaxWithTensor:maskedMM axis:3 name:nil];
+          MPSGraphTensor* correctedSM = [mpsGraph selectWithPredicateTensor:zero_mask
+                                                        truePredicateTensor:zeroTensor
+                                                       falsePredicateTensor:sm
+                                                                       name:nil];
 
-              // upcasting to float32 if needed to improve precision when multiplying by the scale factor
-              maskedMM = castMPSTensor(mpsGraph, maskedMM, MPSDataTypeFloat32);
-              maskedMM = [mpsGraph multiplicationWithPrimaryTensor:maskedMM secondaryTensor:scaleTensor name:nil];
-
-              if (is_causal) {
-                auto causalMask = [mpsGraph constantWithScalar:1.0f
-                                                         shape:getMPSShape({qSize, maxSeqLength})
-                                                      dataType:MPSDataTypeBool];
-                causalMask = [mpsGraph bandPartWithTensor:causalMask numLower:-1 numUpper:0 name:nil];
-                auto minusInf = [mpsGraph constantWithScalar:-1e20 shape:maskedMM.shape dataType:maskedMM.dataType];
-                maskedMM = [mpsGraph selectWithPredicateTensor:causalMask
-                                           truePredicateTensor:maskedMM
-                                          falsePredicateTensor:minusInf
-                                                          name:nil];
-              } else if (attn_mask) {
-                graph->maskTensor = mpsGraphRankedPlaceHolder(mpsGraph, *attn_mask);
-                maskedMM = [mpsGraph additionWithPrimaryTensor:maskedMM
-                                               secondaryTensor:castMPSTensor(mpsGraph, graph->maskTensor, maskedMM.dataType)
-                                                          name:nil];
-              }
-
-              // Account for case where all values were masked causing division by 0 in softmax (issue:#156707)
-              // Overwrites expected NANs in sm with zeros.
-              auto negInfTensor = [mpsGraph constantWithScalar:-INFINITY shape:maskedMM.shape dataType:maskedMM.dataType];
-              auto elem_neg_inf = [mpsGraph equalWithPrimaryTensor:maskedMM secondaryTensor:negInfTensor name:nil];
-              auto all_neg_infs_along_axis = [mpsGraph reductionAndWithTensor:elem_neg_inf axis:3 name:nil];
-              auto zero_mask = [mpsGraph broadcastTensor:all_neg_infs_along_axis toShape:maskedMM.shape name:nil];
-              auto zeroTensor = [mpsGraph constantWithScalar:0.0 shape:maskedMM.shape dataType:maskedMM.dataType];
-
-              auto sm = [mpsGraph softMaxWithTensor:maskedMM axis:3 name:nil];
-              MPSGraphTensor* correctedSM = [mpsGraph selectWithPredicateTensor:zero_mask
-                                                            truePredicateTensor:zeroTensor
-                                                           falsePredicateTensor:sm
-                                                                           name:nil];
-
-              auto output = [mpsGraph matrixMultiplicationWithPrimaryTensor:correctedSM secondaryTensor:vTensor name:nil];
-              graph->qTensor = qTensor;
-              graph->kTensor = kTensor;
-              graph->vTensor = vTensor;
-              graph->outputTensor = castMPSTensor(mpsGraph, output, qTensor.dataType);
-              graph->attnTensor = castMPSTensor(mpsGraph, sm, qTensor.dataType);
-            });
-    }
+          auto output = [mpsGraph matrixMultiplicationWithPrimaryTensor:correctedSM secondaryTensor:vTensor name:nil];
+          graph->qTensor = qTensor;
+          graph->kTensor = kTensor;
+          graph->vTensor = vTensor;
+          graph->outputTensor = castMPSTensor(mpsGraph, output, qTensor.dataType);
+          graph->attnTensor = castMPSTensor(mpsGraph, sm, qTensor.dataType);
+        });
     auto qPlaceholder = Placeholder(cachedGraph->qTensor, query);
     auto kPlaceholder = Placeholder(cachedGraph->kTensor, key);
     auto vPlaceholder = Placeholder(cachedGraph->vTensor, value);
     auto outputPlaceholder = Placeholder(cachedGraph->outputTensor, out);
-//    auto attnPlaceholder = Placeholder(cachedGraph->attnTensor, attn);
+    auto attnPlaceholder = Placeholder(cachedGraph->attnTensor, attn);
     NSDictionary* feeds = nil;
     if (!attn_mask) {
       feeds = dictionaryFromPlaceholders(qPlaceholder, kPlaceholder, vPlaceholder);
@@ -209,8 +145,7 @@ static std::tuple<Tensor, Tensor> sdpa_general_mps(const Tensor& query,
       auto mPlaceholder = Placeholder(cachedGraph->maskTensor, *attn_mask);
       feeds = dictionaryFromPlaceholders(qPlaceholder, kPlaceholder, vPlaceholder, mPlaceholder);
     }
-//    NSDictionary* outs = dictionaryFromPlaceholders(outputPlaceholder, attnPlaceholder);
-    NSDictionary* outs = dictionaryFromPlaceholders(outputPlaceholder);
+    NSDictionary* outs = dictionaryFromPlaceholders(outputPlaceholder, attnPlaceholder);
     runMPSGraph(getCurrentMPSStream(), cachedGraph->graph(), feeds, outs);
   }
 

