Automated submodule update: FBGEMM

This PR fixes code to use c10::filesystem functionality instead of manually implemented functions.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/167821
Approved by: https://github.com/Skylion007

.ci/aarch64_linux/README.md

@@ -1,19 +0,0 @@
-# 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

@@ -1,53 +0,0 @@
-#!/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

@@ -1,21 +0,0 @@
-#!/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

@@ -1,333 +0,0 @@
-#!/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

@@ -1,999 +0,0 @@
-#!/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

@@ -1,87 +0,0 @@
-#!/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/manywheel/build.sh

@@ -4,14 +4,17 @@ 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 | cuda-aarch64)
         bash "${SCRIPTPATH}/build_cuda.sh"
         ;;
     rocm)
         bash "${SCRIPTPATH}/build_rocm.sh"
         ;;
-    cpu | cpu-cxx11-abi | cpu-s390x)
+    cpu | cpu-cxx11-abi | cpu-aarch64 | cpu-s390x)
         bash "${SCRIPTPATH}/build_cpu.sh"
         ;;
     xpu)

.ci/manywheel/build_common.sh

@@ -18,12 +18,31 @@ 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
-    PLATFORM="manylinux_2_28_x86_64"
+    # 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: $ARCH"
+            exit 1
+            ;;
+    esac
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
     retry dnf install -q -y zip openssl
 elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
@@ -38,6 +57,8 @@ else
     exit 1
 fi
 
+echo "Platform set to: $PLATFORM"
+
 # We use the package name to test the package by passing this to 'pip install'
 # This is the env variable that setup.py uses to name the package. Note that
 # pip 'normalizes' the name first by changing all - to _
@@ -299,8 +320,8 @@ for pkg in /$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 $destpath)
-            # Keep the so number for XPU dependencies and libgomp.so.1 to avoid twice load
-            elif [[ "$DESIRED_CUDA" == *"xpu"* || "$filename" == "libgomp.so.1" ]]; then
+            # Keep the so number for XPU dependencies, libgomp.so.1, ACL libraries, and NVPL libraries to avoid twice load
+            elif [[ "$DESIRED_CUDA" == *"xpu"* || "$filename" == "libgomp.so.1" || "$filename" == libarm_compute* || "$filename" == libnvpl* || "$filename" == "libgfortran.so.5" ]]; then
                 patchedpath=$destpath
             else
                 patchedpath=$(fname_with_sha256 $destpath)
@@ -346,9 +367,22 @@ for pkg in /$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
-    if [[ $PLATFORM == "manylinux_2_28_x86_64" && $GPU_ARCH_TYPE != "cpu-s390x" && $GPU_ARCH_TYPE != "xpu" ]]; then
+    # Support all architectures (x86_64, aarch64, s390x)
+    if [[ "$IS_MANYLINUX2_28" == "1" && $GPU_ARCH_TYPE != "xpu" ]]; then
         wheel_file=$(echo $(basename $pkg) | sed -e 's/-cp.*$/.dist-info\/WHEEL/g')
-        sed -i -e s#linux_x86_64#"${PLATFORM}"# $wheel_file;
+        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
     fi
 
     # regenerate the RECORD file with new hashes

.ci/manywheel/build_cpu.sh

@@ -15,6 +15,10 @@ 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
@@ -34,8 +38,10 @@ elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
 elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
     LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
 elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
-    if [[ "$(uname -m)" == "s390x" ]]; then
+    if [[ "$ARCH" == "s390x" ]]; then
         LIBGOMP_PATH="/usr/lib/s390x-linux-gnu/libgomp.so.1"
+    elif [[ "$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
@@ -49,6 +55,32 @@ 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"
+    )
+    DEPS_SONAME+=(
+        "libgfortran.so.5"
+    )
+fi
+
 rm -rf /usr/local/cuda*
 
 SOURCE_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null && pwd )"

.ci/manywheel/build_cuda.sh

@@ -29,6 +29,10 @@ if [[ -z "$EXTRA_CAFFE2_CMAKE_FLAGS" ]]; then
     EXTRA_CAFFE2_CMAKE_FLAGS=()
 fi
 
+# Detect architecture
+ARCH=$(uname -m)
+echo "Building for architecture: $ARCH"
+
 # Determine CUDA version and architectures to build for
 #
 # NOTE: We should first check `DESIRED_CUDA` when determining `CUDA_VERSION`,
@@ -53,34 +57,60 @@ fi
 cuda_version_nodot=$(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 ${CUDA_VERSION} in
-    #removing sm_50-sm_60 as these architectures are deprecated in CUDA 12.8/9 and will be removed in future releases
-    #however we would like to keep sm_70 architecture see: https://github.com/pytorch/pytorch/issues/157517
-    12.8)
-        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0"
-        ;;
-    12.9)
-        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0+PTX"
-        # WAR to resolve the ld error in libtorch build with CUDA 12.9
+    12.6) TORCH_CUDA_ARCH_LIST="5.0;6.0;${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
         if [[ "$PACKAGE_TYPE" == "libtorch" ]]; then
-            TORCH_CUDA_ARCH_LIST="7.5;8.0;9.0;10.0;12.0+PTX"
+            TORCH_CUDA_ARCH_LIST="${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
         fi
         ;;
     13.0)
-        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 $CUDA_VERSION"
-        exit 1
+        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;$([[ "$ARCH" == "aarch64" ]] && echo "11.0;" || echo "")12.0+PTX"
+        export TORCH_NVCC_FLAGS="-compress-mode=size"
+        export BUILD_BUNDLE_PTXAS=1
         ;;
+    *) 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"
@@ -244,6 +274,51 @@ 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"
 
@@ -251,9 +326,11 @@ 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
-rm -rf /usr/local/magma || true
-ln -s /usr/local/cuda-${CUDA_VERSION}/magma /usr/local/magma
+# Switch `/usr/local/magma` to the desired CUDA version (skip for aarch64)
+if [[ "$ARCH" != "aarch64" ]]; then
+    rm -rf /usr/local/magma || true
+    ln -s /usr/local/cuda-${CUDA_VERSION}/magma /usr/local/magma
+fi
 
 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/pytorch/build.sh

@@ -86,10 +86,20 @@ 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/common_utils.sh

@@ -96,7 +96,6 @@ function pip_build_and_install() {
     python3 -m pip wheel \
       --no-build-isolation \
       --no-deps \
-      --no-use-pep517 \
       -w "${wheel_dir}" \
       "${build_target}"
   fi
@@ -308,6 +307,28 @@ function install_torchao() {
   pip_build_and_install "git+https://github.com/pytorch/ao.git@${commit}" dist/ao
 }
 
+function install_flash_attn_cute() {
+  echo "Installing FlashAttention CuTe from GitHub..."
+  # Grab latest main til we have a pinned commit
+  local flash_attn_commit
+  flash_attn_commit=$(git ls-remote https://github.com/Dao-AILab/flash-attention.git HEAD | cut -f1)
+
+  # Clone the repo to a temporary directory
+  rm -rf flash-attention-build
+  git clone --depth 1 --recursive https://github.com/Dao-AILab/flash-attention.git flash-attention-build
+
+  pushd flash-attention-build
+  git checkout "${flash_attn_commit}"
+
+  # Install only the 'cute' sub-directory
+  pip_install -e flash_attn/cute/
+  popd
+
+  # remove the local repo
+  rm -rf flash-attention-build
+  echo "FlashAttention CuTe installation complete."
+}
+
 function print_sccache_stats() {
   echo 'PyTorch Build Statistics'
   sccache --show-stats

.ci/pytorch/smoke_test/smoke_test.py

@@ -353,6 +353,17 @@ def test_linalg(device="cpu") -> None:
             torch.linalg.svd(A)
 
 
+def test_sdpa(device="cpu", dtype=torch.float16) -> None:
+    """Regression test for https://github.com/pytorch/pytorch/issues/167602
+    Without nvrtc_builtins on CuDNN-9.13 on CUDA-13 fails with ` No valid execution plans built.`
+    """
+    print(f"Testing SDPA on {device} using type {dtype}")
+    k, q, v = torch.rand(3, 1, 16, 77, 64, dtype=dtype, device=device).unbind(0)
+    attn = torch.rand(1, 1, 77, 77, dtype=dtype, device=device)
+    rc = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn)
+    assert rc.isnan().any().item() is False
+
+
 def smoke_test_compile(device: str = "cpu") -> None:
     supported_dtypes = [torch.float16, torch.float32, torch.float64]
 
@@ -489,10 +500,12 @@ def main() -> None:
     smoke_test_conv2d()
     test_linalg()
     test_numpy()
+    test_sdpa()
 
     if is_cuda_system:
         test_linalg("cuda")
         test_cuda_gds_errors_captured()
+        test_sdpa("cuda")
 
     if options.package == "all":
         smoke_test_modules()

.ci/pytorch/test.sh

@@ -344,8 +344,18 @@ test_python_smoke() {
 }
 
 test_python_smoke_b200() {
-  # Targeted smoke tests for B200 - staged approach to avoid too many failures
-  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  # Targeted smoke tests for B200 including FlashAttention CuTe coverage
+  install_flash_attn_cute
+  time python test/run_test.py \
+    --include \
+      test_matmul_cuda \
+      test_scaled_matmul_cuda \
+      inductor/test_fp8 \
+      nn/attention/test_fa4 \
+      nn/attention/test_open_registry \
+      inductor/test_flex_flash \
+    $PYTHON_TEST_EXTRA_OPTION \
+    --upload-artifacts-while-running
   assert_git_not_dirty
 }
 
@@ -379,6 +389,13 @@ test_lazy_tensor_meta_reference_disabled() {
   export -n TORCH_DISABLE_FUNCTIONALIZATION_META_REFERENCE
 }
 
+test_dynamo_core() {
+  time python test/run_test.py \
+    --include-dynamo-core-tests \
+    --verbose \
+    --upload-artifacts-while-running
+  assert_git_not_dirty
+}
 
 test_dynamo_wrapped_shard() {
   if [[ -z "$NUM_TEST_SHARDS" ]]; then
@@ -1670,6 +1687,22 @@ test_operator_microbenchmark() {
   done
 }
 
+test_attention_microbenchmark() {
+  TEST_REPORTS_DIR=$(pwd)/test/test-reports
+  mkdir -p "$TEST_REPORTS_DIR"
+  TEST_DIR=$(pwd)
+
+  # Install attention-gym dependency
+  echo "Installing attention-gym..."
+  python -m pip install git+https://github.com/meta-pytorch/attention-gym.git@main
+  pip show triton
+
+  cd "${TEST_DIR}"/benchmarks/transformer
+
+  $TASKSET python score_mod.py --config configs/config_basic.yaml \
+    --output-json-for-dashboard "${TEST_REPORTS_DIR}/attention_microbenchmark.json"
+}
+
 if ! [[ "${BUILD_ENVIRONMENT}" == *libtorch* || "${BUILD_ENVIRONMENT}" == *-bazel-* ]]; then
   (cd test && python -c "import torch; print(torch.__config__.show())")
   (cd test && python -c "import torch; print(torch.__config__.parallel_info())")
@@ -1727,6 +1760,8 @@ elif [[ "${TEST_CONFIG}" == *operator_benchmark* ]]; then
   fi
 elif [[ "${TEST_CONFIG}" == *operator_microbenchmark* ]]; then
   test_operator_microbenchmark
+elif [[ "${TEST_CONFIG}" == *attention_microbenchmark* ]]; then
+  test_attention_microbenchmark
 elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
   test_inductor_distributed
 elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
@@ -1786,6 +1821,8 @@ elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   test_inductor_shard "${SHARD_NUMBER}"
 elif [[ "${TEST_CONFIG}" == *einops* ]]; then
   test_einops
+elif [[ "${TEST_CONFIG}" == *dynamo_core* ]]; then
+  test_dynamo_core
 elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then
   install_torchvision
   test_dynamo_wrapped_shard "${SHARD_NUMBER}"

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-ad5816f0eee1c873df1b7d371c69f1f811a89387
+07b6cbde121417a70e4dc871adb6d27030e0ce3f

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-ccb801b88af136454798b945175c4c87e636ac33
+acccf86477759b2d3500f1ae1be065f7b1e409ec

.github/pytorch-probot.yml

@@ -7,6 +7,7 @@ ciflow_push_tags:
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
+- ciflow/dynamo
 - ciflow/h100
 - ciflow/h100-cutlass-backend
 - ciflow/h100-distributed

.github/scripts/generate_pytorch_version.py

@@ -50,7 +50,7 @@ def get_tag() -> str:
 
 def get_base_version() -> str:
     root = get_pytorch_root()
-    dirty_version = open(root / "version.txt").read().strip()
+    dirty_version = Path(root / "version.txt").read_text().strip()
     # Strips trailing a0 from version.txt, not too sure why it's there in the
     # first place
     return re.sub(LEGACY_BASE_VERSION_SUFFIX_PATTERN, "", dirty_version)

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

@@ -260,11 +260,8 @@ jobs:
             "${DOCKER_IMAGE}"
           )
           docker exec -t -w "${PYTORCH_ROOT}" "${container_name}" bash -c "bash .circleci/scripts/binary_populate_env.sh"
-          if [[ ${BUILD_ENVIRONMENT} == *"aarch64"* ]]; then
-            docker exec -t "${container_name}" bash -c "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
+          # 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"
 
       - name: Chown artifacts
         if: ${{ steps.filter.outputs.is-test-matrix-empty == 'False' && inputs.build_environment != 'linux-s390x-binary-manywheel' }}

.github/workflows/_linux-test.yml

@@ -326,7 +326,7 @@ jobs:
           SCCACHE_BUCKET: ${{ !contains(matrix.runner, 'b200') && 'ossci-compiler-cache-circleci-v2' || '' }}
           SCCACHE_REGION: ${{ !contains(matrix.runner, 'b200') && 'us-east-1' || '' }}
           SHM_SIZE: ${{ contains(inputs.build-environment, 'cuda') && '2g' || '1g' }}
-          DOCKER_IMAGE: ${{ inputs.docker-image }}
+          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
           XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
           XLA_CLANG_CACHE_S3_BUCKET_NAME: ossci-compiler-clang-cache-circleci-xla
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}

.github/workflows/attention_op_microbenchmark.yml

@@ -0,0 +1,73 @@
+name: attention_op_microbenchmark
+
+on:
+  push:
+    tags:
+      - ciflow/op-benchmark/*
+  workflow_dispatch:
+  schedule:
+    # Run at 06:00 UTC everyday
+    - cron: 0 7 * * *
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
+  cancel-in-progress: true
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+  attn-microbenchmark-build:
+    if: github.repository_owner == 'pytorch'
+    uses: ./.github/workflows/_linux-build.yml
+    with:
+      runner: linux.12xlarge.memory
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
+      cuda-arch-list: '8.0 9.0'
+      test-matrix: |
+        { include: [
+          { config: "attention_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.aws.a100" },
+          { config: "attention_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.aws.h100" },
+        ]}
+    secrets: inherit
+
+  attn-microbenchmark-test:
+    name: attn-microbenchmark-test
+    uses: ./.github/workflows/_linux-test.yml
+    needs: attn-microbenchmark-build
+    with:
+      timeout-minutes: 500
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      docker-image: ${{ needs.attn-microbenchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.attn-microbenchmark-build.outputs.test-matrix }}
+    secrets: inherit
+
+  # B200 runner
+  opmicrobenchmark-build-b200:
+    if: github.repository_owner == 'pytorch'
+    name: opmicrobenchmark-build-b200
+    uses: ./.github/workflows/_linux-build.yml
+    with:
+      runner: linux.12xlarge.memory
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
+      cuda-arch-list: '10.0'
+      test-matrix: |
+        { include: [
+          { config: "operator_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
+        ]}
+    secrets: inherit
+
+  opmicrobenchmark-test-b200:
+    name: opmicrobenchmark-test-b200
+    uses: ./.github/workflows/_linux-test.yml
+    needs: opmicrobenchmark-build-b200
+    with:
+      timeout-minutes: 500
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      docker-image: ${{ needs.opmicrobenchmark-build-b200.outputs.docker-image }}
+      test-matrix: ${{ needs.opmicrobenchmark-build-b200.outputs.test-matrix }}
+      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+    secrets: inherit

.github/workflows/b200-distributed.yml

@@ -37,6 +37,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.12xlarge.memory
       build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '10.0'

.github/workflows/b200-symm-mem.yml

@@ -37,6 +37,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.12xlarge.memory
       build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '10.0'

.github/workflows/docker-builds.yml

@@ -119,6 +119,22 @@ jobs:
         with:
           docker-image: ${{ steps.build-docker-image.outputs.docker-image }}
 
+      - name: Generate output
+        if: contains(matrix.docker-image-name, 'rocm')
+        id: generate_output
+        run: |
+          docker_image_name="${{ matrix.docker-image-name }}"
+          docker_image_tag="${{ steps.build-docker-image.outputs.docker-image }}"
+          echo "${docker_image_name}=${docker_image_tag}" >> docker-builds-output-${docker_image_name}.txt
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4.4.0
+        if: contains(matrix.docker-image-name, 'rocm')
+        with:
+          name: docker-builds-artifacts-${{ matrix.docker-image-name }}
+          retention-days: 14
+          path: ./docker-builds-output-${{ matrix.docker-image-name }}.txt
+
       - uses: nick-fields/retry@7152eba30c6575329ac0576536151aca5a72780e # v3.0.0
         name: Push to https://ghcr.io/
         id: push-to-ghcr-io

.github/workflows/docker-cache-rocm.yml

@@ -3,8 +3,7 @@ name: docker-cache-rocm
 on:
   workflow_run:
     workflows: [docker-builds]
-    # TODO: Uncomment before merging
-    #branches: [main, release]
+    branches: [main, release]
     types:
       - completed
   workflow_dispatch:
@@ -98,9 +97,7 @@ jobs:
           elif [[ $ref_name == "main" ]]; then
             ref_suffix="main"
           else
-            # TODO: Remove below
-            ref_suffix="main"
-            # echo "Unexpected branch in ref_name: ${ref_name}" && exit 1
+            echo "Unexpected branch in ref_name: ${ref_name}" && exit 1
           fi
           docker tag ${{ steps.ghcr-io-tag.outputs.ghcr_image }} ${{ matrix.docker-image }}
           # mv is atomic operation, so we use intermediate tar.tmp file to prevent read-write contention