aten/src/ATen/native/quantized/cpu/qnnpack/test/avgpool-microkernel-tester.h

@@ -301,12 +301,12 @@ class AvgPoolMicrokernelTester {
           ASSERT_NEAR(
               float(int32_t(y[i * yStride() + k])), yFP[i * kc() + k], 0.5001f)
               << "at pixel " << i << ", channel " << k << ", n = " << n()
-              << ", ks = " << kh() << "x" << kw() << " (" << ks()
+              << ", ks = " << kh() << 'x' << kw() << " (" << ks()
               << "), kc = " << kc() << ", acc = " << yAcc[i * kc() + k];
           ASSERT_EQ(
               uint32_t(yRef[i * kc() + k]), uint32_t(y[i * yStride() + k]))
               << "at pixel " << i << ", channel " << k << ", n = " << n()
-              << ", ks = " << kh() << "x" << kw() << " (" << ks()
+              << ", ks = " << kh() << 'x' << kw() << " (" << ks()
               << "), kc = " << kc() << ", acc = " << yAcc[i * kc() + k];
         }
       }
@@ -396,12 +396,12 @@ class AvgPoolMicrokernelTester {
           ASSERT_NEAR(
               float(int32_t(y[i * yStride() + k])), yFP[i * kc() + k], 0.5001f)
               << "at pixel " << i << ", channel " << k << ", n = " << n()
-              << ", ks = " << kh() << "x" << kw() << " (" << ks()
+              << ", ks = " << kh() << 'x' << kw() << " (" << ks()
               << "), kc = " << kc() << ", acc = " << yAcc[i * kc() + k];
           ASSERT_EQ(
               uint32_t(yRef[i * kc() + k]), uint32_t(y[i * yStride() + k]))
               << "at pixel " << i << ", channel " << k << ", n = " << n()
-              << ", ks = " << kh() << "x" << kw() << " (" << ks()
+              << ", ks = " << kh() << 'x' << kw() << " (" << ks()
               << "), kc = " << kc() << ", acc = " << yAcc[i * kc() + k];
         }
       }

aten/src/ATen/native/quantized/cpu/qnnpack/test/maxpool-microkernel-tester.h

@@ -232,7 +232,7 @@ class MaxPoolMicrokernelTester {
           ASSERT_EQ(
               uint32_t(yRef[i * kc() + k]), uint32_t(y[i * yStride() + k]))
               << "at pixel " << i << ", channel " << k << ", n = " << n()
-              << ", ks = " << kh() << "x" << kw() << " (" << ks()
+              << ", ks = " << kh() << 'x' << kw() << " (" << ks()
               << "), kc = " << kc();
         }
       }

aten/src/ATen/native/utils/ParamUtils.h

@@ -17,7 +17,7 @@ inline std::vector<T> _expand_param_if_needed(
     std::ostringstream ss;
     ss << "expected " << param_name << " to be a single integer value or a "
        << "list of " << expected_dim << " values to match the convolution "
-       << "dimensions, but got " << param_name << "=" << list_param;
+       << "dimensions, but got " << param_name << '=' << list_param;
     TORCH_CHECK(false, ss.str());
   } else {
     return list_param.vec();

aten/src/ATen/native/vulkan/api/Adapter.cpp

@@ -358,9 +358,9 @@ std::string Adapter::stringize() const {
   std::string device_type = get_device_type_str(properties.deviceType);
   VkPhysicalDeviceLimits limits = properties.limits;
 
-  ss << "{" << std::endl;
+  ss << '{' << std::endl;
   ss << "  Physical Device Info {" << std::endl;
-  ss << "    apiVersion:    " << v_major << "." << v_minor << std::endl;
+  ss << "    apiVersion:    " << v_major << '.' << v_minor << std::endl;
   ss << "    driverversion: " << properties.driverVersion << std::endl;
   ss << "    deviceType:    " << device_type << std::endl;
   ss << "    deviceName:    " << properties.deviceName << std::endl;
@@ -371,7 +371,7 @@ std::string Adapter::stringize() const {
 