.github/workflows/dynamo-unittest.yml

@@ -0,0 +1,70 @@
+# Workflow: Dynamo Unit Test
+# runs unit tests for dynamo.
+name: dynamo-unittest
+
+on:
+  push:
+    tags:
+      - ciflow/dynamo/*
+  workflow_call:
+  schedule:
+    - cron: 29 8 * * * # about 1:29am PDT
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+  get-label-type:
+    name: get-label-type
+    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
+    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
+    with:
+      triggering_actor: ${{ github.triggering_actor }}
+      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
+      curr_branch: ${{ github.head_ref || github.ref_name }}
+      curr_ref_type: ${{ github.ref_type }}
+      opt_out_experiments: lf
+
+  dynamo-build:
+    name: dynamo-build
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    strategy:
+      matrix:
+        python-version: ['3.11', '3.12']
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-py${{ matrix.python-version }}-clang12
+      docker-image-name: ci-image:pytorch-linux-jammy-py${{ matrix.python-version }}-clang12
+      test-matrix: |
+        { include: [
+          { config: "dynamo_core", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "dynamo_wrapped", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "dynamo_wrapped", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "dynamo_wrapped", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+        ]}
+    secrets: inherit
+
+  dynamo-test:
+    name: dynamo-test
+    uses: ./.github/workflows/_linux-test.yml
+    needs: [get-label-type, dynamo-build]
+    strategy:
+      matrix:
+        python-version: ['3.11', '3.12']
+    with:
+      build-environment: linux-jammy-py${{ matrix.python-version }}-clang12
+      docker-image: ci-image:pytorch-linux-jammy-py${{ matrix.python-version }}-clang12
+      test-matrix: |
+        { include: [
+          { config: "dynamo_core", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "dynamo_wrapped", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "dynamo_wrapped", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "dynamo_wrapped", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+        ]}
+    secrets: inherit

.github/workflows/test-b200.yml

@@ -5,7 +5,9 @@
 # Flow:
 # 1. Builds PyTorch with CUDA 12.8+ and sm100 architecture for B200
 # 2. Runs smoke tests on linux.dgx.b200 runner
-# 3. Tests executed are defined in .ci/pytorch/test.sh -> test_python_smoke() function
+# 3. Tests executed are defined in .ci/pytorch/test.sh -> test_python_smoke_b200() function
+#    - Includes matmul, scaled_matmul, FP8, and FlashAttention CuTe tests
+#    - FlashAttention CuTe DSL is installed as part of test execution
 #
 # Triggered by:
 # - Pull requests modifying this workflow file
@@ -52,6 +54,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.12xlarge.memory
       build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '10.0'
@@ -72,4 +75,4 @@ jobs:
       docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build.outputs.test-matrix }}
       aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-    secrets: inherit
+    secrets: inherit

.github/workflows/trunk-rocm-mi300.yml

@@ -0,0 +1,83 @@
+name: trunk-rocm-mi300
+
+on:
+  push:
+    branches:
+      - main
+      - release/*
+  workflow_dispatch:
+  schedule:
+    - cron: 29 8 * * *  # about 1:29am PDT
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
+  cancel-in-progress: true
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+  llm-td:
+    if: github.repository_owner == 'pytorch'
+    name: before-test
+    uses: ./.github/workflows/llm_td_retrieval.yml
+    permissions:
+      id-token: write
+      contents: read
+
+  target-determination:
+    name: before-test
+    uses: ./.github/workflows/target_determination.yml
+    needs: llm-td
+    permissions:
+      id-token: write
+      contents: read
+
+  get-label-type:
+    name: get-label-type
+    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
+    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
+    with:
+      triggering_actor: ${{ github.triggering_actor }}
+      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
+      curr_branch: ${{ github.head_ref || github.ref_name }}
+      curr_ref_type: ${{ github.ref_type }}
+
+  linux-jammy-rocm-py3_10-build:
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
+      sync-tag: rocm-build
+      test-matrix: |
+        { include: [
+          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
+          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
+          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
+          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
+          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
+          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4.b" },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4.b" },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4.b" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+    secrets: inherit

.github/workflows/upload-test-stats.yml

@@ -5,6 +5,7 @@ on:
     workflows:
       - pull
       - trunk
+      - trunk-rocm-mi300
       - periodic
       - periodic-rocm-mi200
       - periodic-rocm-mi300

.spin/cmds.py

@@ -0,0 +1,330 @@
+import hashlib
+import subprocess
+import sys
+from pathlib import Path
+
+import click
+import spin
+
+
+def file_digest(file, algorithm: str):
+    try:
+        return hashlib.file_digest(file, algorithm)
+    except AttributeError:
+        pass  # Fallback to manual implementation below
+    hash = hashlib.new(algorithm)
+    while chunk := file.read(8192):
+        hash.update(chunk)
+    return hash
+
+
+def _hash_file(file):
+    with open(file, "rb") as f:
+        hash = file_digest(f, "sha256")
+    return hash.hexdigest()
+
+
+def _hash_files(files):
+    hashes = {file: _hash_file(file) for file in files}
+    return hashes
+
+
+def _read_hashes(hash_file: Path):
+    if not hash_file.exists():
+        return {}
+    with hash_file.open("r") as f:
+        lines = f.readlines()
+    hashes = {}
+    for line in lines:
+        hash = line[:64]
+        file = line[66:].strip()
+        hashes[file] = hash
+    return hashes
+
+
+def _updated_hashes(hash_file, files_to_hash):
+    old_hashes = _read_hashes(hash_file)
+    new_hashes = _hash_files(files_to_hash)
+    if new_hashes != old_hashes:
+        return new_hashes
+    return None
+
+
+@click.command()
+def regenerate_version():
+    """Regenerate version.py."""
+    cmd = [
+        sys.executable,
+        "-m",
+        "tools.generate_torch_version",
+        "--is-debug=false",
+    ]
+    spin.util.run(cmd)
+
+
+TYPE_STUBS = [
+    (
+        "Pytorch type stubs",
+        Path(".lintbin/.pytorch-type-stubs.sha256"),
+        [
+            "aten/src/ATen/native/native_functions.yaml",
+            "aten/src/ATen/native/tags.yaml",
+            "tools/autograd/deprecated.yaml",
+        ],
+        [
+            sys.executable,
+            "-m",
+            "tools.pyi.gen_pyi",
+            "--native-functions-path",
+            "aten/src/ATen/native/native_functions.yaml",
+            "--tags-path",
+            "aten/src/ATen/native/tags.yaml",
+            "--deprecated-functions-path",
+            "tools/autograd/deprecated.yaml",
+        ],
+    ),
+    (
+        "Datapipes type stubs",
+        None,
+        [],
+        [
+            sys.executable,
+            "torch/utils/data/datapipes/gen_pyi.py",
+        ],
+    ),
+]
+
+
+@click.command()
+def regenerate_type_stubs():
+    """Regenerate type stubs."""
+    for name, hash_file, files_to_hash, cmd in TYPE_STUBS:
+        if hash_file:
+            if hashes := _updated_hashes(hash_file, files_to_hash):
+                click.echo(
+                    f"Changes detected in type stub files for {name}. Regenerating..."
+                )
+                spin.util.run(cmd)
+                hash_file.parent.mkdir(parents=True, exist_ok=True)
+                with hash_file.open("w") as f:
+                    for file, hash in hashes.items():
+                        f.write(f"{hash}  {file}\n")
+                click.echo("Type stubs and hashes updated.")
+            else:
+                click.echo(f"No changes detected in type stub files for {name}.")
+        else:
+            click.echo(f"No hash file for {name}. Regenerating...")
+            spin.util.run(cmd)
+            click.echo("Type stubs regenerated.")
+
+
+@click.command()
+def regenerate_clangtidy_files():
+    """Regenerate clang-tidy files."""
+    cmd = [
+        sys.executable,
+        "-m",
+        "tools.linter.clang_tidy.generate_build_files",
+    ]
+    spin.util.run(cmd)
+
+
+#: These linters are expected to need less than 3s cpu time total
+VERY_FAST_LINTERS = {
+    "ATEN_CPU_GPU_AGNOSTIC",
+    "BAZEL_LINTER",
+    "C10_NODISCARD",
+    "C10_UNUSED",
+    "CALL_ONCE",
+    "CMAKE_MINIMUM_REQUIRED",
+    "CONTEXT_DECORATOR",
+    "COPYRIGHT",
+    "CUBINCLUDE",
+    "DEPLOY_DETECTION",
+    "ERROR_PRONE_ISINSTANCE",
+    "EXEC",
+    "HEADER_ONLY_LINTER",
+    "IMPORT_LINTER",
+    "INCLUDE",
+    "LINTRUNNER_VERSION",
+    "MERGE_CONFLICTLESS_CSV",
+    "META_NO_CREATE_UNBACKED",
+    "NEWLINE",
+    "NOQA",
+    "NO_WORKFLOWS_ON_FORK",
+    "ONCE_FLAG",
+    "PYBIND11_INCLUDE",
+    "PYBIND11_SPECIALIZATION",
+    "PYPIDEP",
+    "PYPROJECT",
+    "RAWCUDA",
+    "RAWCUDADEVICE",
+    "ROOT_LOGGING",
+    "TABS",
+    "TESTOWNERS",
+    "TYPEIGNORE",
+    "TYPENOSKIP",
+    "WORKFLOWSYNC",
+}
+
+
+#: These linters are expected to take a few seconds, but less than 10s cpu time total
+FAST_LINTERS = {
+    "CMAKE",
+    "DOCSTRING_LINTER",
+    "GHA",
+    "NATIVEFUNCTIONS",
+    "RUFF",
+    "SET_LINTER",
+    "SHELLCHECK",
+    "SPACES",
+}
+
+
+#: These linters are expected to take more than 10s cpu time total;
+#: some need more than 1 hour.
+SLOW_LINTERS = {
+    "ACTIONLINT",
+    "CLANGFORMAT",
+    "CLANGTIDY",
+    "CODESPELL",
+    "FLAKE8",
+    "GB_REGISTRY",
+    "PYFMT",
+    "PYREFLY",
+    "TEST_DEVICE_BIAS",
+    "TEST_HAS_MAIN",
+}
+
+
+ALL_LINTERS = VERY_FAST_LINTERS | FAST_LINTERS | SLOW_LINTERS
+
+
+LINTRUNNER_CACHE_INFO = (
+    Path(".lintbin/.lintrunner.sha256"),
+    [
+        "requirements.txt",
+        "pyproject.toml",
+        ".lintrunner.toml",
+    ],
+)
+
+
+LINTRUNNER_BASE_CMD = [
+    "uvx",
+    "--python",
+    "3.10",
+    "lintrunner@0.12.7",
+]
+
+
+@click.command()
+def setup_lint():
+    """Set up lintrunner with current CI version."""
+    cmd = LINTRUNNER_BASE_CMD + ["init"]
+    subprocess.run(cmd, check=True, capture_output=True, text=True)
+
+
+def _check_linters():
+    cmd = LINTRUNNER_BASE_CMD + ["list"]
+    ret = spin.util.run(cmd, output=False, stderr=subprocess.PIPE)
+    linters = {l.strip() for l in ret.stdout.decode().strip().split("\n")[1:]}
+    unknown_linters = linters - ALL_LINTERS
+    missing_linters = ALL_LINTERS - linters
+    if unknown_linters:
+        click.secho(
+            f"Unknown linters found; please add them to the correct category "
+            f"in .spin/cmds.py: {', '.join(unknown_linters)}",
+            fg="yellow",
+        )
+    if missing_linters:
+        click.secho(
+            f"Missing linters found; please update the corresponding category "
+            f"in .spin/cmds.py: {', '.join(missing_linters)}",
+            fg="yellow",
+        )
+    return unknown_linters, missing_linters
+
+
+@spin.util.extend_command(
+    setup_lint,
+    doc=f"""
+        If configuration has changed, update lintrunner.
+
+        Compares the stored old hashes of configuration files with new ones and
+        performs setup via setup-lint if the hashes have changed.
+        Hashes are stored in {LINTRUNNER_CACHE_INFO[0]}; the following files are
+        considered: {", ".join(LINTRUNNER_CACHE_INFO[1])}.
+        """,
+)
+@click.pass_context
+def lazy_setup_lint(ctx, parent_callback, **kwargs):
+    if hashes := _updated_hashes(*LINTRUNNER_CACHE_INFO):
+        click.echo(
+            "Changes detected in lint configuration files. Setting up linting tools..."
+        )
+        parent_callback(**kwargs)
+        hash_file = LINTRUNNER_CACHE_INFO[0]
+        hash_file.parent.mkdir(parents=True, exist_ok=True)
+        with hash_file.open("w") as f:
+            for file, hash in hashes.items():
+                f.write(f"{hash}  {file}\n")
+        click.echo("Linting tools set up and hashes updated.")
+    else:
+        click.echo("No changes detected in lint configuration files. Skipping setup.")
+    click.echo("Regenerating version...")
+    ctx.invoke(regenerate_version)
+    click.echo("Regenerating type stubs...")
+    ctx.invoke(regenerate_type_stubs)
+    click.echo("Done.")
+    _check_linters()
+
+
+@click.command()
+@click.option("-a", "--apply-patches", is_flag=True)
+@click.pass_context
+def lint(ctx, apply_patches, **kwargs):
+    """Lint all files."""
+    ctx.invoke(lazy_setup_lint)
+    all_files_linters = VERY_FAST_LINTERS | FAST_LINTERS
+    changed_files_linters = SLOW_LINTERS
+    cmd = LINTRUNNER_BASE_CMD
+    if apply_patches:
+        cmd += ["--apply-patches"]
+    all_files_cmd = cmd + [
+        "--take",
+        ",".join(all_files_linters),
+        "--all-files",
+    ]
+    spin.util.run(all_files_cmd)
+    changed_files_cmd = cmd + [
+        "--take",
+        ",".join(changed_files_linters),
+    ]
+    spin.util.run(changed_files_cmd)
+
+
+@click.command()
+@click.pass_context
+def fixlint(ctx, **kwargs):
+    """Autofix all files."""
+    ctx.invoke(lint, apply_patches=True)
+
+
+@click.command()
+@click.option("-a", "--apply-patches", is_flag=True)
+@click.pass_context
+def quicklint(ctx, apply_patches, **kwargs):
+    """Lint changed files."""
+    ctx.invoke(lazy_setup_lint)
+    cmd = LINTRUNNER_BASE_CMD
+    if apply_patches:
+        cmd += ["--apply-patches"]
+    spin.util.run(cmd)
+
+
+@click.command()
+@click.pass_context
+def quickfix(ctx, **kwargs):
+    """Autofix changed files."""
+    ctx.invoke(quicklint, apply_patches=True)

aten/src/ATen/core/ivalue.h

@@ -18,6 +18,8 @@
 #include <unordered_set>
 #include <utility>
 
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
+
 namespace torch {
 class TORCH_API CustomClassHolder : public c10::intrusive_ptr_target {};
 namespace jit {
@@ -1630,4 +1632,6 @@ struct TORCH_API WeakOrStrongTypePtr {
 
 } // namespace c10
 
+C10_DIAGNOSTIC_POP()
+
 #include <ATen/core/ivalue_inl.h> // IWYU pragma: keep

aten/src/ATen/core/ivalue_inl.h

@@ -29,6 +29,8 @@
 #include <c10/util/intrusive_ptr.h>
 #include <c10/util/irange.h>
 
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
+
 namespace torch {
 namespace jit {
 struct Function;
@@ -2567,3 +2569,5 @@ TypePtr IValue::type() const {
 }
 
 } // namespace c10
+
+C10_DIAGNOSTIC_POP()

aten/src/ATen/cpu/vec/vec128/vec128_convert.h

@@ -223,6 +223,62 @@ CONVERT_FROM_BF16_TEMPLATE(double)
 CONVERT_FROM_BF16_TEMPLATE(float16_t)
 #endif
 
+#ifdef __ARM_FEATURE_BF16
+
+// clang-[17, 20] crashes when autovectorizing static cast to bf16
+// Below is a workaround to have some vectorization
+// Works decently well for smaller int types
+template <typename from_type>
+inline void convertToBf16Impl(
+    const from_type* __restrict src,
+    c10::BFloat16* __restrict dst,
+    uint64_t n) {
+  bfloat16_t* dstPtr = reinterpret_cast<bfloat16_t*>(dst);
+  uint64_t loopBound = n - (n % 16);
+  uint64_t i = 0;
+  for (; i < loopBound; i += 16) {
+    float32x4_t a, b, c, d;
+    a[0] = static_cast<float>(src[i]);
+    a[1] = static_cast<float>(src[i + 1]);
+    a[2] = static_cast<float>(src[i + 2]);
+    a[3] = static_cast<float>(src[i + 3]);
+    b[0] = static_cast<float>(src[i + 4]);
+    b[1] = static_cast<float>(src[i + 5]);
+    b[2] = static_cast<float>(src[i + 6]);
+    b[3] = static_cast<float>(src[i + 7]);
+    c[0] = static_cast<float>(src[i + 8]);
+    c[1] = static_cast<float>(src[i + 9]);
+    c[2] = static_cast<float>(src[i + 10]);
+    c[3] = static_cast<float>(src[i + 11]);
+    d[0] = static_cast<float>(src[i + 12]);
+    d[1] = static_cast<float>(src[i + 13]);
+    d[2] = static_cast<float>(src[i + 14]);
+    d[3] = static_cast<float>(src[i + 15]);
+
+    vst1q_bf16(dstPtr + i, vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(a), b));
+    vst1q_bf16(dstPtr + i + 8, vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(c), d));
+  }
+
+#pragma clang loop vectorize(disable) interleave(disable) unroll(disable)
+  for (; i < n; i++) {
+    float a = static_cast<float>(src[i]);
+    dstPtr[i] = vcvth_bf16_f32(a);
+  }
+}
+
+#define CONVERT_TO_BF16_TEMPLATE(from_type)                                  \
+  template <>                                                                \
+  inline void convert(const from_type* src, c10::BFloat16* dst, int64_t n) { \
+    return convertToBf16Impl<from_type>(src, dst, n);                        \
+  }
+
+CONVERT_TO_BF16_TEMPLATE(uint8_t)
+CONVERT_TO_BF16_TEMPLATE(int8_t)
+CONVERT_TO_BF16_TEMPLATE(int16_t)
+CONVERT_TO_BF16_TEMPLATE(int32_t)
+
+#endif
+
 inline void convertBoolToBfloat16Impl(
     const bool* __restrict src,
     c10::BFloat16* __restrict dst,

aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h

@@ -11,6 +11,8 @@
 #include <sleef.h>
 #endif
 
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
+
 // Sleef offers vectorized versions of some transcedentals
 // such as sin, cos, tan etc..
 // However for now opting for STL, since we are not building
@@ -650,3 +652,5 @@ inline Vectorized<float> Vectorized<float>::erf() const {
 