 #define PRINT_LIMIT_PROP_VEC3(name)                                       \
   ss << "      " << std::left << std::setw(36) << #name << limits.name[0] \
-     << "," << limits.name[1] << "," << limits.name[2] << std::endl;
+     << ',' << limits.name[1] << ',' << limits.name[2] << std::endl;
 
   ss << "    Physical Device Limits {" << std::endl;
   PRINT_LIMIT_PROP(maxImageDimension1D);
@@ -425,7 +425,7 @@ std::string Adapter::stringize() const {
     ;
   }
   ss << "  ]" << std::endl;
-  ss << "}";
+  ss << '}';
 
   return ss.str();
 }

aten/src/ATen/native/vulkan/api/Exception.cpp

@@ -33,7 +33,7 @@ std::ostream& operator<<(std::ostream& out, const VkResult result) {
     VK_RESULT_CASE(VK_ERROR_FORMAT_NOT_SUPPORTED)
     VK_RESULT_CASE(VK_ERROR_FRAGMENTED_POOL)
     default:
-      out << "VK_ERROR_UNKNOWN (VkResult " << result << ")";
+      out << "VK_ERROR_UNKNOWN (VkResult " << result << ')';
       break;
   }
   return out;
@@ -46,7 +46,7 @@ std::ostream& operator<<(std::ostream& out, const VkResult result) {
 //
 
 std::ostream& operator<<(std::ostream& out, const SourceLocation& loc) {
-  out << loc.function << " at " << loc.file << ":" << loc.line;
+  out << loc.function << " at " << loc.file << ':' << loc.line;
   return out;
 }
 
@@ -66,7 +66,7 @@ Error::Error(SourceLocation source_location, const char* cond, std::string msg)
     : msg_(std::move(msg)), source_location_{source_location} {
   std::ostringstream oss;
   oss << "Exception raised from " << source_location_ << ": ";
-  oss << "(" << cond << ") is false! ";
+  oss << '(' << cond << ") is false! ";
   oss << msg_;
   what_ = oss.str();
 }

aten/src/ATen/native/vulkan/api/QueryPool.cpp

@@ -173,8 +173,8 @@ void QueryPool::extract_results() {
 
 static std::string stringize(const VkExtent3D& extents) {
   std::stringstream ss;
-  ss << "{" << extents.width << ", " << extents.height << ", " << extents.depth
-     << "}";
+  ss << '{' << extents.width << ", " << extents.height << ", " << extents.depth
+     << '}';
   return ss.str();
 }
 

aten/src/ATen/native/vulkan/api/Runtime.cpp

@@ -149,7 +149,7 @@ VKAPI_ATTR VkBool32 VKAPI_CALL debug_report_callback_fn(
   (void)flags;
 
   std::stringstream stream;
-  stream << layer_prefix << " " << message_code << " " << message << std::endl;
+  stream << layer_prefix << ' ' << message_code << ' ' << message << std::endl;
   const std::string log = stream.str();
 
   std::cout << log;

aten/src/ATen/native/vulkan/api/Utils.h

@@ -253,7 +253,7 @@ using vec4 = vec<4u>;
 
 // uvec3 is the type representing tensor extents. Useful for debugging.
 inline std::ostream& operator<<(std::ostream& os, const uvec3& v) {
-  os << "(" << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ")";
+  os << '(' << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ')';
   return os;
 }
 

aten/src/ATen/test/basic.cpp

@@ -246,7 +246,7 @@ void TestToCFloat() {
 void TestToString() {
   Tensor b = ones({3, 7}) * .0000001f;
   std::stringstream s;
-  s << b << "\n";
+  s << b << '\n';
   std::string expect = "1e-07 *";
   ASSERT_EQ_RESOLVED(s.str().substr(0, expect.size()), expect);
 }

aten/src/ATen/test/scalar_test.cpp

@@ -33,7 +33,7 @@ struct Foo {
   static void apply(Tensor a, Tensor b) {
     scalar_type s = 1;
     std::stringstream ss;
-    ss << "hello, dispatch: " << a.toString() << s << "\n";
+    ss << "hello, dispatch: " << a.toString() << s << '\n';
     auto data = (scalar_type*)a.data_ptr();
     (void)data;
   }
@@ -73,8 +73,8 @@ TEST(TestScalar, TestScalar) {
   Scalar bar = 3.0;
   Half h = bar.toHalf();
   Scalar h2 = h;
-  cout << "H2: " << h2.toDouble() << " " << what.toFloat() << " "
-       << bar.toDouble() << " " << what.isIntegral(false) << "\n";
+  cout << "H2: " << h2.toDouble() << ' ' << what.toFloat() << ' '
+       << bar.toDouble() << ' ' << what.isIntegral(false) << '\n';
   auto gen = at::detail::getDefaultCPUGenerator();
   {
     // See Note [Acquire lock when using random generators]
@@ -84,7 +84,7 @@ TEST(TestScalar, TestScalar) {
   }
   if (at::hasCUDA()) {
     auto t2 = zeros({4, 4}, at::kCUDA);
-    cout << &t2 << "\n";
+    cout << &t2 << '\n';
   }
   auto t = ones({4, 4});
 