 } // namespace CPU_CAPABILITY
 } // namespace at::vec
+
+C10_DIAGNOSTIC_POP()

aten/src/ATen/cuda/CUDAGraph.cpp

@@ -1,6 +1,7 @@
 #include <ATen/cuda/CUDAGeneratorImpl.h>
 #include <ATen/cuda/CUDAGraph.h>
 #include <ATen/cuda/Exceptions.h>
+#include <ATen/cuda/MemPool.h>
 #include <ATen/Functions.h>
 #include <c10/cuda/CUDAFunctions.h>
 
@@ -13,7 +14,7 @@ static bool _cuda_graphs_debug = false;
 MempoolId_t graph_pool_handle() {
   // Sets just the second value, to distinguish it from MempoolId_ts created from
   // cudaStreamGetCaptureInfo id_s in capture_begin.
-  return c10::cuda::MemPool::graph_pool_handle();
+  return at::cuda::MemPool::graph_pool_handle();
 }
 
 /**
@@ -90,7 +91,7 @@ void CUDAGraph::capture_begin(MempoolId_t pool/*=0*/, cudaStreamCaptureMode capt
   } else {
     // User did not ask us to share a mempool. Create graph pool handle using is_user_created=false.
     // Sets just the first value, to distinguish it from MempoolId_ts created by graph_pool_handle().
-    mempool_id_ = c10::cuda::MemPool::graph_pool_handle(false);
+    mempool_id_ = at::cuda::MemPool::graph_pool_handle(false);
     TORCH_INTERNAL_ASSERT(mempool_id_.first > 0);
   }
 
@@ -174,17 +175,24 @@ void CUDAGraph::instantiate() {
     // Trailing NULL, NULL, 0 arguments were recommended by Cuda driver people,
     // who prefer not to report error message through these arguments moving forward
     // (they prefer return value, or errors on api calls internal to the capture)
-#if (defined(CUDA_VERSION) && CUDA_VERSION >= 12000)
-    AT_CUDA_CHECK(cudaGraphInstantiate(&graph_exec_, graph_, 0));
+    // ROCM appears to fail with HIP error: invalid argument
+#if (defined(CUDA_VERSION) && CUDA_VERSION >= 12000) && !defined(USE_ROCM)
+    AT_CUDA_CHECK(cudaGraphInstantiate(&graph_exec_, graph_, cudaGraphInstantiateFlagUseNodePriority));
 #else
     AT_CUDA_CHECK(cudaGraphInstantiate(&graph_exec_, graph_, NULL, NULL, 0));
 #endif
 //Since ROCm 6.2, we want to go down this path as hipGraphExecDestroy in the destructor will not immediately free the memory.
 //It will wait for the next sync operation. cudaGraphInstantiateFlagAutoFreeOnLaunch will add async frees after graph launch.
   } else {
+#if !defined(USE_ROCM)
+    AT_CUDA_CHECK(cudaGraphInstantiateWithFlags(&graph_exec_,
+                                                graph_,
+                                                cudaGraphInstantiateFlagAutoFreeOnLaunch | cudaGraphInstantiateFlagUseNodePriority));
+#else
     AT_CUDA_CHECK(cudaGraphInstantiateWithFlags(&graph_exec_,
                                                 graph_,
                                                 cudaGraphInstantiateFlagAutoFreeOnLaunch));
+#endif
   }
   has_graph_exec_ = true;
 }

aten/src/ATen/cuda/MemPool.cpp

@@ -0,0 +1,69 @@
+#include <ATen/core/CachingHostAllocator.h>
+#include <ATen/cuda/MemPool.h>
+
+namespace at::cuda {
+
+// uid_ is incremented when a user creates a MemPool,
+// for example: using graph_pool_handle() or c10::cuda::MemPool().
+//
+// uuid_ is incremented when CUDAGraph creates a MemPool
+// as a result of a user not providing a pool.
+//
+// MempoolId_t of {0, 0} is used to denote when no MemPool has been
+// passed to a function, either by user or CUDAGraphs. For example,
+// default value of MempoolId_t for capture_begin function is {0, 0}.
+// That's why uid_ and uuid_ start at 1.
+std::atomic<CaptureId_t> MemPool::uid_{1};
+std::atomic<CaptureId_t> MemPool::uuid_{1};
+
+MemPool::MemPool(
+    CUDACachingAllocator::CUDAAllocator* allocator,
+    bool is_user_created,
+    bool use_on_oom)
+    : allocator_(allocator), is_user_created_(is_user_created) {
+  if (is_user_created_) {
+    id_ = {0, uid_++};
+  } else {
+    id_ = {uuid_++, 0};
+  }
+  device_ = c10::cuda::current_device();
+  CUDACachingAllocator::createOrIncrefPool(device_, id_, allocator);
+  if (use_on_oom) {
+    CUDACachingAllocator::setUseOnOOM(device_, id_);
+  }
+}
+
+MemPool::~MemPool() {
+  // TORCH_INTERNAL_ASSERT(use_count() == 1);
+  // We used to assert that TORCH_INTERNAL_ASSERT(use_count() == 1);
+  // However, this assertion is not true if a memory pool is shared
+  // with a cuda graph. That CUDAGraph will increase the use count
+  // until it is reset.
+  CUDACachingAllocator::releasePool(device_, id_);
+  c10::cuda::CUDACachingAllocator::emptyCache(id_);
+}
+
+MempoolId_t MemPool::id() {
+  return id_;
+}
+
+CUDACachingAllocator::CUDAAllocator* MemPool::allocator() {
+  return allocator_;
+}
+
+int MemPool::use_count() {
+  return CUDACachingAllocator::getPoolUseCount(device_, id_);
+}
+
+c10::DeviceIndex MemPool::device() {
+  return device_;
+}
+
+MempoolId_t MemPool::graph_pool_handle(bool is_user_created) {
+  if (is_user_created) {
+    return {0, uid_++};
+  }
+  return {uuid_++, 0};
+}
+
+} // namespace at::cuda

aten/src/ATen/cuda/MemPool.h

@@ -0,0 +1,44 @@
+#pragma once
+
+#include <c10/core/Allocator.h>
+#include <c10/cuda/CUDACachingAllocator.h>
+
+namespace at::cuda {
+
+// Keep BC only
+using c10::CaptureId_t;
+using c10::MempoolId_t;
+
+// MemPool represents a pool of memory in a caching allocator. Currently,
+// it's just the ID of the pool object maintained in the CUDACachingAllocator.
+//
+// An allocator pointer can be passed to the MemPool to define how the
+// allocations should be done in the pool. For example: using a different
+// system allocator such as ncclMemAlloc.
+struct TORCH_CUDA_CPP_API MemPool {
+  MemPool(
+      c10::cuda::CUDACachingAllocator::CUDAAllocator* allocator = nullptr,
+      bool is_user_created = true,
+      bool use_on_oom = false);
+  MemPool(const MemPool&) = delete;
+  MemPool(MemPool&&) = default;
+  MemPool& operator=(const MemPool&) = delete;
+  MemPool& operator=(MemPool&&) = default;
+  ~MemPool();
+
+  MempoolId_t id();
+  c10::cuda::CUDACachingAllocator::CUDAAllocator* allocator();
+  int use_count();
+  c10::DeviceIndex device();
+  static MempoolId_t graph_pool_handle(bool is_user_created = true);
+
+ private:
+  static std::atomic<CaptureId_t> uid_;
+  static std::atomic<CaptureId_t> uuid_;
+  c10::cuda::CUDACachingAllocator::CUDAAllocator* allocator_;
+  bool is_user_created_;
+  MempoolId_t id_;
+  c10::DeviceIndex device_;
+};
+
+} // namespace at::cuda

aten/src/ATen/native/LinearAlgebra.cpp

@@ -1936,7 +1936,7 @@ static bool should_fold(const Tensor& tensor1, const Tensor& tensor2, bool has_o
 
   // We order the tensors. t1 will be the larger tensor
   // We can always transpose tensor2 as the dimensions are always >= 1 (precondition from matmul)
-  // and tensor1_larger iff tensor2.dim() > tensor1.dim(9
+  // and tensor1_larger iff tensor2.dim() > tensor1.dim()
   const auto t1 = tensor1_larger ? MaybeOwned<Tensor>::borrowed(tensor1)
                                  : MaybeOwned<Tensor>::owned(tensor2.mT());
   const int64_t dim_t1 = t1->dim();
@@ -1948,20 +1948,11 @@ static bool should_fold(const Tensor& tensor1, const Tensor& tensor2, bool has_o
     return false;
   }
 
-  // In this case we *do* incur in an extra copy to avoid creating an unnecessary large tensor in the backward
-  // Suppose we don't fold here. Let t1.shape = [b, m, n] t2.shape = [n, k] like in a transformer
-  // t2 will be expanded to a tensor of shape [b, n, k] and then we do t1.bmm(t2_expanded)
-  // The issue appears in the backward.
-  // The output gradient g of this operation would have shape [b, m, k]
-  // The backward wrt. t2 of bmm would be given by t1.mH @ g, which has shape [b, n, k]
-  // Then, the backward of expand is simply `sum(0)`. As such, we are instantiating a tensor
-  // of shape [b, n, k] unnecessarily, which may cause a large memory footprint, and in the
-  // worst case, an OOM
-  bool t2_requires_grad = tensor1_larger ? tensor2.requires_grad() : tensor1.requires_grad();
-  if (t2_requires_grad && !has_out) {
-    // We should be checking !at::GradMode::is_enabled(), but apparently
-    // this regresses performance in some cases:
-    // https://github.com/pytorch/pytorch/issues/118548#issuecomment-1916022394
+  // If we require a gradient, we should fold to minimize backward memory usage - even if this
+  // leads to a copy in forward because is needed in backward,
+  // only time we avoid this strict pre-allocated memory usage (has_out = True)
+  bool requires_grad = tensor1.requires_grad() || tensor2.requires_grad();
+  if (requires_grad && !has_out) {
     return true;
   }
 

aten/src/ATen/native/TransposeType.h

@@ -1,6 +1,8 @@
 #pragma once
 #include <c10/util/Exception.h>
 
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
+
 namespace at::native {
 
 // Used as an interface between the different BLAS-like libraries
@@ -21,3 +23,5 @@ static inline char to_blas(TransposeType trans) {
 }
 
 }  // namespace at::native
+
+C10_DIAGNOSTIC_POP()

aten/src/ATen/native/UnaryOps.cpp

@@ -904,19 +904,11 @@ Tensor mvlgamma(const Tensor& self, int64_t p) {
   return args.lgamma_().sum(-1).add_(p2_sub_p * std::log(c10::pi<double>) * QUARTER);
 }
 
+// since mvlgamma_ has different signature from its
+// out and functional variant, we explicitly
+// define it (instead of using structured kernel).
 Tensor& mvlgamma_(Tensor& self, int64_t p) {
-  mvlgamma_check(self, p);
-  Tensor args = native::arange(
-      -p *HALF  + HALF,
-      HALF,
-      HALF,
-      optTypeMetaToScalarType(self.options().dtype_opt()),
-      self.options().layout_opt(),
-      self.options().device_opt(),
-      self.options().pinned_memory_opt());
-  args = args.add(self.unsqueeze(-1));
-  const auto p2_sub_p = static_cast<double>(p * (p - 1));
-  return self.copy_(args.lgamma_().sum(-1).add_(p2_sub_p * std::log(c10::pi<double>) * QUARTER));
+  return at::mvlgamma_out(self, self, p);
 }
 
 Tensor& mvlgamma_out(const Tensor& self, int64_t p, Tensor& result) {

aten/src/ATen/native/cuda/CompositeRandomAccessor.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include <ATen/native/CompositeRandomAccessorCommon.h>
+#include <thrust/swap.h>
 #include <thrust/tuple.h>
 
 namespace at { namespace native {

aten/src/ATen/native/cuda/DilatedMaxPool2d.cu

@@ -75,30 +75,52 @@ static inline bool can_use_int32_nhwc(
   return true;
 }
 
+static inline bool can_use_int32_nchw(
+    int64_t nbatch, int64_t channels,
+    int64_t height, int64_t width,
+    int64_t pooled_height, int64_t pooled_width) {
+  int64_t hw = height * width;
+  return can_use_int32_nhwc(
+      nbatch, channels, height, width,
+      pooled_height, pooled_width,
+      channels * hw,  // in_stride_n
+      hw, // in_stride_c
+      width, // in_stride_h
+      1 // in_stride_w
+  );
+}
+
 // kernels borrowed from Caffe
-template <typename scalar_t>
-__global__ void max_pool_forward_nchw(const int nthreads, const scalar_t* bottom_data,
-    const int64_t channels, const int64_t height,
-    const int64_t width, const int pooled_height, const int pooled_width,
-    const int kernel_h, const int kernel_w, const int stride_h,
-    const int stride_w, const int pad_h, const int pad_w,
-    const int dilation_h, const int dilation_w, scalar_t* top_data,
+template <typename scalar_t, typename index_t>
+__global__ void max_pool_forward_nchw(
+    const index_t nthreads,
+    const scalar_t* bottom_data,
+    const int64_t channels,
+    const int64_t height,
+    const int64_t width,
+    const int pooled_height,
+    const int pooled_width,
+    const int kernel_h, const int kernel_w,
+    const int stride_h, const int stride_w,
+    const int pad_h, const int pad_w,
+    const int dilation_h, const int dilation_w,
+    scalar_t* top_data,
     int64_t* top_mask) {
-  CUDA_KERNEL_LOOP(index, nthreads) {
-    int pw = index % pooled_width;
-    int ph = (index / pooled_width) % pooled_height;
-    int c = (index / pooled_width / pooled_height) % channels;
-    int n = index / pooled_width / pooled_height / channels;
-    int hstart = ph * stride_h - pad_h;
-    int wstart = pw * stride_w - pad_w;
-    int hend = min(hstart + (kernel_h - 1) * dilation_h + 1, height);
-    int wend = min(wstart + (kernel_w - 1) * dilation_w + 1, width);
+  CUDA_KERNEL_LOOP_TYPE(index, nthreads, index_t) {
+    index_t pw = index % pooled_width;
+    index_t ph = (index / pooled_width) % pooled_height;
+    index_t c = (index / pooled_width / pooled_height) % channels;
+    index_t n = index / pooled_width / pooled_height / channels;
+    index_t hstart = ph * stride_h - pad_h;
+    index_t wstart = pw * stride_w - pad_w;
+    index_t hend = min(hstart + (kernel_h - 1) * dilation_h + 1, height);
+    index_t wend = min(wstart + (kernel_w - 1) * dilation_w + 1, width);
     while(hstart < 0)
       hstart += dilation_h;
     while(wstart < 0)
       wstart += dilation_w;
     scalar_t maxval = at::numeric_limits<scalar_t>::lower_bound(); // -Infinity
-    int maxidx = hstart * width + wstart;
+    index_t maxidx = hstart * width + wstart;
     const scalar_t* btm_data = bottom_data + (n * channels + c) * height * width;
     for (int h = hstart; h < hend; h += dilation_h) {
       for (int w = wstart; w < wend; w += dilation_w) {
@@ -251,32 +273,39 @@ __global__ void max_pool_forward_nhwc(
 
 static constexpr int BLOCK_THREADS = 256;
 
-template <typename scalar_t, typename accscalar_t>
+template <typename scalar_t, typename accscalar_t, typename index_t>
 #if defined (USE_ROCM)
 C10_LAUNCH_BOUNDS_2(BLOCK_THREADS, 4)
 #else
 C10_LAUNCH_BOUNDS_2(BLOCK_THREADS, 8)
 #endif
-__global__ void max_pool_backward_nchw(const scalar_t* top_diff,
-    const int64_t* top_mask, const int num, const int64_t channels,
-    const int64_t height, const int64_t width, const int pooled_height,
-    const int pooled_width, const int kernel_h, const int kernel_w,
-    const int stride_h, const int stride_w, const int pad_h, const int pad_w,
+__global__ void max_pool_backward_nchw(
+    const scalar_t* top_diff,
+    const int64_t* top_mask,
+    const index_t num,
+    const index_t channels,
+    const index_t height,
+    const index_t width,
+    const index_t pooled_height,
+    const index_t pooled_width,
+    const int kernel_h, const int kernel_w,
+    const int stride_h, const int stride_w,
+    const int pad_h, const int pad_w,
     const int dilation_h, const int dilation_w,
     scalar_t* bottom_diff) {
-  CUDA_KERNEL_LOOP(index, height*width) {
-    int h = index / width;
-    int w = index - h * width;
-    int phstart = p_start(h, pad_h, kernel_h, dilation_h, stride_h);
-    int phend = p_end(h, pad_h, pooled_height, stride_h);
-    int pwstart = p_start(w, pad_w, kernel_w, dilation_w, stride_w);
-    int pwend = p_end(w, pad_w, pooled_width, stride_w);
-    for (int n = blockIdx.y; n < num; n += gridDim.y) {
-      for (int c = blockIdx.z; c < channels; c+= gridDim.z) {
+  CUDA_KERNEL_LOOP_TYPE(index, height*width, index_t) {
+    index_t h = index / width;
+    index_t w = index - h * width;
+    index_t phstart = p_start(h, pad_h, kernel_h, dilation_h, stride_h);
+    index_t phend = p_end(h, pad_h, pooled_height, stride_h);
+    index_t pwstart = p_start(w, pad_w, kernel_w, dilation_w, stride_w);
+    index_t pwend = p_end(w, pad_w, pooled_width, stride_w);
+    for (index_t n = blockIdx.y; n < num; n += gridDim.y) {
+      for (index_t c = blockIdx.z; c < channels; c += gridDim.z) {
         accscalar_t gradient = accscalar_t(0);
-        int offset = (n * channels + c) * pooled_height * pooled_width;
-        for (int ph = phstart; ph < phend; ++ph) {
-          for (int pw = pwstart; pw < pwend; ++pw) {
+        index_t offset = (n * channels + c) * pooled_height * pooled_width;
+        for (index_t ph = phstart; ph < phend; ++ph) {
+          for (index_t pw = pwstart; pw < pwend; ++pw) {
             if (top_mask[ph * pooled_width + pw + offset] == h * width + w) {
               gradient += static_cast<accscalar_t>(top_diff[ph * pooled_width + pw + offset]);
             }
@@ -469,8 +498,6 @@ const Tensor& indices) {
   const int64_t in_stride_h = input.stride(-2);
   const int64_t in_stride_w = input.stride(-1);
 