@@ -129,7 +129,7 @@ TEST(TestScalar, TestScalar) {
       std::stringstream ss;
       // NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
       ASSERT_NO_THROW(
-          ss << "hello, dispatch" << x.toString() << s << "\n");
+          ss << "hello, dispatch" << x.toString() << s << '\n');
       auto data = (scalar_t*)x.data_ptr();
       (void)data;
     });

aten/src/ATen/test/test_install/main.cpp

@@ -1,5 +1,5 @@
 #include <ATen/ATen.h>
 
 int main() {
-  std::cout << at::ones({3,4}, at::CPU(at::kFloat)) << "\n";
+  std::cout << at::ones({3,4}, at::CPU(at::kFloat)) << '\n';
 }

aten/src/ATen/test/vec_test_all_types.cpp

@@ -1828,9 +1828,9 @@ namespace {
       #endif
 
         EXPECT_EQ(u16, c10::detail::fp16_ieee_from_fp32_value(f32s[i]))
-            << "Test failed for float to uint16 " << f32s[i] << "\n";
+            << "Test failed for float to uint16 " << f32s[i] << '\n';
         EXPECT_EQ(x, c10::detail::fp16_ieee_to_fp32_value(u16))
-            << "Test failed for uint16 to float " << u16 << "\n";
+            << "Test failed for uint16 to float " << u16 << '\n';
       }
     }
     TEST(FP8E4M3Test, FP8E4M3ConversionFloat) {
@@ -1848,10 +1848,10 @@ namespace {
           EXPECT_TRUE(std::isnan(f32));
         } else {
           EXPECT_EQ(f32, c10::detail::fp8e4m3fn_to_fp32_value(input))
-              << "Test failed for u8 to float " << input << "\n";
+              << "Test failed for u8 to float " << input << '\n';
         }
         EXPECT_EQ(u8, c10::detail::fp8e4m3fn_from_fp32_value(f32))
-            << "Test failed for float to u8 " << f32 << "\n";
+            << "Test failed for float to u8 " << f32 << '\n';
       }
     }
     TEST(FP8E4M3Test, FP8E4M3BinaryAdd) {
@@ -2015,10 +2015,10 @@ namespace {
           EXPECT_TRUE(std::isnan(f32));
         } else {
           EXPECT_EQ(f32, c10::detail::fp8e5m2_to_fp32_value(input))
-              << "Test failed for u8 to float " << input << "\n";
+              << "Test failed for u8 to float " << input << '\n';
         }
         EXPECT_EQ(u8, c10::detail::fp8e5m2_from_fp32_value(f32))
-            << "Test failed for float to u8 " << f32 << "\n";
+            << "Test failed for float to u8 " << f32 << '\n';
       }
     }
     TEST(FP8E5M2Test, FP8E5M2BinaryAdd) {

aten/src/ATen/test/vitals.cpp

@@ -19,7 +19,7 @@ TEST(Vitals, Basic) {
     c10::utils::set_env("TORCH_VITAL", "1");
     TORCH_VITAL_DEFINE(Testing);
     TORCH_VITAL(Testing, Attribute0) << 1;
-    TORCH_VITAL(Testing, Attribute1) << "1";
+    TORCH_VITAL(Testing, Attribute1) << '1';
     TORCH_VITAL(Testing, Attribute2) << 1.0f;
     TORCH_VITAL(Testing, Attribute3) << 1.0;
     auto t = at::ones({1, 1});

aten/src/ATen/test/vulkan_api_test.cpp

@@ -129,14 +129,14 @@ void showRtol(const at::Tensor& a, const at::Tensor& b) {
   std::cout << "Max Diff allowed: " << maxDiff << std::endl;
   if (diff.sizes().size() == 2) {
     for (const auto y : c10::irange(diff.sizes()[0])) {
-      std::cout << y << ":";
+      std::cout << y << ':';
       for (const auto x : c10::irange(diff.sizes()[1])) {
         float diff_xy = diff[y][x].item<float>();
         if (diff_xy > maxDiff) {
           std::cout << std::setw(5) << x;
         }
         else {
-          std::cout << std::setw(5) << " ";
+          std::cout << std::setw(5) << ' ';
         }
       }
       std::cout << std::endl;
@@ -3276,7 +3276,7 @@ TEST_F(VulkanAPITest, masked_fill_invalidinputs_exceptions) {
 
 void print_shape(const std::vector<int64_t>& shape) {
   for (const auto& num : shape) {
-    std::cout << num << " ";
+    std::cout << num << ' ';
   }
 }
 
@@ -3367,7 +3367,7 @@ void test_masked_fill_scalar(
             print_shape(tmp_curr_input_shape);
             std::cout << "], and mask of shape [";
             print_shape(tmp_curr_mask_shape);
-            std::cout << "]" << std::endl;
+            std::cout << ']' << std::endl;
           }
 