-  const int count = safe_downcast<int, int64_t>(output.numel());
-
   AT_DISPATCH_FLOATING_TYPES_AND2(kHalf, kBFloat16, input.scalar_type(),
     "max_pool2d_with_indices_out_cuda_frame",
     [&] {
@@ -553,14 +580,42 @@ const Tensor& indices) {
           break;
         }
         case MemoryFormat::Contiguous: {
-          const int num_threads = std::min(at::cuda::getCurrentDeviceProperties()->maxThreadsPerBlock,
-                                            BLOCK_THREADS);
-          max_pool_forward_nchw<scalar_t>
-              <<<ceil_div(count, num_threads), num_threads, 0, at::cuda::getCurrentCUDAStream()>>>(
-              count, input_data,
-                  nInputPlane, inputHeight, inputWidth, outputHeight, outputWidth,
-                  kH, kW, dH, dW, padH, padW, dilationH, dilationW,
-                  output_data, indices_data);
+          const int threads = std::min(
+              at::cuda::getCurrentDeviceProperties()->maxThreadsPerBlock,
+              BLOCK_THREADS);
+          const int64_t nthreads = output.numel();
+          bool use_int32 = can_use_int32_nchw(
+              nbatch, nInputPlane, inputHeight, inputWidth, outputHeight, outputWidth);
+          const int maxGridX = at::cuda::getCurrentDeviceProperties()->maxGridSize[0];
+          const int blocks = static_cast<int>(std::min<int64_t>(
+              ceil_div(nthreads, static_cast<int64_t>(threads)),
+              static_cast<int64_t>(maxGridX)));
+          auto stream = at::cuda::getCurrentCUDAStream();
+          if (use_int32) {
+            max_pool_forward_nchw<scalar_t, int32_t>
+                <<<blocks, threads, 0, stream>>>(
+                    static_cast<int32_t>(nthreads),
+                    input_data,
+                    static_cast<int32_t>(nInputPlane),
+                    static_cast<int32_t>(inputHeight),
+                    static_cast<int32_t>(inputWidth),
+                    static_cast<int32_t>(outputHeight),
+                    static_cast<int32_t>(outputWidth),
+                    kH, kW, dH, dW, padH, padW, dilationH, dilationW,
+                    output_data, indices_data);
+          } else {
+            max_pool_forward_nchw<scalar_t, int64_t>
+                <<<blocks, threads, 0, stream>>>(
+                    nthreads,
+                    input_data,
+                    nInputPlane,
+                    inputHeight,
+                    inputWidth,
+                    outputHeight,
+                    outputWidth,
+                    kH, kW, dH, dW, padH, padW, dilationH, dilationW,
+                    output_data, indices_data);
+          }
           C10_CUDA_KERNEL_LAUNCH_CHECK();
           break;
         }
@@ -633,8 +688,6 @@ const Tensor& gradInput) {
 
   gradInput.zero_();
 
-  int64_t count = input.numel();
-
   AT_DISPATCH_FLOATING_TYPES_AND2(kHalf, kBFloat16, input.scalar_type(),
     "max_pool2d_with_indices_out_cuda_frame",
     [&] {
@@ -692,25 +745,45 @@ const Tensor& gradInput) {
           break;
         }
         case MemoryFormat::Contiguous: {
-          int imgcount = inputWidth * inputHeight;
-          dim3 grid;
-          const int blocks = (imgcount + BLOCK_THREADS - 1) / BLOCK_THREADS;
-          grid.x = blocks;
-          grid.y = nbatch;
-          uint64_t maxGridY = at::cuda::getCurrentDeviceProperties()->maxGridSize[1];
-          if (maxGridY < grid.y) grid.y = maxGridY;
-          grid.z = nInputPlane;
-          uint64_t maxGridZ = at::cuda::getCurrentDeviceProperties()->maxGridSize[2];
-          if (maxGridZ < grid.z) grid.z = maxGridZ;
-
-          max_pool_backward_nchw<scalar_t, accscalar_t>
-          <<<grid, BLOCK_THREADS, 0, at::cuda::getCurrentCUDAStream()>>>(
-                  gradOutput_data,
-                  indices_data,
-                  nbatch,
-                  nInputPlane, inputHeight, inputWidth, outputHeight, outputWidth,
-                  kH, kW, dH, dW, padH, padW, dilationH, dilationW,
-                  gradInput_data);
+          const int threads = std::min(
+              at::cuda::getCurrentDeviceProperties()->maxThreadsPerBlock,
+              BLOCK_THREADS);
+          const int imgcount = inputWidth * inputHeight;
+          const int maxGridX = at::cuda::getCurrentDeviceProperties()->maxGridSize[0];
+          const int maxGridY = at::cuda::getCurrentDeviceProperties()->maxGridSize[1];
+          const int maxGridZ = at::cuda::getCurrentDeviceProperties()->maxGridSize[2];
+          const int blocks_x = std::min(ceil_div(imgcount, threads), maxGridX);
+          dim3 grid(blocks_x, static_cast<unsigned>(std::min<int64_t>(nbatch, maxGridY)), static_cast<unsigned>(std::min<int64_t>(nInputPlane, maxGridZ)));
+          bool use_int32 = can_use_int32_nchw(
+              nbatch, nInputPlane, inputHeight, inputWidth, outputHeight, outputWidth);
+          auto stream = at::cuda::getCurrentCUDAStream();
+          if (use_int32) {
+            max_pool_backward_nchw<scalar_t, accscalar_t, int32_t>
+                <<<grid, threads, 0, stream>>>(
+                    gradOutput_data,
+                    indices_data,
+                    static_cast<int32_t>(nbatch),
+                    static_cast<int32_t>(nInputPlane),
+                    static_cast<int32_t>(inputHeight),
+                    static_cast<int32_t>(inputWidth),
+                    static_cast<int32_t>(outputHeight),
+                    static_cast<int32_t>(outputWidth),
+                    kH, kW, dH, dW, padH, padW, dilationH, dilationW,
+                    gradInput_data);
+          } else {
+            max_pool_backward_nchw<scalar_t, accscalar_t, int64_t>
+                <<<grid, threads, 0, stream>>>(
+                    gradOutput_data,
+                    indices_data,
+                    nbatch,
+                    nInputPlane,
+                    inputHeight,
+                    inputWidth,
+                    outputHeight,
+                    outputWidth,
+                    kH, kW, dH, dW, padH, padW, dilationH, dilationW,
+                    gradInput_data);
+          }
           C10_CUDA_KERNEL_LAUNCH_CHECK();
           break;
         }

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

@@ -78,9 +78,18 @@ __global__ void EmbeddingBag_updateOutputKernel_max(
       scalar_t weightFeatMax = 0;
       int64_t bag_size_ = 0;
       int64_t maxWord = -1;
+
+      // Separate validation loop reduces register pressure in the main loop below.
+      // No early exit (break) on invalid input as benchmarking shows it degrades performance.
+      bool has_invalid_index = false;
+      for (int64_t emb = begin; emb < end; emb++) {
+        index_t input_idx = input[emb];
+        has_invalid_index = has_invalid_index || (input_idx < 0 || input_idx >= numRows);
+      }
+      CUDA_KERNEL_ASSERT(!has_invalid_index && "Invalid input index in EmbeddingBag: index out of range [0, numRows)");
+
       for (int64_t emb = begin; emb < end; emb++) {
         bool pad = (input[emb] == padding_idx);
-        CUDA_KERNEL_ASSERT(input[emb] < numRows);
         const int64_t weightRow = input[emb] * weight_stride0;
         scalar_t weightValue = weightFeat[weightRow];
         if (bag_size_ == 0 || weightValue > weightFeatMax) {
@@ -129,10 +138,19 @@ __global__ void EmbeddingBag_updateOutputKernel_sum_mean(
       CUDA_KERNEL_ASSERT(end >= begin);
       accscalar_t weightFeatSum = 0;
       int64_t bag_size_ = 0;
+
+      // Separate validation loop reduces register pressure in the main loop below.
+      // No early exit (break) on invalid input as benchmarking shows it degrades performance.
+      bool has_invalid_index = false;
+      for (int64_t emb = begin; emb < end; emb++) {
+        index_t input_idx = input[emb];
+        has_invalid_index = has_invalid_index || (input_idx < 0 || input_idx >= numRows);
+      }
+      CUDA_KERNEL_ASSERT(!has_invalid_index && "Invalid input index in EmbeddingBag: index out of range [0, numRows)");
+
       for (int64_t emb = begin; emb < end; emb++) {
         index_t input_idx = input[emb];
         bool pad = (input_idx == padding_idx);
-        CUDA_KERNEL_ASSERT(0 <= input_idx && input_idx < numRows);
         const int64_t weightRow = input_idx * weight_stride0;
         scalar_t weightValue = weightFeat[weightRow];
         weightValue = pad ? static_cast<scalar_t>(0) : weightValue;

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

@@ -78,9 +78,9 @@ _mx8_mx8_bf16_grouped_mm_fbgemm(
         const Tensor& mat_a,
         const Tensor& mat_b,
         const Tensor& scale_a,
-        const SwizzleType& swizzle_a,
+        const SwizzleType swizzle_a,
         const Tensor& scale_b,
-        const SwizzleType& swizzle_b,
+        const SwizzleType swizzle_b,
         const std::optional<at::Tensor>& offs,
         Tensor& out) {
     const bool a_is_2d = mat_a.dim() == 2;

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

@@ -740,7 +740,12 @@ _scaled_rowwise_rowwise(
   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));
+  // if we have a scale of shape [256, 1] (say), then stride can be [1, 0] - handle this case
+  TORCH_CHECK_VALUE(
+      scale_a.stride(1) == 1 ||
+      scale_a.size(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;

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

@@ -267,15 +267,15 @@ void scan_dim_with_indices(const TensorBase& self, const TensorBase& values, con
  * outer dimensions, which contains several "inner rows").
  * Each thread processes a single inner row at a time.
  */
-template<typename scalar_t, class BinaryOp>
+template<typename scalar_t, typename index_t, class BinaryOp>
 __global__ void tensor_kernel_scan_outer_dim(scalar_t *tgt_, const scalar_t *src_,
                                               const uint32_t num_orows, const uint32_t num_irows, const uint32_t row_size,
                                               const scalar_t init, BinaryOp binary_op)
 {
   for (uint32_t orow = blockIdx.x; orow < num_orows; orow += gridDim.x) {
     for (uint32_t irow = blockIdx.y * blockDim.x + threadIdx.x; irow < num_irows; irow += gridDim.y * blockDim.x) {
-      const scalar_t *src = src_ + orow * row_size * num_irows + irow;
-      scalar_t *tgt = tgt_ + orow * row_size * num_irows + irow;
+      const scalar_t *src = src_ + static_cast<index_t>(orow) * row_size * num_irows + irow;
+      scalar_t *tgt = tgt_ + (index_t) orow * row_size * num_irows + irow;
       scalar_t acc = init;
 
       for (uint32_t col = 0; col < row_size; ++col) {
@@ -409,10 +409,15 @@ __host__ void scan_outer_dim(const TensorBase& self, const TensorBase& result,
   check_fits_in_unsigned(num_irows, "num_irows");
   check_fits_in_unsigned(num_orows, "num_orows");
   check_fits_in_unsigned(row_size, "row_size");
-
-  tensor_kernel_scan_outer_dim<scalar_t><<<grid, threads, 0, at::cuda::getCurrentCUDAStream()>>>(
+  if (static_cast<size_t>(num_irows) * num_orows * row_size <= UINT_MAX) {
+  tensor_kernel_scan_outer_dim<scalar_t, uint32_t><<<grid, threads, 0, at::cuda::getCurrentCUDAStream()>>>(
     result.mutable_data_ptr<scalar_t>(), self.const_data_ptr<scalar_t>(),
     num_orows, num_irows, row_size, init, binary_op);
+  } else  {
+  tensor_kernel_scan_outer_dim<scalar_t, size_t><<<grid, threads, 0, at::cuda::getCurrentCUDAStream()>>>(
+    result.mutable_data_ptr<scalar_t>(), self.const_data_ptr<scalar_t>(),
+    num_orows, num_irows, row_size, init, binary_op);
+  }
   C10_CUDA_KERNEL_LAUNCH_CHECK();
 }
 

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

@@ -337,10 +337,6 @@ Tensor _convolution_out(
   TORCH_CHECK(
       3 == ndim || 4 == ndim || 5 == ndim,
       "convolution only supports 3D, 4D, 5D tensor");
-  // get computation format for Conv/TransposedConv
-  bool is_channels_last_suggested =
-      use_channels_last_for_conv(input_r, weight_r);
-
   Tensor input = input_r, weight = weight_r;
   // PyTorch does not support ChannelsLast1D case,
   // thus we need the transformation here
@@ -348,13 +344,8 @@ Tensor _convolution_out(
     input = view4d(input_r);
     weight = view4d(weight_r);
   }
-  // ensure the input/weight/bias/output are congituous in desired format
-  at::MemoryFormat mfmt = is_channels_last_suggested
-      ? get_cl_tag_by_ndim(input.ndimension())
-      : at::MemoryFormat::Contiguous;
-  auto bias = bias_r.defined() ? bias_r.contiguous() : bias_r;
-  input = input.contiguous(mfmt);
-  weight = weight.contiguous(mfmt);
+  // get computation format for Conv/TransposedConv
+  bool is_channels_last_suggested = use_channels_last_for_conv(input, weight);
 
   auto k = weight.ndimension();
   if (k == input.ndimension() + 1) {
@@ -388,6 +379,14 @@ Tensor _convolution_out(
         expand_param_if_needed(output_padding_, "output_padding", dim);
     params.groups = groups_;
   }
+
+  // ensure the input/weight/bias/output are congituous in desired format
+  at::MemoryFormat mfmt = is_channels_last_suggested
+      ? get_cl_tag_by_ndim(input.ndimension())
+      : at::MemoryFormat::Contiguous;
+  auto bias = bias_r.defined() ? bias_r.contiguous() : bias_r;
+  input = input.contiguous(mfmt);
+  weight = weight.contiguous(mfmt);
   check_shape_forward(input, weight, bias, params, true);
 
   Tensor output;
@@ -514,18 +513,9 @@ Tensor convolution_overrideable(
       at::borrow_from_optional_tensor(bias_r_opt);
   const Tensor& bias_r = *bias_r_maybe_owned;
 
-  auto k = weight_r.ndimension();
-  at::MemoryFormat backend_memory_format = at::MemoryFormat::Contiguous;
-  if (xpu_conv_use_channels_last(input_r, weight_r)) {
-    backend_memory_format = (k == 5) ? at::MemoryFormat::ChannelsLast3d
-                                     : at::MemoryFormat::ChannelsLast;
-  }
-  Tensor input_c = input_r.contiguous(backend_memory_format);
-  Tensor weight_c = weight_r.contiguous(backend_memory_format);
-
   return _convolution(
-      input_c,
-      weight_c,
+      input_r,
+      weight_r,
       bias_r,
       stride_,
       padding_,