           ASSERT_TRUE(check);
@@ -4542,9 +4542,9 @@ void test_softmax(const at::IntArrayRef shape, bool log_softmax = false) {
     if (!check) {
       std::cout << "Softmax test failed on axis " << dim << "for tensor dims {";
       for (uint32_t place = 0; place < shape.size() - 1; place++) {
-        std::cout << shape[place] << " ";
+        std::cout << shape[place] << ' ';
       }
-      std::cout << shape.back() << "}" << std::endl;
+      std::cout << shape.back() << '}' << std::endl;
       showRtol(out_cpu, out_vulkan.cpu());
     }
     ASSERT_TRUE(check);

aten/src/ATen/test/vulkan_quantized_api_test.cpp

@@ -95,7 +95,7 @@ void showRtol(
   std::cout << "Max Diff found is: " << diff.max().item<double>() << std::endl;
   if (diff.sizes().size() == 2) {
     for (const auto y : c10::irange(diff.sizes()[0])) {
-      std::cout << y << ":";
+      std::cout << y << ':';
       for (const auto x : c10::irange(diff.sizes()[1])) {
         double diff_xy = diff[y][x].item<double>();
         if (diff_xy > maxDiff) {
@@ -109,7 +109,7 @@ void showRtol(
             }
           }
         } else {
-          std::cout << std::setw(5) << " ";
+          std::cout << std::setw(5) << ' ';
         }
       }
       std::cout << std::endl;
@@ -148,19 +148,19 @@ using at::native::vulkan::api::utils::ivec4;
 using at::native::vulkan::api::utils::vec4;
 
 std::ostream& operator<<(std::ostream& os, const vec4& v) {
-  os << "(" << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ", "
-     << v.data[3u] << ")";
+  os << '(' << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ", "
+     << v.data[3u] << ')';
   return os;
 }
 
 std::ostream& operator<<(std::ostream& os, const ivec3& v) {
-  os << "(" << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ")";
+  os << '(' << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ')';
   return os;
 }
 
 std::ostream& operator<<(std::ostream& os, const ivec4& v) {
-  os << "(" << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ", "
-     << v.data[3u] << ")";
+  os << '(' << v.data[0u] << ", " << v.data[1u] << ", " << v.data[2u] << ", "
+     << v.data[3u] << ')';
   return os;
 }
 
@@ -3379,51 +3379,51 @@ bool _test_quantized_linear(
       showRtol(out_cpu_dequant, out_vk_to_cpu_dequant);
     }
     if (xpos != -1 && ypos != -1) {
-      std::cout << "\nFailure caused on row/col: " << ypos << "/" << xpos
-                << "\n";
+      std::cout << "\nFailure caused on row/col: " << ypos << '/' << xpos
+                << '\n';
       std::cout << "Input tensor scale: " << scale << " zerop: " << zero_point
-                << "\n";
-      std::cout << "Input tensor row " << ypos << "\n";
+                << '\n';
+      std::cout << "Input tensor row " << ypos << '\n';
       for (int i = 0; i < input_cpu.sizes()[1]; i++) {
         std::cout << input_cpu[ypos][i].item<double>() << ", ";
       }
-      std::cout << "\n";
+      std::cout << '\n';
 
       std::cout << "Weight tensor scale: " << w_scale
-                << " zerop: " << w_zero_point << "\n";
-      std::cout << "Weight tensor col " << xpos << "\n";
+                << " zerop: " << w_zero_point << '\n';
+      std::cout << "Weight tensor col " << xpos << '\n';
       for (int i = 0; i < weight.sizes()[1]; i++) {
         std::cout << weight[xpos][i].item<double>() << ", ";
       }
-      std::cout << "\n";
+      std::cout << '\n';
 
       std::cout << "Input tensor quantized row " << ypos << " with dtype "
-                << (input_quant_dtype_int8 ? "QInt8" : "QUInt8") << "\n";
+                << (input_quant_dtype_int8 ? "QInt8" : "QUInt8") << '\n';
       for (int i = 0; i < input_cpu.sizes()[1]; i++) {
         std::cout << input_cpu_quantized[ypos][i].item<double>() << ", ";
       }
-      std::cout << "\n";
+      std::cout << '\n';
 
       std::cout << "Weight tensor quantized col " << xpos << " with dtype "
-                << (weight_quant_dtype_int8 ? "QInt8" : "QUInt8") << "\n";
+                << (weight_quant_dtype_int8 ? "QInt8" : "QUInt8") << '\n';
       for (int i = 0; i < weight.sizes()[1]; i++) {
         std::cout << weight_cpu_quantized[xpos][i].item<double>() << ", ";
       }
-      std::cout << "\n";
+      std::cout << '\n';
 
       std::cout << "bias tensor\n";
       for (int i = 0; i < bias.sizes()[0]; i++) {
         std::cout << bias[i].item<double>() << ", ";
       }
-      std::cout << "\n";
+      std::cout << '\n';
 
       std::cout << "out_scale: " << out_scale
-                << " out_zero_point: " << out_zero_point << "\n";
+                << " out_zero_point: " << out_zero_point << '\n';
 
       std::cout << "cpu unmatched output: "
-                << out_cpu_dequant[ypos][xpos].item<double>() << "\n";
+                << out_cpu_dequant[ypos][xpos].item<double>() << '\n';
       std::cout << "vk unmatched output: "
-                << out_vk_to_cpu_dequant[ypos][xpos].item<double>() << "\n";
+                << out_vk_to_cpu_dequant[ypos][xpos].item<double>() << '\n';
     }
   }
   return check;

aten/src/ATen/xpu/XPUScaledBlas.cpp

@@ -0,0 +1,122 @@
+#include <c10/core/Scalar.h>
+#include <c10/core/ScalarType.h>
+#include <c10/util/Exception.h>
+#include <c10/util/SmallVector.h>
+#include <c10/util/typeid.h>
+#include <cstdint>
+#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
+#include <ATen/BlasBackend.h>
+#include <ATen/Dispatch.h>
+#include <ATen/ExpandUtils.h>
+#include <ATen/OpMathType.h>
+#include <ATen/TensorUtils.h>
+#include <ATen/core/NamedTensor.h>
+#include <ATen/core/Tensor.h>
+#include <ATen/native/GroupedMMUtils.h>
+#include <ATen/native/Resize.h>
+#include <c10/util/MaybeOwned.h>
+
+#include <ATen/ceil_div.h>
+#include <ATen/xpu/XPUScaledBlas.h>
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/Functions.h>
+#include <ATen/NativeFunctions.h>
+#else
+#include <ATen/ops/_addmm_activation_native.h>
+#include <ATen/ops/_efficientzerotensor.h>
+#include <ATen/ops/_scaled_mm_native.h>
+#include <ATen/ops/_unsafe_view_native.h>
+#include <ATen/ops/abs.h>
+#include <ATen/ops/addmm_native.h>
+#include <ATen/ops/addmv_native.h>
+#include <ATen/ops/baddbmm_native.h>
+#include <ATen/ops/bmm_native.h>
+#include <ATen/ops/copy_native.h>
+#include <ATen/ops/dot_native.h>
+#include <ATen/ops/empty.h>
+#include <ATen/ops/empty_strided.h>
+#include <ATen/ops/gelu.h>
+#include <ATen/ops/max.h>
+#include <ATen/ops/mm_native.h>
+#include <ATen/ops/mul.h>
+#include <ATen/ops/ones.h>
+#include <ATen/ops/relu.h>
+#include <ATen/ops/scalar_tensor_native.h>
+#include <ATen/ops/vdot_native.h>
+#endif
+
+using at::blas::ScalingType;
+
+namespace at::native::onednn::scaled {
+
+/**
+ * Both inputs must be fp8,
+ * Each needs a single scale, {Tensorwise (float)}
+ */
+bool check_tensorwise_recipe(
+    c10::ScalarType type_a,
+    std::vector<ScalingType>& recipe_a,
+    ArrayRef<Tensor>& scales_a,
+    c10::ScalarType type_b,
+    std::vector<ScalingType>& recipe_b,
+    ArrayRef<Tensor>& scales_b) {
+  // both types must be fp8
+  if (!isFloat8Type(type_a) || !isFloat8Type(type_b)) {
+    return false;
+  }
+
+  // 1 scale each, {Tensorwise, float}
+  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 ||
+      recipe_b.size() != 1) {
+    return false;
+  }
+  // Need {Blockwise_1x32, e8m0} for A & B
+  if (recipe_a[0] != ScalingType::TensorWise)
+    return false;
+  if (scales_a[0].scalar_type() != ScalarType::Float)
+    return false;
+  if (recipe_b[0] != ScalingType::TensorWise)
+    return false;
+  if (scales_b[0].scalar_type() != ScalarType::Float)
+    return false;
+
+  return true;
+}
+
+/**
+ * Both inputs must be fp8,
+ * Each needs scales, {Rowwise (float)}
+ */
+bool check_rowwise_recipe(
+    c10::ScalarType type_a,
+    std::vector<ScalingType>& recipe_a,
+    ArrayRef<Tensor>& scales_a,
+    c10::ScalarType type_b,
+    std::vector<ScalingType>& recipe_b,
+    ArrayRef<Tensor>& scales_b) {
+  // both types must be fp8
+  if (!isFloat8Type(type_a) || !isFloat8Type(type_b)) {
+    return false;
+  }
+
+  // 1 scale each, {Tensorwise, float}
+  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 ||
+      recipe_b.size() != 1) {
+    return false;
+  }
+
+  // Need {RowWise, dp32} for A & B
+  if (recipe_a[0] != ScalingType::RowWise)
+    return false;
+  if (scales_a[0].scalar_type() != ScalarType::Float)
+    return false;
+  if (recipe_b[0] != ScalingType::RowWise)
+    return false;
+  if (scales_b[0].scalar_type() != ScalarType::Float)
+    return false;
+
+  return true;
+}
+
+} // namespace at::native::onednn::scaled