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

@@ -0,0 +1,342 @@
+#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
+#include <ATen/BlasBackend.h>
+#include <ATen/WrapDimUtilsMulti.h>
+#include <ATen/ceil_div.h>
+#include <ATen/native/Resize.h>
+#include <ATen/native/mkldnn/xpu/detail/oneDNN.h>
+#include <ATen/native/xpu/Blas.h>
+#include <torch/library.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
+
+namespace at::native {
+
+using at::blas::ScalingType;
+using at::blas::SwizzleType;
+
+namespace {
+/*
+ * Scaling Type Determination:
+ * ---------------------------
+ * Conditions and corresponding Scaling Types:
+ *
+ * - If scale tensor is `Float8_e8m0fnu` or `Float8_e4m3fn`:
+ *   - Returns BlockWise (with additional size checks).
+ *
+ * - Else if scale.numel() == 1:
+ *   - Returns TensorWise.
+ *
+ * - Else if scale.dim() == 2 && scale.size(0) == outer_dim && scale.size(1) ==
+ * 1:
+ *   - Returns RowWise.
+ *
+ * - Otherwise:
+ *   - Returns Error.
+ */
+
+bool is_tensorwise_scaling(const at::Tensor& t, const at::Tensor& scale) {
+  return at::isFloat8Type(t.scalar_type()) &&
+      scale.scalar_type() == at::kFloat && scale.numel() == 1;
+}
+
+bool is_rowwise_scaling(const at::Tensor& t, const at::Tensor& scale) {
+  return (
+      at::isFloat8Type(t.scalar_type()) && scale.scalar_type() == at::kFloat &&
+      scale.dim() == 2 && scale.size(0) == t.size(0) && scale.size(1) == 1 &&
+      scale.is_contiguous());
+}
+
+bool is_desired_scaling(
+    const at::Tensor& t,
+    const at::Tensor& scale,
+    ScalingType desired_scaling) {
+  auto result = desired_scaling == ScalingType::TensorWise
+      ? is_tensorwise_scaling(t, scale)
+      : is_rowwise_scaling(t, scale);
+  return result;
+}
+
+std::pair<ScalingType, ScalingType> get_joint_scaling(
+    std::initializer_list<std::pair<ScalingType, ScalingType>> options,
+    const at::Tensor& a,
+    const at::Tensor& b,
+    const at::Tensor& scale_a,
+    const at::Tensor& scale_b) {
+  for (auto [lhs, rhs] : options) {
+    if (is_desired_scaling(a, scale_a, lhs) &&
+        is_desired_scaling(b.t(), scale_b.t(), rhs)) {
+      return {lhs, rhs};
+    }
+  }
+  TORCH_CHECK(
+      false,
+      "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 (",
+      a.size(0),
+      ", 1) and scale_b should be (1, ",
+      b.size(1),
+      "), and both should be contiguous.\n"
+      "Got a.dtype()=",
+      a.scalar_type(),
+      ", scale_a.dtype()=",
+      scale_a.scalar_type(),
+      ", scale_a.size()=",
+      scale_a.sizes(),
+      ", scale_a.stride()=",
+      scale_a.strides(),
+      ", ",
+      "b.dtype()=",
+      b.scalar_type(),
+      ", scale_b.dtype()=",
+      scale_b.scalar_type(),
+      ", scale_b.size()=",
+      scale_b.sizes(),
+      " and scale_b.stride()=",
+      scale_b.strides());
+}
+
+Tensor& _scaled_gemm(
+    const Tensor& mat1,
+    const Tensor& mat2,
+    const Tensor& scale_a,
+    const Tensor& scale_b,
+    const ScalingType scaling_choice_a,
+    const ScalingType scaling_choice_b,
+    const std::optional<Tensor>& bias,
+    const bool use_fast_accum,
+    Tensor& out,
+    const std::optional<Tensor>& alpha = std::nullopt) {
+  // TODO: scale_result and alpha is not defined or used!
+  std::optional<Tensor> scaled_result = std::nullopt;
+  at::native::onednn::scaled_matmul(
+      mat1,
+      mat2,
+      out,
+      scale_a,
+      scale_b,
+      scaling_choice_a,
+      scaling_choice_b,
+      bias,
+      scaled_result,
+      use_fast_accum);
+
+  return out;
+}
+
+} // namespace
+
+// 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:
+//    - `mat1`: the first operand of the matrix multiply, can be type
+//    `torch.float8_e4m3fn` or `torch.float8_e5m2`
+//    - `mat2`: the second operand of the matrix multiply, can be type
+//    `torch.float8_e4m3fn` or `torch.float8_e5m2`
+//    - `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
+//    - `scale_a`: a tensor with the inverse scale of `mat1`, whose
+//    shape/strides/dtype depend on the scaling scheme
+//    - `scale_b`: a tensor with the inverse scale of `mat2`, whose
+//    shape/strides/dtype depend on the scaling scheme
+//    - `scale_result`: a scalar tensor with the scale of the output, only
+//    utilized if the output is a float8 type
+//    - `use_fast_accum`: Not applicable for XPU. For now, it should always be
+//    false.
+//    - `out`: a reference to the output tensor
+
+Tensor& _scaled_mm_out_xpu(
+    const Tensor& mat1,
+    const Tensor& mat2,
+    const Tensor& scale_a,
+    const Tensor& scale_b,
+    const std::optional<at::Tensor>& bias,
+    const std::optional<at::Tensor>& scale_result,
+    std::optional<c10::ScalarType> out_dtype,
+    bool use_fast_accum,
+    Tensor& out) {
+  // Note: fast_accum is not supported in XPU for now.
+  TORCH_CHECK(!use_fast_accum, "fast_accum is not supported in XPU for now.");
+
+  TORCH_CHECK(mat1.dim() == 2, "mat1 must be a matrix");
+  TORCH_CHECK(mat2.dim() == 2, "mat2 must be a matrix");
+
+  TORCH_CHECK(
+      mat1.sizes()[1] == mat2.sizes()[0],
+      "mat1 and mat2 shapes cannot be multiplied (",
+      mat1.sizes()[0],
+      "x",
+      mat1.sizes()[1],
+      " and ",
+      mat2.sizes()[0],
+      "x",
+      mat2.sizes()[1],
+      ")");
+
+  // Check what type of scaling we are doing based on inputs. This list is
+  // sorted by decreasing priority.
+
+  // List of supported datatypes for XPU with oneDNN:
+  // https://uxlfoundation.github.io/oneDNN/dev_guide_matmul.html#data-types
+  auto [scaling_choice_a, scaling_choice_b] = get_joint_scaling(
+      {
+          std::make_pair(ScalingType::TensorWise, ScalingType::TensorWise),
+          std::make_pair(ScalingType::RowWise, ScalingType::RowWise),
+      },
+      mat1,
+      mat2,
+      scale_a,
+      scale_b);
+  TORCH_CHECK(
+      !scale_result ||
+          (scale_result->numel() == 1 && scale_result->scalar_type() == kFloat),
+      "scale_result must be a float scalar");
+  TORCH_CHECK(
+      !bias || bias->numel() == mat2.sizes()[1],
+      "Bias must be size ",
+      mat2.sizes()[1],
+      " but got ",
+      bias->numel());
+  TORCH_CHECK(
+      mat1.sizes()[1] % 16 == 0,
+      "Expected trailing dimension of mat1 to be divisible by 16 ",
+      "but got mat1 shape: (",
+      mat1.sizes()[0],
+      "x",
+      mat1.sizes()[1],
+      ").");
+  TORCH_CHECK(
+      mat2.sizes()[0] % 16 == 0 && mat2.sizes()[1] % 16 == 0,
+      "mat2 shape (",
+      mat2.sizes()[0],
+      "x",
+      mat2.sizes()[1],
+      ") must be divisible by 16");
+  // Check types
+  TORCH_CHECK(
+      !out_dtype || *out_dtype == out.scalar_type(),
+      "out_dtype must match output matrix type");
+  TORCH_CHECK(
+      at::isFloat8Type(mat1.scalar_type()),
+      "Expected mat1 to be Float8 matrix got ",
+      mat1.scalar_type());
+  TORCH_CHECK(
+      at::isFloat8Type(mat2.scalar_type()),
+      "Expected mat2 to be Float8 matrix got ",
+      mat2.scalar_type());
+  // TODO: oneDNN Currently only supports e4m3 with group scales on BMG. Not
+  // support 2D scales, only 1D. Needs to add more checks there.
+
+  if (bias) {
+    TORCH_CHECK(
+        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());
+    auto scale_result_ = scale_result.value_or(Tensor());
+
+    // NOLINTNEXTLINE(*c-array*)
+    TensorArg targs[]{
+        {out, "out", 0},
+        {mat1, "mat1", 1},
+        {mat2, "mat2", 2},
+        {bias_, "bias", 3},
+        {scale_a, "scale_a", 4},
+        {scale_b, "scale_b", 5},
+        {scale_result_, "scale_result", 6}};
+    checkAllSameGPU(__func__, targs);
+  }
+
+  // Validation checks have passed lets resize the output to actual size
+  IntArrayRef mat1_sizes = mat1.sizes();
+  IntArrayRef mat2_sizes = mat2.sizes();
+  at::native::resize_output(out, {mat1_sizes[0], mat2_sizes[1]});
+
+  // If any of M, K, N is 0 - return early (the tensorwise/rowwise float8 gemm
+  // kernels do not support this case).
+  if (mat1_sizes[0] == 0 || mat1_sizes[1] == 0 || mat2_sizes[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.
+    if (mat1_sizes[1] == 0) {
+      out.zero_();
+    }
+
+    return out;
+  }
+
+  // TODO: Scale_result is not supported by now!!
+  return _scaled_gemm(
+      mat1,
+      mat2,
+      scale_a,
+      scale_b,
+      scaling_choice_a,
+      scaling_choice_b,
+      bias,
+      use_fast_accum,
+      out);
+}
+
+Tensor _scaled_mm_xpu(
+    const Tensor& mat_a,
+    const Tensor& mat_b,
+    const Tensor& scale_a,
+    const Tensor& scale_b,
+    const std::optional<at::Tensor>& bias,
+    const std::optional<at::Tensor>& scale_result,
+    std::optional<c10::ScalarType> out_dtype,
+    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_out_xpu(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      bias,
+      scale_result,
+      out_dtype,
+      use_fast_accum,
+      out);
+}
+
+} // namespace at::native

aten/src/ATen/native/mkldnn/xpu/detail/QMatmul.cpp

@@ -1,3 +1,4 @@
+#include <ATen/BlasBackend.h>
 #include <ATen/Tensor.h>
 #include <ATen/core/Tensor.h>
 #include <c10/core/ScalarType.h>
@@ -8,7 +9,6 @@
 #include <oneapi/dnnl/dnnl.hpp>
 
 namespace at::native::onednn {
-
 at::Tensor broadcast_bias2D(
     at::Tensor& dst,
     at::Tensor& bias,
@@ -328,4 +328,236 @@ void quantized_matmul(
     result.copy_(dst);
 }
 
+// Describes how to configure oneDNN scales for a given role/ScalingType
+struct ScaleSpec {
+  // specifies the way scale values will be applied to an ARG tensor.
+  int mask;
+  // specifies how scales are grouped along dimensions where
+  // multiple scale factors are used.
+  dnnl::memory::dims groups;
+  // specifies data type for scale factors.
+  dnnl::memory::data_type dtype;
+
+  // Helper to compute expected number of elements for scale tensors
+  // arg_type: "src" for SRC (groups pattern {1, X}),
+  // "wei" for WEIGHTS (groups pattern {X, 1})
+  int64_t expected_numel(
+      int64_t outer_dim,
+      int64_t inner_dim,
+      const std::string& arg_type) const {
+    if (groups == dnnl::memory::dims{1, 1})
+      return 1; // tensorwise scaling
+
+    TORCH_CHECK(
+        arg_type == "src" || arg_type == "wei",
+        "Expected arg_type to be 'src' or 'wei', but got '",
+        arg_type,
+        "'");
+
+    // For rowwise: SRC groups={1, K}, WEI groups={K, 1}
+    TORCH_INTERNAL_ASSERT(
+        (groups == dnnl::memory::dims{1, inner_dim} ||
+         groups == dnnl::memory::dims{inner_dim, 1}),
+        "The groups must be either {1, inner_dim} or {inner_dim, 1}. But got ",
+        groups,
+        ".");
+    return outer_dim;
+  }
+
+  // Normalize an incoming scale tensor to contiguous storage and appropriate
+  // dtype/view
+  at::Tensor normalize(const at::Tensor& scale) const {
+    TORCH_INTERNAL_ASSERT(
+        dtype == dnnl::memory::data_type::f32,
+        "tensor scale currently must be f32, but got scale dtype: ",
+        scale.scalar_type());
+    return scale.to(at::kFloat).contiguous();
+  }
+};
+
+// This function defines how to set scales mask and groups according to:
+// https://github.com/uxlfoundation/oneDNN/blob/main/tests/benchdnn/doc/knobs_attr.md#--attr-scales
+// The returned value will be used in
+// `set_scales(arg, mask, groups, data_type)`.
+inline ScaleSpec make_scale_spec(
+    at::blas::ScalingType scaling_type,
+    int64_t M,
+    int64_t K,
+    int64_t N,
+    const std::string& arg_type) {
+  TORCH_CHECK(
+      arg_type == "src" || arg_type == "wei",
+      "Expected arg_type to be 'src' or 'wei', but got '",
+      arg_type,
+      "'");
+  TORCH_INTERNAL_ASSERT(
+      (scaling_type == at::blas::ScalingType::TensorWise ||
+       scaling_type == at::blas::ScalingType::RowWise),
+      "Currently only support scaling_type for TensorWise or RowWise");
+  int64_t dim = K; // Currently only K is used for grouping
+  bool is_src = (arg_type == "src");
+  if (scaling_type == at::blas::ScalingType::TensorWise) {
+    // Scale tensorwise. The same as `--attr-scales=common`.
+    // mask=0 : scale whole tensor
+    // groups={1, 1}: indicates that there is only one group for scaling
+    return {0, {1, 1}, dnnl::memory::data_type::f32};
+  } else {
+    // (scaling_type == at::blas::ScalingType::RowWise)
+    // Scale RowWise. The same as `--attr-scales=per_dim_01`.
+    // mask={(1 << 0) | (1 << 1)}: Scale on both dim0 and dim1
+    // SRC: groups={1, K}, WEIGHTS: groups={K, 1}
+    return {
+        (1 << 0) | (1 << 1),
+        is_src ? dnnl::memory::dims{1, dim} : dnnl::memory::dims{dim, 1},
+        dnnl::memory::data_type::f32};
+  }
+}
+
+sycl::event scaled_matmul(
+    const Tensor& mat1,
+    const Tensor& mat2,
+    Tensor& result,
+    const Tensor& scale_a,
+    const Tensor& scale_b,
+    at::blas::ScalingType scaling_choice_a,
+    at::blas::ScalingType scaling_choice_b,
+    const std::optional<at::Tensor>& bias,
+    const std::optional<at::Tensor>& scale_result,
+    bool use_fast_accum) {
+  auto& engine = GpuEngineManager::Instance().get_engine();
+  auto& stream = GpuStreamManager::Instance().get_stream();
+
+  // This function will do steps with following steps
+  // 1. create memory descriptor
+  // 2. call write_to_dnnl_memory() to actually write memory
+  // 3. execute
+
+  const int64_t M = mat1.size(0);
+  const int64_t K = mat1.size(1);
+  const int64_t N = mat2.size(1);
+
+  // 1.1 Create memory descriptor
+  dnnl::memory::desc src_md = get_onednn_md(mat1);
+  dnnl::memory::desc weights_md = get_onednn_md(mat2);
+  dnnl::memory::desc dst_md = get_onednn_md(result);
+
+  // scale_a and scale_b has already be checked in `is_desired_scaling()` call.
+  // So we could directly get their memory desc and set later.
+  dnnl::memory::desc scale_a_md = get_onednn_md(scale_a);
+  dnnl::memory::desc scale_b_md = get_onednn_md(scale_b);
+
+  dnnl::memory::desc bias_md;
+  bool with_bias = bias.has_value();
+  at::Tensor possible_reshaped_bias = bias.value_or(at::Tensor());
+  if (with_bias) {
+    if (possible_reshaped_bias.dim() == 1) {
+      possible_reshaped_bias =
+          possible_reshaped_bias.reshape({1, possible_reshaped_bias.size(0)});
+      bias_md = get_onednn_md(possible_reshaped_bias);
+    } else {
+      bias_md = get_onednn_md(possible_reshaped_bias);
+    }
+  }
+
+  // 1.2 Create primitive descriptor and set scales mask
+  const ScaleSpec src_spec = make_scale_spec(scaling_choice_a, M, K, N, "src");
+  const ScaleSpec wei_spec = make_scale_spec(scaling_choice_b, M, K, N, "wei");
+
+  dnnl::primitive_attr op_attr = dnnl::primitive_attr();
+
+#if ONEDNN_SUPPORT_DETERMINISTIC
+  if (at::globalContext().deterministicAlgorithms() ||
+      at::globalContext().deterministicMkldnn())
+    op_attr.set_deterministic(true);
+#endif
+
+  std::vector<int64_t> default_groups;
+  op_attr.set_scales(
+      DNNL_ARG_SRC, src_spec.mask, src_spec.groups, src_spec.dtype);
+  op_attr.set_scales(
+      DNNL_ARG_WEIGHTS, wei_spec.mask, wei_spec.groups, wei_spec.dtype);
+  // scale_result tensor currently only supports scalar(TensorWise Scaling).
+  bool with_dst_scale = scale_result && scale_result->defined();
+  if (with_dst_scale) {
+    op_attr.set_scales(DNNL_ARG_DST, 0, {1}, dnnl::memory::data_type::f32);
+  }
+
+  op_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+
+  // 1.3 Create the matmul primitive descriptor
+  dnnl::matmul::primitive_desc matmul_pd = with_bias
+      ? dnnl::matmul::primitive_desc(
+            engine, src_md, weights_md, bias_md, dst_md, op_attr)
+      : dnnl::matmul::primitive_desc(
+            engine, src_md, weights_md, dst_md, op_attr);
+
+  // 1.4 (Possible) Additional Checks
+  // TODO: In case there are memory desc does not align with the actual tensor,
+  // we might need to reorder weights similar to CPU's reorder_if_differ_in()
+  // call. For example, weights not the same as matmul_pd.weights_desc(),
+
+  // 2. Prepare memory
+
+  // Create memory
+  auto src_usr_m = make_onednn_memory(src_md, engine, mat1.data_ptr());
+  auto weights_usr_m = make_onednn_memory(weights_md, engine, mat2.data_ptr());
+  auto dst_usr_m = make_onednn_memory(dst_md, engine, result.data_ptr());
+  dnnl::memory b_usr_m;
+  if (with_bias) {
+    b_usr_m =
+        make_onednn_memory(bias_md, engine, possible_reshaped_bias.data_ptr());
+  }
+
+  // Prepare runtime scale memories (flat 1-D views) using the specs
+  auto make_scale_mem_from_spec = [&](const ScaleSpec& spec,
+                                      int64_t expected_numel,
+                                      const at::Tensor& scale_tensor) {
+    at::Tensor prepared = spec.normalize(scale_tensor);
+    TORCH_CHECK(
+        prepared.numel() == expected_numel,
+        "Scale buffer length mismatch. Expected ",
+        expected_numel,
+        ", got ",
+        prepared.numel());
+    dnnl::memory::desc scale_md(
+        {prepared.numel()}, spec.dtype, dnnl::memory::format_tag::x);
+    return make_onednn_memory(scale_md, engine, prepared.data_ptr());
+  };
+
+  auto scratchpad =
+      make_onednn_memory(matmul_pd.scratchpad_desc(), engine, nullptr);
+
+  // 3. Setup Args for exec
+  std::unordered_map<int, dnnl::memory> args;
+  args.insert({DNNL_ARG_SRC, src_usr_m});
+  args.insert({DNNL_ARG_WEIGHTS, weights_usr_m});
+  args.insert({DNNL_ARG_DST, dst_usr_m});
+  args.insert({DNNL_ARG_SCRATCHPAD, scratchpad});
+  if (with_bias) {
+    args.insert({DNNL_ARG_BIAS, b_usr_m});
+  }
+
+  // Attach runtime scales using specs
+  auto src_sc_mem = make_scale_mem_from_spec(
+      src_spec, src_spec.expected_numel(M, K, "src"), scale_a);
+  auto wei_sc_mem = make_scale_mem_from_spec(
+      wei_spec, wei_spec.expected_numel(N, K, "wei"), scale_b);
+  args.insert({DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, src_sc_mem});
+  args.insert({DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, wei_sc_mem});
+  if (with_dst_scale) {
+    // Bind single f32 scalar as DST scale
+    at::Tensor dst_scale_f32 = scale_result->to(at::kFloat).contiguous();
+    dnnl::memory::desc dst_sc_md(
+        {1}, dnnl::memory::data_type::f32, dnnl::memory::format_tag::x);
+    auto dst_sc_mem =
+        make_onednn_memory(dst_sc_md, engine, dst_scale_f32.data_ptr());
+    args.insert({DNNL_ARG_ATTR_SCALES | DNNL_ARG_DST, dst_sc_mem});
+  }
+
+  dnnl::matmul matmul_p = dnnl::matmul(matmul_pd);
+  sycl::event matmul_fwd_event =
+      dnnl::sycl_interop::execute(matmul_p, stream, args);
+  return matmul_fwd_event;
+}
+
 } // namespace at::native::onednn

aten/src/ATen/native/mkldnn/xpu/detail/Utils.cpp

@@ -78,6 +78,10 @@ dnnl::memory::data_type get_onednn_dtype(
       return dnnl::memory::data_type::f32;
     case at::ScalarType::BFloat16:
       return dnnl::memory::data_type::bf16;
+    case at::ScalarType::Float8_e4m3fn:
+      return dnnl::memory::data_type::f8_e4m3;
+    case at::ScalarType::Float8_e5m2:
+      return dnnl::memory::data_type::f8_e5m2;
     default:
       if (!allow_undef) {
         TORCH_CHECK(

aten/src/ATen/native/mkldnn/xpu/detail/oneDNN.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include <ATen/ATen.h>
+#include <ATen/BlasBackend.h>
 #include <ATen/native/mkldnn/xpu/detail/Attr.h>
 #include <ATen/native/mkldnn/xpu/detail/Utils.h>
 #include <ATen/native/mkldnn/xpu/detail/oneDNNContext.h>
@@ -202,4 +203,16 @@ void sdpa_backward(
     Tensor& grad_query,
     Tensor& grad_key,
     Tensor& grad_value);
+
+sycl::event scaled_matmul(
+    const Tensor& mat1,
+    const Tensor& mat2,
+    Tensor& result,
+    const Tensor& scale_a,
+    const Tensor& scale_b,
+    at::blas::ScalingType scaling_choice_a,
+    at::blas::ScalingType scaling_choice_b,
+    const std::optional<at::Tensor>& bias,
+    const std::optional<at::Tensor>& scale_result,
+    bool use_fast_accum);
 } // namespace at::native::onednn

aten/src/ATen/native/mps/OperationUtils.h

@@ -82,6 +82,7 @@ NSArray<NSNumber*>* getTensorAxes(const TensorBase& t);
 NSArray<NSNumber*>* getTensorAxes(const IntArrayRef& sizes, at::OptionalIntArrayRef dim);
 std::string getMPSShapeString(MPSShape* shape);
 std::string getTensorsStringKey(const TensorList& tensors, bool short_dtype = true, bool exclude_shape = false);
+std::string to_hex_key(float);
 std::string getArrayRefString(const IntArrayRef s);
 // use has_storage() on the returned tensor to determine if src actually is a view
 Tensor gatherViewTensor(const Tensor& src, Tensor& dst);

aten/src/ATen/native/mps/OperationUtils.mm

@@ -301,6 +301,10 @@ std::string getArrayRefString(const IntArrayRef s) {
   return fmt::to_string(fmt::join(s, ","));
 }
 