aten/src/ATen/xpu/XPUScaledBlas.h

@@ -0,0 +1,95 @@
+#include <c10/core/Scalar.h>
+#include <c10/core/ScalarType.h>
+#include <c10/util/Exception.h>
+#include <c10/util/SmallVector.h>
+#include <c10/util/typeid.h>
+#include <cstdint>
+#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
+#include <ATen/Dispatch.h>
+#include <ATen/ExpandUtils.h>
+#include <ATen/OpMathType.h>
+#include <ATen/TensorUtils.h>
+#include <ATen/core/NamedTensor.h>
+#include <ATen/core/Tensor.h>
+#include <ATen/native/Resize.h>
+#include <c10/util/MaybeOwned.h>
+
+#include <ATen/BlasBackend.h>
+#include <ATen/ceil_div.h>
+
+#ifdef USE_FBGEMM_GENAI
+#include <fbgemm_gpu/torch_ops.h>
+#endif
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/Functions.h>
+#include <ATen/NativeFunctions.h>
+#else
+#include <ATen/ops/_addmm_activation_native.h>
+#include <ATen/ops/_efficientzerotensor.h>
+#include <ATen/ops/_scaled_mm_native.h>
+#include <ATen/ops/_unsafe_view_native.h>
+#include <ATen/ops/abs.h>
+#include <ATen/ops/addmm_native.h>
+#include <ATen/ops/addmv_native.h>
+#include <ATen/ops/baddbmm_native.h>
+#include <ATen/ops/bmm_native.h>
+#include <ATen/ops/copy_native.h>
+#include <ATen/ops/dot_native.h>
+#include <ATen/ops/empty.h>
+#include <ATen/ops/empty_strided.h>
+#include <ATen/ops/gelu.h>
+#include <ATen/ops/max.h>
+#include <ATen/ops/mm_native.h>
+#include <ATen/ops/mul.h>
+#include <ATen/ops/ones.h>
+#include <ATen/ops/relu.h>
+#include <ATen/ops/scalar_tensor_native.h>
+#include <ATen/ops/vdot_native.h>
+#endif
+
+using at::blas::ScalingType;
+
+namespace at::native::onednn::scaled {
+
+/**
+ * Track concrete implementations available
+ */
+enum class ScaledGemmImplementation {
+  NONE = 0,
+  TENSORWISE_TENSORWISE = 1,
+  ROWWISE_ROWWISE = 2,
+};
+
+/**
+ * Convert passed int (enum) from python back into a
+ * strictly-typed enum
+ */
+template <class EnumType, class ArrayType>
+std::vector<EnumType> convert_int_to_enum(ArrayType& v) {
+  std::vector<EnumType> converted;
+  converted.reserve(v.size());
+
+  for (auto vi : v) {
+    converted.push_back(static_cast<EnumType>(vi));
+  }
+  return converted;
+}
+
+bool check_tensorwise_recipe(
+    c10::ScalarType,
+    std::vector<ScalingType>&,
+    ArrayRef<Tensor>&,
+    c10::ScalarType,
+    std::vector<ScalingType>&,
+    ArrayRef<Tensor>&);
+
+bool check_rowwise_recipe(
+    c10::ScalarType,
+    std::vector<ScalingType>&,
+    ArrayRef<Tensor>&,
+    c10::ScalarType,
+    std::vector<ScalingType>&,
+    ArrayRef<Tensor>&);
+
+} // namespace at::native::onednn::scaled

c10/core/DispatchKeySet.cpp

@@ -176,7 +176,7 @@ std::ostream& operator<<(std::ostream& os, DispatchKeySet ts) {
     os << k;
     first = false;
   }
-  os << ")";
+  os << ')';
   return os;
 }
 

c10/core/TensorOptions.cpp

@@ -33,7 +33,7 @@ std::ostream& operator<<(std::ostream& stream, const TensorOptions& options) {
   } else {
     stream << "(nullopt)";
   }
-  stream << ")";
+  stream << ')';
 
   return stream;
 }

c10/cuda/CUDADeviceAssertionHost.cpp

@@ -136,7 +136,7 @@ std::string c10_retrieve_device_side_assertion_info() {
     // Something failed, let's talk about that
     oss << failures_found
         << " CUDA device-side assertion failures were found on GPU #"
-        << device_num << "!" << std::endl;
+        << device_num << '!' << std::endl;
     if (assertion_data_for_device.assertion_count >
         C10_CUDA_DSA_ASSERTION_COUNT) {
       oss << "But at least " << assertion_data_for_device.assertion_count
@@ -151,17 +151,17 @@ std::string c10_retrieve_device_side_assertion_info() {
       oss << "Assertion failure " << i << std::endl;
       oss << "  GPU assertion failure message = " << self.assertion_msg
           << std::endl;
-      oss << "  File containing assertion = " << self.filename << ":"
+      oss << "  File containing assertion = " << self.filename << ':'
           << self.line_number << std::endl;
       oss << "  Device function containing assertion = " << self.function_name
           << std::endl;
-      oss << "  Thread ID that failed assertion = [" << self.thread_id[0] << ","
-          << self.thread_id[1] << "," << self.thread_id[2] << "]" << std::endl;
-      oss << "  Block ID that failed assertion = [" << self.block_id[0] << ","
-          << self.block_id[1] << "," << self.block_id[2] << "]" << std::endl;
+      oss << "  Thread ID that failed assertion = [" << self.thread_id[0] << ','
+          << self.thread_id[1] << ',' << self.thread_id[2] << ']' << std::endl;
+      oss << "  Block ID that failed assertion = [" << self.block_id[0] << ','
+          << self.block_id[1] << ',' << self.block_id[2] << ']' << std::endl;
       if (launch_info.generation_number == self.caller) {
         oss << "  File containing kernel launch = "
-            << launch_info.launch_filename << ":" << launch_info.launch_linenum
+            << launch_info.launch_filename << ':' << launch_info.launch_linenum
             << std::endl;
         oss << "  Function containing kernel launch = "
             << launch_info.launch_function << std::endl;
@@ -175,7 +175,7 @@ std::string c10_retrieve_device_side_assertion_info() {
         if (launch_registry.gather_launch_stacktrace) {
           oss << "Launch stacktracing disabled." << std::endl;
         } else {
-          oss << "\n" << launch_info.launch_stacktrace << std::endl;
+          oss << '\n' << launch_info.launch_stacktrace << std::endl;
         }
       } else {
         oss << "  CPU launch site info: Unavailable, the circular queue wrapped around. Increase `CUDAKernelLaunchRegistry::max_size`."

c10/cuda/driver_api.h

@@ -20,6 +20,22 @@
     }                                                                      \
   } while (0)
 
+#define C10_CUDA_DRIVER_CHECK_GOTO(EXPR, NEXT)                             \
+  do {                                                                     \
+    CUresult __err = EXPR;                                                 \
+    if (__err != CUDA_SUCCESS) {                                           \
+      const char* err_str;                                                 \
+      CUresult get_error_str_err [[maybe_unused]] =                        \
+          c10::cuda::DriverAPI::get()->cuGetErrorString_(__err, &err_str); \
+      if (get_error_str_err != CUDA_SUCCESS) {                             \
+        TORCH_WARN("CUDA driver error: unknown error");                    \
+      } else {                                                             \
+        TORCH_WARN("CUDA driver error: ", err_str);                        \
+      }                                                                    \
+      goto NEXT;                                                           \
+    }                                                                      \
+  } while (0)
+
 // The integer in the second column specifies the requested CUDA Driver API
 // version. The dynamic loader will accept a driver with a newer version, but it
 // ensures that the requested symbol exists in *at least* the specified version

c10/test/core/DispatchKeySet_test.cpp

@@ -435,7 +435,7 @@ TEST(DispatchKeySet, TestFunctionalityDispatchKeyToString) {
     if (i > 0) {
       ASSERT_TRUE(res.find("Unknown") == std::string::npos)
           << i << " (before is " << toString(static_cast<DispatchKey>(i - 1))
-          << ")";
+          << ')';
     } else {
       ASSERT_TRUE(res.find("Unknown") == std::string::npos) << i;
     }

c10/test/util/Half_test.cpp

@@ -96,10 +96,10 @@ TEST(HalfConversionTest, TestPorableConversion) {
   for (auto x : inputs) {
     auto target = c10::detail::fp16_ieee_to_fp32_value(x);
     EXPECT_EQ(halfbits2float(x), target)
-        << "Test failed for uint16 to float " << x << "\n";
+        << "Test failed for uint16 to float " << x << '\n';
     EXPECT_EQ(
         float2halfbits(target), c10::detail::fp16_ieee_from_fp32_value(target))
-        << "Test failed for float to uint16" << target << "\n";
+        << "Test failed for float to uint16" << target << '\n';
   }
 }
 

c10/test/util/logging_test.cpp

@@ -98,7 +98,7 @@ struct Noncopyable {
 };
 
 std::ostream& operator<<(std::ostream& out, const Noncopyable& nc) {
-  out << "Noncopyable(" << nc.x << ")";
+  out << "Noncopyable(" << nc.x << ')';
   return out;
 }
 } // namespace

c10/util/ArrayRef.h

@@ -204,13 +204,13 @@ ArrayRef(const std::initializer_list<T>&) -> ArrayRef<T>;
 template <typename T>
 std::ostream& operator<<(std::ostream& out, ArrayRef<T> list) {
   int i = 0;
-  out << "[";
+  out << '[';
   for (const auto& e : list) {
     if (i++ > 0)
       out << ", ";
     out << e;
   }
-  out << "]";
+  out << ']';
   return out;
 }