+std::string to_hex_key(float f) {
+  return fmt::format("{:a}", f);
+}
+
 std::string getTensorsStringKey(const TensorList& tensors, bool short_dtype, bool exclude_shape) {
   fmt::basic_memory_buffer<char, 100> buffer;
   auto buf_iterator = std::back_inserter(buffer);

aten/src/ATen/native/mps/kernels/LinearAlgebra.metal

@@ -40,7 +40,7 @@ inline c10::metal::opmath_t<T> matmul_inner(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     for (uint k = 0; k < TILE_DIM; k++) {
-      sum += A_tile[tid.y][k] * B_tile[k][tid.x];
+      sum += c10::metal::mul(A_tile[tid.y][k], B_tile[k][tid.x]);
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
@@ -96,7 +96,9 @@ kernel void addmm(
     auto bias =
         biasData[thread_id.y * strides[3].x + thread_id.x * strides[3].y];
     outputData[thread_id.y * strides[2].x + thread_id.x * strides[2].y] =
-        static_cast<T>(alpha_beta[0] * sum + alpha_beta[1] * bias);
+        static_cast<T>(
+            c10::metal::mul(alpha_beta[0], sum) +
+            c10::metal::mul(alpha_beta[1], bias));
   }
 }
 
@@ -832,6 +834,10 @@ INSTANTIATE_MM_OPS(float);
 INSTANTIATE_MM_OPS(half);
 INSTANTIATE_MM_OPS(bfloat);
 
+// Complex MM
+INSTANTIATE_MM_OPS(float2);
+INSTANTIATE_MM_OPS(half2);
+
 // Integral MM
 INSTANTIATE_MM_OPS(long);
 INSTANTIATE_MM_OPS(int);

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

@@ -121,7 +121,7 @@ Tensor& do_metal_addmm(const Tensor& self,
                        const Scalar& alpha,
                        const Scalar& beta,
                        const Tensor& bias) {
-  if (beta.toDouble() == 0 && alpha.toDouble() == 1) {
+  if (beta.isFloatingPoint() && alpha.isFloatingPoint() && beta.toDouble() == 0 && alpha.toDouble() == 1) {
     return do_metal_mm(self, other, output);
   }
   auto stream = getCurrentMPSStream();
@@ -147,13 +147,15 @@ Tensor& do_metal_addmm(const Tensor& self,
         std::array<int64_t, 2> i64;
         std::array<int32_t, 2> i32;
         std::array<float, 2> f32;
-      } alpha_beta;
+        std::array<c10::complex<float>, 2> c64;
+      } alpha_beta{};
       if (output.scalar_type() == kLong) {
         alpha_beta.i64 = {alpha.toLong(), beta.toLong()};
       } else if (c10::isIntegralType(output.scalar_type(), true)) {
         alpha_beta.i32 = {alpha.toInt(), beta.toInt()};
+      } else if (c10::isComplexType(output.scalar_type())) {
+        alpha_beta.c64 = {alpha.toComplexFloat(), beta.toComplexFloat()};
       } else {
-        TORCH_INTERNAL_ASSERT(c10::isFloatingType(output.scalar_type()));
         alpha_beta.f32 = {alpha.toFloat(), beta.toFloat()};
       }
       constexpr uint32_t TILE_DIM = 16; // fastest performance from tests on multiple macs
@@ -190,10 +192,16 @@ std::tuple<MPSGraphTensor*, MPSGraphTensor*, MPSGraphTensor*> do_mm(MPSGraph* gr
 bool use_metal_mm(const Tensor& self, const Tensor& other, const Tensor& output) {
   static bool always_use_metal = c10::utils::has_env("PYTORCH_MPS_PREFER_METAL");
   constexpr auto max_stride_size = 32768;
+  constexpr auto max_complex_inner_size = 2048;
   static bool is_macos_14_4_or_newer = is_macos_13_or_newer(MacOSVersion::MACOS_VER_14_4_PLUS);
   if (always_use_metal || c10::isIntegralType(self.scalar_type(), true)) {
     return true;
   }
+  // multiplicationWithPrimaryTensor: returns incorrect results if inner size exceeds 2048
+  // See https://github.com/pytorch/pytorch/issues/167727#issuecomment-3529308548
+  if (c10::isComplexType(self.scalar_type()) && self.size(1) > max_complex_inner_size) {
+    return true;
+  }
   return !is_macos_14_4_or_newer &&
       (self.stride(0) > max_stride_size || self.stride(1) > max_stride_size || self.size(0) > max_stride_size ||
        self.size(1) > max_stride_size || other.stride(0) > max_stride_size || other.stride(1) > max_stride_size ||

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

@@ -244,8 +244,8 @@ static void clamp_scalar_out_mps(const Tensor& input_t,
 
   @autoreleasepool {
     // the optional min/max refs could affect how we build the cached graph
-    std::string key = op_name + (has_min ? ("_min:" + std::to_string(min_scalar)) : "") +
-        (has_max ? ("_max:" + std::to_string(max_scalar)) : "") + "_scalar:" + getTensorsStringKey({input_t});
+    std::string key = op_name + (has_min ? ("_min:" + to_hex_key(min_scalar)) : "") +
+        (has_max ? ("_max:" + to_hex_key(max_scalar)) : "") + "_scalar:" + getTensorsStringKey({input_t});
     auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
       if (has_min)
         newCachedGraph->minTensor = [mpsGraph constantWithScalar:min_scalar

aten/src/ATen/native/native_functions.yaml

@@ -7518,7 +7518,7 @@
 - func: _sparse_mask_projection(Tensor self, Tensor mask, bool accumulate_matches=False) -> Tensor
   variants: method
   dispatch:
-    SparseCPU, SparseCUDA: sparse_mask_projection
+    SparseCPU, SparseCUDA, SparseMPS: sparse_mask_projection
   autogen: _sparse_mask_projection.out
 
 - func: _to_cpu(Tensor[] tensors) -> Tensor[]

aten/src/ATen/native/sparse/cuda/SoftMax.cu

@@ -30,10 +30,12 @@
 
 #include <thrust/binary_search.h>
 #include <thrust/device_ptr.h>
+#include <thrust/distance.h>
+#include <thrust/iterator/constant_iterator.h>
+#include <thrust/scan.h>
 #include <thrust/sequence.h>
 #include <thrust/sort.h>
 #include <thrust/system/cuda/execution_policy.h>
-#include <thrust/iterator/constant_iterator.h>
 
 #include <cuda_runtime_api.h>
 #include <cusparse.h>

aten/src/ATen/native/sparse/mps/SparseMPSTensorMath.mm

@@ -445,6 +445,33 @@ static SparseTensor& mul_out_dense_sparse_mps(
   return out;
 }
 
+static std::tuple<Tensor, Tensor, int64_t> mps_intersect_binary_search(
+    const Tensor& A_keys,
+    const Tensor& B_keys,
+    int64_t lenA,
+    int64_t lenB,
+    bool boolean_flag) {
+
+  auto stream = getCurrentMPSStream();
+  auto outA_idx = at::empty({lenA}, A_keys.options().dtype(at::kLong));
+  auto outB_idx = at::empty({lenA}, A_keys.options().dtype(at::kLong));
+  auto counter = at::zeros({1}, A_keys.options().dtype(at::kInt));
+
+  dispatch_sync_with_rethrow(stream->queue(), ^() {
+    @autoreleasepool {
+      auto pso = lib.getPipelineStateForFunc("intersect_binary_search");
+      auto enc = stream->commandEncoder();
+      [enc setComputePipelineState:pso];
+      mtl_setArgs(enc, A_keys, B_keys, outA_idx, outB_idx, counter,
+                  static_cast<uint32_t>(lenB), boolean_flag);
+      mtl_dispatch1DJob(enc, pso, static_cast<uint32_t>(lenA));
+    }
+  });
+
+  const auto match_count = static_cast<int64_t>(counter.item<int32_t>());
+  return std::make_tuple(std::move(outA_idx), std::move(outB_idx), match_count);
+}
+
 
 SparseTensor& mul_out_sparse_mps(const Tensor& t_, const Tensor& src_, SparseTensor& r_) {
   TORCH_CHECK(r_.is_mps(), "mul: expected 'out' to be MPS, but got ", r_.device());
@@ -523,22 +550,10 @@ SparseTensor& mul_out_sparse_mps(const Tensor& t_, const Tensor& src_, SparseTen
   auto A_keys = A_is_lhs ? lhs_keys : rhs_keys;
   auto B_keys = A_is_lhs ? rhs_keys : lhs_keys;
 
-  auto outA_idx = at::empty({lenA}, at::device(device).dtype(kLong));
-  auto outB_idx = at::empty({lenA}, at::device(device).dtype(kLong));
-  auto counter = at::zeros({1}, at::device(device).dtype(kInt));
+  auto [outA_idx, outB_idx, M_int64] = mps_intersect_binary_search(
+      A_keys, B_keys, lenA, lenB, A_is_lhs);
 
-  dispatch_sync_with_rethrow(stream->queue(), ^() {
-    @autoreleasepool {
-      auto pso = lib.getPipelineStateForFunc("intersect_binary_search");
-      auto enc = stream->commandEncoder();
-      [enc setComputePipelineState:pso];
-      mtl_setArgs(enc, A_keys, B_keys, outA_idx, outB_idx, counter,
-                  static_cast<uint32_t>(lenB), A_is_lhs);
-      mtl_dispatch1DJob(enc, pso, static_cast<uint32_t>(lenA));
-    }
-  });
-
-  const uint32_t M = counter.item<int32_t>(); // number of structural matches
+  const auto M = static_cast<uint32_t>(M_int64); // number of structural matches
 
   r_.resize_as_(lhs);
 
@@ -762,6 +777,14 @@ SparseTensor& add_out_sparse_mps(const SparseTensor& self,
 
 using OptTensor = std::optional<Tensor>;
 
+static Tensor create_sparse_output_values(
+    const Tensor& template_values,
+    int64_t output_nnz,
+    ScalarType dtype) {
+  auto out_val_sizes = template_values.sizes().vec();
+  out_val_sizes[0] = output_nnz;
+  return at::zeros(out_val_sizes, template_values.options().dtype(dtype));
+}
 
 static void sparse_mask_apply_out_mps_kernel(
     Tensor& result,
@@ -783,9 +806,9 @@ static void sparse_mask_apply_out_mps_kernel(
   auto src  = src_in.coalesce();
   auto mask = coalesce_mask ? mask_in.coalesce() : mask_in;
 
-  const int64_t src_nnz = src._nnz();
-  const int64_t mask_nnz = mask._nnz();
-  const int64_t sd = src.sparse_dim();
+  const auto src_nnz = src._nnz();
+  const auto mask_nnz = mask._nnz();
+  const auto sd = src.sparse_dim();
   result.sparse_resize_(mask.sizes(), mask.sparse_dim(), mask.dense_dim());
 
   auto commonDtype = at::result_type(src, mask);
@@ -814,53 +837,27 @@ static void sparse_mask_apply_out_mps_kernel(
     return;
   }
 
+  auto mask_indices = mask._indices().contiguous();
+  auto src_values = src._values().to(commonDtype).contiguous();
+  auto out_values = create_sparse_output_values(src_values, mask_nnz, commonDtype);
+
   if (src_nnz == 0) {
-    auto out_indices = mask._indices().contiguous();
-    auto src_values  = src._values().to(commonDtype);
-    auto out_val_sizes = src_values.sizes().vec();
-    out_val_sizes[0] = mask_nnz;
-    auto out_values = at::zeros(out_val_sizes, src_values.options());
-    alias_into_sparse(result, out_indices, out_values);
+    alias_into_sparse(result, mask_indices, out_values);
     result._coalesced_(mask.is_coalesced());
     return;
   }
 
-  auto mask_indices = mask._indices().contiguous();
-  auto src_indices = src._indices().contiguous();
-  auto src_values = src._values().to(commonDtype).contiguous();
+  auto mask_keys = flatten_indices(mask._indices().contiguous(), mask.sizes().slice(0, sd)).contiguous();
+  auto src_keys  = flatten_indices(src._indices().contiguous(), src.sizes().slice(0, sd)).contiguous();
 
-  auto mask_keys = flatten_indices(mask_indices, mask.sizes().slice(0, sd)).contiguous();
-  auto src_keys  = flatten_indices(src_indices,  src.sizes().slice(0, sd)).contiguous();
-
-  const bool A_is_src = (src_nnz <= mask_nnz);
-  const int64_t lenA = A_is_src ? src_nnz  : mask_nnz;
-  const int64_t lenB = A_is_src ? mask_nnz : src_nnz;
+  const auto A_is_src = (src_nnz <= mask_nnz);
+  const auto lenA = A_is_src ? src_nnz  : mask_nnz;
+  const auto lenB = A_is_src ? mask_nnz : src_nnz;
   auto A_keys = A_is_src ? src_keys  : mask_keys;
   auto B_keys = A_is_src ? mask_keys : src_keys;
 
-  const auto device = result.device();
-  auto stream = getCurrentMPSStream();
-
-  auto outA_idx = at::empty({lenA}, at::device(device).dtype(at::kLong));
-  auto outB_idx = at::empty({lenA}, at::device(device).dtype(at::kLong));
-  auto counter = at::zeros({1}, at::device(device).dtype(at::kInt));
-
-  dispatch_sync_with_rethrow(stream->queue(), ^() {
-    @autoreleasepool {
-      auto pso = lib.getPipelineStateForFunc("intersect_binary_search");
-      auto enc = stream->commandEncoder();
-      [enc setComputePipelineState:pso];
-      mtl_setArgs(enc, A_keys, B_keys, outA_idx, outB_idx, counter,
-                  static_cast<uint32_t>(lenB), A_is_src);
-      mtl_dispatch1DJob(enc, pso, static_cast<uint32_t>(lenA));
-    }
-  });
-
-  const int64_t M = static_cast<int64_t>(counter.item<int32_t>());
-
-  auto out_val_sizes = src_values.sizes().vec();
-  out_val_sizes[0] = mask_nnz;
-  auto out_values = at::zeros(out_val_sizes, src_values.options());
+  auto [outA_idx, outB_idx, M] = mps_intersect_binary_search(
+      A_keys, B_keys, lenA, lenB, A_is_src);
 
   if (M > 0) {
     auto src_match = outA_idx.narrow(0, 0, M);
@@ -878,6 +875,70 @@ static void sparse_mask_apply_out_mps_kernel(
   result._coalesced_(mask.is_coalesced());
 }
 
+static void sparse_mask_projection_out_mps_kernel(
+    Tensor& result,
+    const Tensor& lhs,
+    const Tensor& rhs,
+    const OptTensor& /*x_hash_opt*/,
+    bool accumulate_matches) {
+
+  TORCH_CHECK(lhs.is_sparse() && rhs.is_sparse(), "sparse_mask_projection: expected sparse COO");
+  TORCH_CHECK(lhs.is_mps() && rhs.is_mps(), "sparse_mask_projection: expected MPS tensors");
+  TORCH_CHECK(lhs.sparse_dim() == rhs.sparse_dim(), "sparse_dim mismatch");
+
+  auto lhs_c = lhs.coalesce();
+  auto rhs_c = rhs.coalesce();
+
+  const auto sd = lhs_c.sparse_dim();
+  const auto lhs_nnz = lhs_c._nnz();
+  const auto rhs_nnz = rhs_c._nnz();
+
+  auto commonDtype = at::result_type(lhs_c, rhs_c);
+  TORCH_CHECK(canCast(commonDtype, result.scalar_type()),
+              "Can't convert ", commonDtype, " to output ", result.scalar_type());
+
+  result.sparse_resize_(lhs.sizes(), lhs.sparse_dim(), lhs.dense_dim());
+
+  auto lhs_indices = lhs_c._indices().contiguous();
+  auto rhs_values  = rhs_c._values().to(commonDtype).contiguous();
+  auto out_values = create_sparse_output_values(rhs_values, lhs_nnz, commonDtype);
+
+  if (lhs_nnz > 0 && rhs_nnz > 0) {
+    auto lhs_keys = flatten_indices(lhs_indices, lhs_c.sizes().slice(0, sd)).contiguous();
+    auto rhs_keys = flatten_indices(rhs_c._indices().contiguous(), rhs_c.sizes().slice(0, sd)).contiguous();
+
+    const auto A_is_lhs = (lhs_nnz <= rhs_nnz);
+    const auto lenA = A_is_lhs ? lhs_nnz : rhs_nnz;
+    const auto lenB = A_is_lhs ? rhs_nnz : lhs_nnz;
+    auto A_keys = A_is_lhs ? lhs_keys : rhs_keys;
+    auto B_keys = A_is_lhs ? rhs_keys : lhs_keys;
+
+    auto [outA_idx, outB_idx, M] = mps_intersect_binary_search(
+        A_keys, B_keys, lenA, lenB, A_is_lhs);
+
+    if (M > 0) {
+      auto idx_in_A = outA_idx.narrow(0, 0, M);
+      auto idx_in_B = outB_idx.narrow(0, 0, M);
+      auto idx_in_lhs = A_is_lhs ? idx_in_A : idx_in_B;
+      auto idx_in_rhs = A_is_lhs ? idx_in_B : idx_in_A;
+
+      const auto view_cols = rhs_values.numel() / std::max<int64_t>(rhs_nnz, 1);
+      auto rhs_rows = rhs_values.index_select(0, idx_in_rhs).contiguous();
+      auto rhs_rows_2d = rhs_rows.view({M, view_cols});
+      auto out_2d = out_values.view({lhs_nnz, view_cols});
+
+      if (accumulate_matches) {
+        out_2d.index_add_(0, idx_in_lhs, rhs_rows_2d);
+      } else {
+        out_2d.index_copy_(0, idx_in_lhs, rhs_rows_2d);
+      }
+    }
+  }
+
+  alias_into_sparse(result, lhs._indices(), out_values);
+  result._coalesced_(lhs.is_coalesced());
+}
+
 static void sparse_mask_intersection_out_mps_kernel(
     Tensor& result,
     const Tensor& lhs,
@@ -1002,4 +1063,5 @@ Tensor sparse_sparse_matmul_mps(const Tensor& mat1_, const Tensor& mat2_) {
 }
 
 REGISTER_MPS_DISPATCH(sparse_mask_intersection_out_stub, &sparse_mask_intersection_out_mps_kernel);
+REGISTER_MPS_DISPATCH(sparse_mask_projection_out_stub, &sparse_mask_projection_out_mps_kernel);
 } // namespace at::native

aten/src/ATen/xpu/XPUEvent.h

@@ -1,191 +1,3 @@
 #pragma once
 #include <ATen/xpu/XPUContext.h>
-
-#include <optional>
-
-namespace at::xpu {
-
-/*
- * XPUEvent are movable not copyable wrappers around SYCL event. XPUEvent are
- * constructed lazily when first recorded. It has a device, and this device is
- * acquired from the first recording stream. Later streams that record the event
- * must match the same device.
- *
- * Currently, XPUEvent does NOT support to export an inter-process event from
- * another process via inter-process communication(IPC). So it means that
- * inter-process communication for event handles between different processes is
- * not available. This could impact some applications that rely on cross-process
- * synchronization and communication.
- */
-struct TORCH_XPU_API XPUEvent {
-  // Constructors
-  XPUEvent(bool enable_timing = false) noexcept
-      : enable_timing_{enable_timing} {}
-
-  ~XPUEvent() {
-    if (isCreated()) {
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_deletion(
-            at::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
-      }
-    }
-  }
-
-  XPUEvent(const XPUEvent&) = delete;
-  XPUEvent& operator=(const XPUEvent&) = delete;
-
-  XPUEvent(XPUEvent&& other) = default;
-  XPUEvent& operator=(XPUEvent&& other) = default;
-
-  operator sycl::event&() const {
-    return event();
-  }
-
-  std::optional<at::Device> device() const {
-    if (isCreated()) {
-      return at::Device(at::kXPU, device_index_);
-    } else {
-      return std::nullopt;
-    }
-  }
-
-  inline bool isCreated() const {
-    return (event_.get() != nullptr);
-  }
-
-  DeviceIndex device_index() const {
-    return device_index_;
-  }
-
-  sycl::event& event() const {
-    return *event_;
-  }
-
-  bool query() const {
-    using namespace sycl::info;
-    if (!isCreated()) {
-      return true;
-    }
-
-    return event().get_info<event::command_execution_status>() ==
-        event_command_status::complete;
-  }
-
-  void record() {
-    record(getCurrentXPUStream());
-  }
-
-  void recordOnce(const XPUStream& stream) {
-    if (!isCreated()) {
-      record(stream);
-    }
-  }
-
-  void record(const XPUStream& stream) {
-    if (!isCreated()) {
-      device_index_ = stream.device_index();
-      assignEvent(stream.queue());
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_creation(
-            at::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
-      }
-    } else {
-      TORCH_CHECK(
-          device_index_ == stream.device_index(),
-          "Event device ",
-          device_index_,
-          " does not match recording stream's device ",
-          stream.device_index(),
-          ".");
-      reassignEvent(stream.queue());
-    }
-    const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-    if (C10_UNLIKELY(interp)) {
-      (*interp)->trace_gpu_event_record(
-          at::kXPU,
-          reinterpret_cast<uintptr_t>(event_.get()),
-          reinterpret_cast<uintptr_t>(&stream.queue()));
-    }
-  }
-
-  void block(const XPUStream& stream) {
-    if (isCreated()) {
-      std::vector<sycl::event> event_list{event()};
-      // Make this stream wait until event_ is completed.
-      stream.queue().ext_oneapi_submit_barrier(event_list);
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_wait(
-            at::kXPU,
-            reinterpret_cast<uintptr_t>(event_.get()),
-            reinterpret_cast<uintptr_t>(&stream.queue()));
-      }
-    }
-  }
-
-  double elapsed_time(const XPUEvent& other) const {
-    TORCH_CHECK(
-        isCreated() && other.isCreated(),
-        "Both events must be recorded before calculating elapsed time.");
-    TORCH_CHECK(
-        query() && other.query(),
-        "Both events must be completed before calculating elapsed time.");
-    TORCH_CHECK(
-        enable_timing_ && other.enable_timing_,
-        "Both events must be created with argument 'enable_timing=True'.");
-
-#if SYCL_COMPILER_VERSION < 20250000
-    TORCH_CHECK_NOT_IMPLEMENTED(
-        false,
-        "elapsed_time of XPUEvent requires PyTorch to be built with SYCL compiler version 2025.0.0 or newer.");
-#endif
-
-    using namespace sycl::info::event_profiling;
-    // Block until both of the recorded events are completed.
-    uint64_t end_time_ns = other.event().get_profiling_info<command_end>();
-    uint64_t start_time_ns = event().get_profiling_info<command_end>();
-    // Return the eplased time in milliseconds.
-    return 1e-6 *
-        (static_cast<double>(end_time_ns) - static_cast<double>(start_time_ns));
-  }
-
-  void synchronize() const {
-    if (isCreated()) {
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_synchronization(
-            at::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
-      }
-      event().wait_and_throw();
-    }
-  }
-
- private:
-  void assignEvent(sycl::queue& queue) {
-#if SYCL_COMPILER_VERSION >= 20250000
-    if (enable_timing_) {
-      event_ = std::make_unique<sycl::event>(
-          sycl::ext::oneapi::experimental::submit_profiling_tag(queue));
-    } else {
-      event_ = std::make_unique<sycl::event>(queue.ext_oneapi_submit_barrier());
-    }
-#else
-    event_ = std::make_unique<sycl::event>(queue.ext_oneapi_submit_barrier());
-#endif
-  }
-
-  void reassignEvent(sycl::queue& queue) {
-    event_.reset();
-    assignEvent(queue);
-  }
-
-  bool enable_timing_ = false;
-  DeviceIndex device_index_ = -1;
-  // Only need to track the last event, as events in an in-order queue are
-  // executed sequentially.
-  std::unique_ptr<sycl::event> event_;
-};
-
-} // namespace at::xpu
+#include <c10/xpu/XPUEvent.h>

benchmarks/dynamo/check_accuracy.py

@@ -50,6 +50,7 @@ def check_accuracy(actual_csv, expected_csv, expected_filename):
                 "mobilenet_v2",
                 "pytorch_CycleGAN_and_pix2pix",
                 "pytorch_stargan",
+                "repvgg_a2",
                 "resnet152",
                 "resnet18",
                 "resnet50",

benchmarks/dynamo/ci_expected_accuracy/dynamic_inductor_timm_training.csv

@@ -10,7 +10,7 @@ beit_base_patch16_224,pass,7
 
 
 
-convnextv2_nano.fcmae_ft_in22k_in1k,pass,7
+convnextv2_nano.fcmae_ft_in22k_in1k,fail_accuracy,7
 
 
 
@@ -66,7 +66,7 @@ visformer_small,pass,7
 
 
 
-vit_base_patch14_dinov2.lvd142m,pass,7
+vit_base_patch14_dinov2.lvd142m,fail_accuracy,7
 
 
 

benchmarks/dynamo/ci_expected_accuracy/rocm/dynamic_inductor_timm_training.csv

@@ -50,7 +50,7 @@ nfnet_l0,pass,7
 
 
 
-repvgg_a2,fail_accuracy,7
+repvgg_a2,pass,7
 
 
 

benchmarks/dynamo/common.py

@@ -2288,11 +2288,9 @@ class BenchmarkRunner:
                     )
                 ):
                     is_same = False
-            except Exception as e:
+            except Exception:
                 # Sometimes torch.allclose may throw RuntimeError
-                exception_string = str(e)
-                accuracy_status = f"fail_exception: {exception_string}"
-                return record_status(accuracy_status, dynamo_start_stats=start_stats)
+                is_same = False
 
             if not is_same:
                 accuracy_status = "eager_two_runs_differ"
@@ -2409,11 +2407,9 @@ class BenchmarkRunner:
                     force_max_multiplier=force_max_multiplier,
                 ):
                     is_same = False
-            except Exception as e:
+            except Exception:
                 # Sometimes torch.allclose may throw RuntimeError
-                exception_string = str(e)
-                accuracy_status = f"fail_exception: {exception_string}"
-                return record_status(accuracy_status, dynamo_start_stats=start_stats)
+                is_same = False
 
             if not is_same:
                 if self.args.skip_accuracy_check:

benchmarks/dynamo/pr_time_benchmarks/benchmarks/dtensor.py

@@ -0,0 +1,62 @@
+import sys
+
+from benchmark_base import BenchmarkBase
+
+import torch
+from torch.distributed._tensor import DTensor, Replicate
+from torch.testing._internal.distributed.fake_pg import FakeStore
+
+
+class BenchmarkDTensorDispatch(BenchmarkBase):
+    def __init__(self, operator, world_size) -> None:
+        super().__init__(
+            category=f"dtensor_dispatch_{operator}",
+            device="cuda",
+        )
+        self.world_size = world_size
+
+    def name(self) -> str:
+        prefix = f"{self.category()}"
+        return prefix
+
+    def description(self) -> str:
+        return f"DTensor dispatch time for {self.category()}"
+
+    def _prepare_once(self) -> None:
+        self.mesh = torch.distributed.device_mesh.init_device_mesh(
+            "cuda", (self.world_size,), mesh_dim_names=("dp",)
+        )
+        self.a = DTensor.from_local(
+            torch.ones(10, 10, device=self.device()), self.mesh, [Replicate()]
+        )
+        self.b = DTensor.from_local(
+            torch.ones(10, 10, device=self.device()), self.mesh, [Replicate()]
+        )
+
+    def _prepare(self) -> None:
+        pass
+
+
+class BenchmarkDetach(BenchmarkDTensorDispatch):
+    def __init__(self, world_size) -> None:
+        super().__init__(operator="detach", world_size=world_size)
+
+    def _work(self) -> None:
+        self.a.detach()
+
+
+def main():
+    world_size = 256
+    fake_store = FakeStore()
+    torch.distributed.init_process_group(
+        "fake", store=fake_store, rank=0, world_size=world_size
+    )
+    result_path = sys.argv[1]
+    BenchmarkDetach(world_size).enable_instruction_count().collect_all().append_results(
+        result_path
+    )
+    torch.distributed.destroy_process_group()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/dynamo/torchbench.yaml

@@ -189,6 +189,10 @@ skip:
     - hf_Whisper
     - hf_distil_whisper
     - timm_vision_transformer_large
+    # https://github.com/pytorch/pytorch/issues/167895
+    - stable_diffusion
+    - stable_diffusion_text_encoder
+    - stable_diffusion_unet
 
   device:
     cpu:

benchmarks/transformer/score_mod.py

@@ -125,6 +125,17 @@ AttentionType = Literal[
 ]
 DtypeString = Literal["bfloat16", "float16", "float32"]
 SpeedupType = Literal["fwd", "bwd"]
+# Operator Name mapping
+backend_to_operator_name = {
+    "math": "math attention kernel",
+    "efficient": "efficient attention kernel",
+    "cudnn": "cudnn attention kernel",
+    "fav2": "flash attention 2 kernel",
+    "fav3": "flash attention 3 kernel",
+    "fakv": "flash attention kv cache kernel",
+    "og-eager": "eager attention kernel",
+    "flex": "flex attention kernel",
+}
 
 
 def benchmark_torch_function_in_microseconds(func: Callable, *args, **kwargs) -> float:
@@ -1265,12 +1276,14 @@ def _output_json_for_dashboard(
                 model: ModelInfo
                 metric: MetricInfo
 
+            operator_name = backend_to_operator_name.get(backend, backend)
+
             # Benchmark extra info
             benchmark_extra_info = {
                 "input_config": input_config,
                 "device": device,
                 "arch": device_arch,
-                "operator_name": backend,
+                "operator_name": operator_name,
                 "attn_type": config.attn_type,
                 "shape": str(config.shape),
                 "max_autotune": config.max_autotune,
@@ -1288,7 +1301,7 @@ def _output_json_for_dashboard(
                     type="attention-benchmark",
                     origins=["pytorch"],
                     extra_info={
-                        "operator_name": backend,
+                        "operator_name": operator_name,
                         "attn_type": config.attn_type,
                     },
                 ),
@@ -1315,7 +1328,7 @@ def _output_json_for_dashboard(
                         type="attention-benchmark",
                         origins=["pytorch"],
                         extra_info={
-                            "operator_name": backend,
+                            "operator_name": operator_name,
                         },
                     ),
                     metric=MetricInfo(
@@ -1341,7 +1354,7 @@ def _output_json_for_dashboard(
                         type="attention-benchmark",
                         origins=["pytorch"],
                         extra_info={
-                            "operator_name": backend,
+                            "operator_name": operator_name,
                         },
                     ),
                     metric=MetricInfo(
@@ -1371,7 +1384,7 @@ def _output_json_for_dashboard(
                         type="attention-benchmark",
                         origins=["pytorch"],
                         extra_info={
-                            "operator_name": backend,
+                            "operator_name": operator_name,
                         },
                     ),
                     metric=MetricInfo(

buckbuild.bzl

@@ -2,6 +2,7 @@
 # These load paths point to different files in internal and OSS environment
 
 load("@bazel_skylib//lib:paths.bzl", "paths")
+load("//tools/build_defs:cell_defs.bzl", "get_fbsource_cell")
 load("//tools/build_defs:fb_native_wrapper.bzl", "fb_native")
 load("//tools/build_defs:fb_xplat_cxx_library.bzl", "fb_xplat_cxx_library")
 load("//tools/build_defs:fb_xplat_genrule.bzl", "fb_xplat_genrule")
@@ -590,6 +591,9 @@ def pt_operator_query_codegen(
         pt_allow_forced_schema_registration = True,
         compatible_with = [],
         apple_sdks = None):
+    if get_fbsource_cell() == "fbcode":
+        return
+
     oplist_dir_name = name + "_pt_oplist"
 
     # @lint-ignore BUCKLINT
@@ -865,6 +869,9 @@ def define_buck_targets(
         pt_xplat_cxx_library = fb_xplat_cxx_library,
         c2_fbandroid_xplat_compiler_flags = [],
         labels = []):
+    if get_fbsource_cell() == "fbcode":
+        return
+
     # @lint-ignore BUCKLINT
     fb_native.filegroup(
         name = "metal_build_srcs",

c10/core/Allocator.h

@@ -19,6 +19,17 @@
 
 namespace c10 {
 
+using CaptureId_t = unsigned long long;
+// first is set if the instance is created by CUDAGraph::capture_begin.
+// second is set if the instance is created by at::cuda::graph_pool_handle.
+using MempoolId_t = std::pair<CaptureId_t, CaptureId_t>;
+
+struct MempoolIdHash {
+  std::size_t operator()(const MempoolId_t& mempool_id) const noexcept {
+    return mempool_id.first != 0 ? mempool_id.first : mempool_id.second;
+  }
+};
+
 // A DataPtr is a unique pointer (with an attached deleter and some
 // context for the deleter) to some memory, which also records what
 // device is for its data.

c10/core/CachingDeviceAllocator.h

@@ -99,7 +99,10 @@ struct C10_API DeviceAllocator : public c10::Allocator {
 
   // Return the free memory size and total memory size in bytes for the
   // specified device.
-  virtual std::pair<size_t, size_t> getMemoryInfo(c10::DeviceIndex device) = 0;
+  virtual std::pair<size_t, size_t> getMemoryInfo(c10::DeviceIndex device) {
+    TORCH_CHECK_NOT_IMPLEMENTED(
+        false, "getMemoryInfo is not implemented for this allocator yet.");
+  }
 };
 
 // This function is used to get the DeviceAllocator for a specific device type

c10/core/ScalarType.h

@@ -27,26 +27,13 @@
 #include <torch/headeronly/core/ScalarType.h>
 
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
 
 namespace c10 {
 
 // See [dtype Macros note] in torch/headeronly/core/ScalarType.h
 // regarding macros.
 
-template <typename T>
-struct CppTypeToScalarType;
-
-#define SPECIALIZE_CppTypeToScalarType(cpp_type, scalar_type)                  \
-  template <>                                                                  \
-  struct CppTypeToScalarType<cpp_type>                                         \
-      : std::                                                                  \
-            integral_constant<c10::ScalarType, c10::ScalarType::scalar_type> { \
-  };
-
-AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(SPECIALIZE_CppTypeToScalarType)
-
-#undef SPECIALIZE_CppTypeToScalarType
-
 #define DEFINE_CONSTANT(_, name) \
   constexpr ScalarType k##name = ScalarType::name;
 
@@ -205,6 +192,12 @@ inline bool isSignedType(ScalarType t) {
       break;
       // Do not add default here, but rather define behavior of every new entry
       // here.  `-Wswitch-enum` would raise a warning in those cases.
+      // TODO: get PyTorch to adopt exhaustive switches by default with a way to
+      // opt specific switches to being non-exhaustive.
+      // Exhaustive:
+      // `-Wswitch-enum`, `-Wswitch-default`, `-Wno-covered-switch-default`
+      // Non-Exhaustive:
+      // `-Wno-switch-enum`, `-Wswitch-default`, `-Wcovered-switch-default`
   }
   TORCH_CHECK(false, "Unknown ScalarType ", t);
 #undef CASE_ISSIGNED

c10/core/TensorImpl.h

@@ -57,6 +57,8 @@ C10_DECLARE_bool(caffe2_keep_on_shrink);
 // respect caffe2_keep_on_shrink.
 C10_DECLARE_int64(caffe2_max_keep_on_shrink_memory);
 
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
+
 namespace at {
 class Tensor;
 class TensorBase;
@@ -3303,3 +3305,5 @@ static_assert(
 #undef C10_GCC_VERSION_MINOR
 
 } // namespace c10
+
+C10_DIAGNOSTIC_POP()

c10/cuda/CUDACachingAllocator.cpp

@@ -1012,12 +1012,6 @@ PrivatePoolState::PrivatePoolState(
   }
 }
 
-struct MempoolIdHash {
-  std::size_t operator()(const MempoolId_t& mempool_id) const noexcept {
-    return mempool_id.first != 0 ? mempool_id.first : mempool_id.second;
-  }
-};
-
 cudaError_t allocPrimitive(void** ptr, size_t size, AllocParams& p) {
   if (p.pool->owner_PrivatePool && p.pool->owner_PrivatePool->allocator()) {
     *ptr = p.pool->owner_PrivatePool->allocator()->raw_alloc(size);
@@ -4510,66 +4504,3 @@ std::atomic<CUDAAllocator*> allocator;
 static BackendStaticInitializer backend_static_initializer;
 } // namespace cuda::CUDACachingAllocator
 } // namespace c10
-
-namespace c10::cuda {
-
-// uid_ is incremented when a user creates a MemPool,
-// for example: using graph_pool_handle() or c10::cuda::MemPool().
-//
-// uuid_ is incremented when CUDAGraph creates a MemPool
-// as a result of a user not providing a pool.
-//
-// MempoolId_t of {0, 0} is used to denote when no MemPool has been
-// passed to a function, either by user or CUDAGraphs. For example,
-// default value of MempoolId_t for capture_begin function is {0, 0}.
-// That's why uid_ and uuid_ start at 1.
-std::atomic<CaptureId_t> MemPool::uid_{1};
-std::atomic<CaptureId_t> MemPool::uuid_{1};
-
-MemPool::MemPool(
-    CUDACachingAllocator::CUDAAllocator* allocator,
-    bool is_user_created,
-    bool use_on_oom)
-    : allocator_(allocator), is_user_created_(is_user_created) {
-  if (is_user_created_) {
-    id_ = {0, uid_++};
-  } else {
-    id_ = {uuid_++, 0};
-  }
-  device_ = c10::cuda::current_device();
-  CUDACachingAllocator::createOrIncrefPool(device_, id_, allocator);
-  if (use_on_oom) {
-    CUDACachingAllocator::setUseOnOOM(device_, id_);
-  }
-}
-
-MemPool::~MemPool() {
-  TORCH_INTERNAL_ASSERT(use_count() == 1);
-  CUDACachingAllocator::releasePool(device_, id_);
-  c10::cuda::CUDACachingAllocator::emptyCache(id_);
-}
-
-MempoolId_t MemPool::id() {
-  return id_;
-}
-
-CUDACachingAllocator::CUDAAllocator* MemPool::allocator() {
-  return allocator_;
-}
-
-int MemPool::use_count() {
-  return CUDACachingAllocator::getPoolUseCount(device_, id_);
-}
-
-c10::DeviceIndex MemPool::device() {
-  return device_;
-}
-
-MempoolId_t MemPool::graph_pool_handle(bool is_user_created) {
-  if (is_user_created) {
-    return {0, uid_++};
-  }
-  return {uuid_++, 0};
-}
-
-} // namespace c10::cuda

c10/cuda/CUDACachingAllocator.h

@@ -562,41 +562,7 @@ inline std::string getUserMetadata() {
 } // namespace c10::cuda::CUDACachingAllocator
 
 namespace c10::cuda {
-
 // Keep BC only
 using c10::CaptureId_t;
 using c10::MempoolId_t;
-
-// MemPool represents a pool of memory in a caching allocator. Currently,
-// it's just the ID of the pool object maintained in the CUDACachingAllocator.
-//
-// An allocator pointer can be passed to the MemPool to define how the
-// allocations should be done in the pool. For example: using a different
-// system allocator such as ncclMemAlloc.
-struct C10_CUDA_API MemPool {
-  MemPool(
-      CUDACachingAllocator::CUDAAllocator* allocator = nullptr,
-      bool is_user_created = true,
-      bool use_on_oom = false);
-  MemPool(const MemPool&) = delete;
-  MemPool(MemPool&&) = default;
-  MemPool& operator=(const MemPool&) = delete;
-  MemPool& operator=(MemPool&&) = default;
-  ~MemPool();
-
-  MempoolId_t id();
-  CUDACachingAllocator::CUDAAllocator* allocator();
-  int use_count();
-  c10::DeviceIndex device();
-  static MempoolId_t graph_pool_handle(bool is_user_created = true);
-
- private:
-  static std::atomic<CaptureId_t> uid_;
-  static std::atomic<CaptureId_t> uuid_;
-  CUDACachingAllocator::CUDAAllocator* allocator_;
-  bool is_user_created_;
-  MempoolId_t id_;
-  c10::DeviceIndex device_;
-};
-
 } // namespace c10::cuda

c10/cuda/CUDADeviceAssertionHost.cpp

@@ -295,11 +295,19 @@ DeviceAssertionsData* CUDAKernelLaunchRegistry::
   C10_CUDA_CHECK_WO_DSA(
       cudaMallocManaged(&uvm_assertions_ptr, sizeof(DeviceAssertionsData)));
 
+#if CUDART_VERSION >= 13000
+  cudaMemLocation cpuDevice;
+  cpuDevice.type = cudaMemLocationTypeDevice;
+  cpuDevice.id = cudaCpuDeviceId;
+#else
+  const auto cpuDevice = cudaCpuDeviceId;
+#endif
+
   C10_CUDA_CHECK_WO_DSA(cudaMemAdvise(
       uvm_assertions_ptr,
       sizeof(DeviceAssertionsData),
       cudaMemAdviseSetPreferredLocation,
-      cudaCpuDeviceId));
+      cpuDevice));
 
   // GPU will establish direct mapping of data in CPU memory, no page faults
   // will be generated
@@ -307,7 +315,7 @@ DeviceAssertionsData* CUDAKernelLaunchRegistry::
       uvm_assertions_ptr,
       sizeof(DeviceAssertionsData),
       cudaMemAdviseSetAccessedBy,
-      cudaCpuDeviceId));
+      cpuDevice));
 
   // Initialize the memory from the CPU; otherwise, pages may have to be created
   // on demand. We think that UVM documentation indicates that first access may

c10/util/Exception.cpp

@@ -1,4 +1,5 @@
 #include <c10/util/Exception.h>
+#include <c10/util/FileSystem.h>
 #include <c10/util/Logging.h>
 #include <c10/util/Type.h>
 
@@ -27,7 +28,7 @@ Error::Error(
     const void* caller)
     : Error(
           str("[enforce fail at ",
-              detail::StripBasename(file),
+              c10::filesystem::path(file).filename(),
               ":",
               line,
               "] ",

c10/util/Exception.h

@@ -379,7 +379,11 @@ C10_API std::string GetExceptionString(const std::exception& e);
 // ----------------------------------------------------------------------------
 
 #ifdef STRIP_ERROR_MESSAGES
-#define TORCH_RETHROW(e, ...) throw
+#define TORCH_RETHROW(e, ...)                       \
+  do {                                              \
+    (void)e; /* Suppress unused variable warning */ \
+    throw;                                          \
+  } while (false)
 #else
 #define TORCH_RETHROW(e, ...)               \
   do {                                      \

c10/util/Logging.cpp

@@ -1,4 +1,5 @@
 #include <c10/util/Backtrace.h>
+#include <c10/util/FileSystem.h>
 #include <c10/util/Flags.h>
 #include <c10/util/Lazy.h>
 #include <c10/util/Logging.h>
@@ -478,8 +479,7 @@ MessageLogger::MessageLogger(
           << std::setfill('0') << " " << std::setw(2) << timeinfo->tm_hour
           << ":" << std::setw(2) << timeinfo->tm_min << ":" << std::setw(2)
           << timeinfo->tm_sec << "." << std::setw(9) << ns << " "
-          << c10::detail::StripBasename(std::string(file)) << ":" << line
-          << "] ";
+          << c10::filesystem::path(file).filename() << ":" << line << "] ";
 }
 
 // Output the contents of the stream to the proper channel on destruction.

c10/xpu/CMakeLists.txt

@@ -24,6 +24,7 @@ set(C10_XPU_HEADERS
     XPUCachingAllocator.h
     XPUDeviceProp.h
     XPUException.h
+    XPUEvent.h
     XPUFunctions.h
     XPUMacros.h
     XPUStream.h

c10/xpu/XPUEvent.h

@@ -0,0 +1,178 @@
+#pragma once
+#include <c10/xpu/XPUStream.h>
+
+namespace c10::xpu {
+
+/*
+ * XPUEvent are movable not copyable wrappers around SYCL event. XPUEvent are
+ * constructed lazily when first recorded. It has a device, and this device is
+ * acquired from the first recording stream. Later streams that record the event
+ * must match the same device.
+ *
+ * Currently, XPUEvent does NOT support to export an inter-process event from
+ * another process via inter-process communication(IPC). So it means that
+ * inter-process communication for event handles between different processes is
+ * not available. This could impact some applications that rely on cross-process
+ * synchronization and communication.
+ */
+struct XPUEvent {
+  // Constructors
+  XPUEvent(bool enable_timing = false) noexcept
+      : enable_timing_{enable_timing} {}
+
+  ~XPUEvent() {
+    if (isCreated()) {
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_deletion(
+            c10::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
+      }
+    }
+  }
+
+  C10_DISABLE_COPY_AND_ASSIGN(XPUEvent);
+
+  XPUEvent(XPUEvent&& other) = default;
+  XPUEvent& operator=(XPUEvent&& other) = default;
+
+  operator sycl::event&() const {
+    return event();
+  }
+
+  std::optional<c10::Device> device() const {
+    if (isCreated()) {
+      return c10::Device(c10::kXPU, device_index_);
+    } else {
+      return std::nullopt;
+    }
+  }
+
+  inline bool isCreated() const {
+    return (event_.get() != nullptr);
+  }
+
+  DeviceIndex device_index() const {
+    return device_index_;
+  }
+
+  sycl::event& event() const {
+    return *event_;
+  }
+
+  bool query() const {
+    using namespace sycl::info;
+    if (!isCreated()) {
+      return true;
+    }
+
+    return event().get_info<event::command_execution_status>() ==
+        event_command_status::complete;
+  }
+
+  void record() {
+    record(getCurrentXPUStream());
+  }
+
+  void recordOnce(const XPUStream& stream) {
+    if (!isCreated()) {
+      record(stream);
+    }
+  }
+
+  void record(const XPUStream& stream) {
+    if (!isCreated()) {
+      device_index_ = stream.device_index();
+      assignEvent(stream.queue());
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_creation(
+            c10::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
+      }
+    } else {
+      TORCH_CHECK(
+          device_index_ == stream.device_index(),
+          "Event device ",
+          device_index_,
+          " does not match recording stream's device ",
+          stream.device_index(),
+          ".");
+      reassignEvent(stream.queue());
+    }
+    const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+    if (C10_UNLIKELY(interp)) {
+      (*interp)->trace_gpu_event_record(
+          c10::kXPU,
+          reinterpret_cast<uintptr_t>(event_.get()),
+          reinterpret_cast<uintptr_t>(&stream.queue()));
+    }
+  }
+
+  void block(const XPUStream& stream) {
+    if (isCreated()) {
+      std::vector<sycl::event> event_list{event()};
+      // Make this stream wait until event_ is completed.
+      stream.queue().ext_oneapi_submit_barrier(event_list);
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_wait(
+            c10::kXPU,
+            reinterpret_cast<uintptr_t>(event_.get()),
+            reinterpret_cast<uintptr_t>(&stream.queue()));
+      }
+    }
+  }
+
+  double elapsed_time(const XPUEvent& other) const {
+    TORCH_CHECK(
+        isCreated() && other.isCreated(),
+        "Both events must be recorded before calculating elapsed time.");
+    TORCH_CHECK(
+        query() && other.query(),
+        "Both events must be completed before calculating elapsed time.");
+    TORCH_CHECK(
+        enable_timing_ && other.enable_timing_,
+        "Both events must be created with argument 'enable_timing=True'.");
+
+    using namespace sycl::info::event_profiling;
+    // Block until both of the recorded events are completed.
+    uint64_t end_time_ns = other.event().get_profiling_info<command_end>();
+    uint64_t start_time_ns = event().get_profiling_info<command_end>();
+    // Return the eplased time in milliseconds.
+    return 1e-6 *
+        (static_cast<double>(end_time_ns) - static_cast<double>(start_time_ns));
+  }
+
+  void synchronize() const {
+    if (isCreated()) {
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_synchronization(
+            c10::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
+      }
+      event().wait_and_throw();
+    }
+  }
+
+ private:
+  void assignEvent(sycl::queue& queue) {
+    if (enable_timing_) {
+      event_ = std::make_unique<sycl::event>(
+          sycl::ext::oneapi::experimental::submit_profiling_tag(queue));
+    } else {
+      event_ = std::make_unique<sycl::event>(queue.ext_oneapi_submit_barrier());
+    }
+  }
+
+  void reassignEvent(sycl::queue& queue) {
+    event_.reset();
+    assignEvent(queue);
+  }
+
+  bool enable_timing_ = false;
+  c10::DeviceIndex device_index_ = -1;
+  // Only need to track the last event, as events in an in-order queue are
+  // executed sequentially.
+  std::unique_ptr<sycl::event> event_;
+};
+
+} // namespace c10::xpu

caffe2/serialize/inline_container.cc

@@ -734,7 +734,7 @@ void PyTorchStreamWriter::setup(const string& file_name) {
           file_name,
           std::ofstream::out | std::ofstream::trunc | std::ofstream::binary
         );
-    } catch (const std::ios_base::failure& e) {
+    } catch (const std::ios_base::failure&) {
 #ifdef _WIN32
       // Windows have verbose error code, we prefer to use it than std errno.
       uint32_t error_code = GetLastError();

cmake/Codegen.cmake

@@ -118,6 +118,11 @@ if(INTERN_BUILD_ATEN_OPS)
             list(APPEND _file_compile_flags "-gencode;arch=compute_120a,code=sm_120a")
           endif()
         endif()
+        if("${_arch}" STREQUAL "121a")
+          if(_existing_arch_flags MATCHES ".*compute_120.*")
+            list(APPEND _file_compile_flags "-gencode;arch=compute_121a,code=sm_121a")
+          endif()
+        endif()
       endforeach()
       list(JOIN _file_compile_flags " " _file_compile_flags)
 
@@ -126,7 +131,7 @@ if(INTERN_BUILD_ATEN_OPS)
 
     _BUILD_FOR_ADDITIONAL_ARCHS(
       "${CMAKE_CURRENT_LIST_DIR}/../aten/src/ATen/native/cuda/RowwiseScaledMM.cu"
-      "89;90a;100a;103a;120a")
+      "89;90a;100a;103a;120a;121a")
     _BUILD_FOR_ADDITIONAL_ARCHS(
       "${CMAKE_CURRENT_LIST_DIR}/../aten/src/ATen/native/cuda/ScaledGroupMM.cu"
       "90a")

cmake/Modules/FindSYCLToolkit.cmake

@@ -1,7 +1,7 @@
 # This will define the following variables:
 # SYCL_FOUND               : True if the system has the SYCL library.
 # SYCL_INCLUDE_DIR         : Include directories needed to use SYCL.
-# SYCL_LIBRARY_DIR         ：The path to the SYCL library.
+# SYCL_LIBRARY_DIR         : The path to the SYCL library.
 # SYCL_LIBRARY             : SYCL library fullname.
 # SYCL_COMPILER_VERSION    : SYCL compiler version.
 

docs/cpp/source/index.rst

@@ -10,7 +10,7 @@ API.  This API can roughly be divided into five parts:
 - **TorchScript**: An interface to the TorchScript JIT compiler and interpreter.
 - **C++ Extensions**: A means of extending the Python API with custom C++ and CUDA routines.
 
-Combining, these building blocks form a research and
+Combined, these building blocks form a research and
 production ready C++ library for tensor computation and dynamic neural
 networks with strong emphasis on GPU acceleration as well as fast CPU
 performance. It is currently in use at Facebook in research and
@@ -76,7 +76,7 @@ C++ Frontend
 ------------
 
 The PyTorch C++ frontend provides a high level, pure C++ modeling interface for
-neural network and general ML(Machine Learning) research and production use cases,
+neural networks and general ML (Machine Learning) research and production use cases,
 largely following the Python API in design and provided functionality. The C++
 frontend includes the following:
 

docs/source/accelerator/device.md

@@ -0,0 +1,113 @@
+# Device Management
+
+## Background
+
+Device management handles basic operations like querying how many devices are available and switching between them. Accelerator backends need to wrap their device runtime's APIs and expose them to PyTorch.
+
+The OpenReg implementation ([`OpenRegFunctions.h/cpp`][OpenReg Device Management]) shows how to wrap a third-party runtime. These functions are used throughout the backend - by streams, events, generators, and Python bindings.
+
+## Design
+
+Accelerator vendors need to implement these core functions:
+
+| Function Name             | Description                                                      | Application Scenarios                                                                                          |
+| ------------------------- | ---------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- |
+| `device_count()`          | Query the total number of available devices in the system        | - Application initialization<br>- Multi-device workload distribution<br>- Validating device indices before use |
+| `current_device()`        | Get the currently active device for the calling thread           | - Debugging and logging<br>- Determining tensor placement<br>- Guard implementations                           |
+| `set_device()`            | Change the active device for subsequent operations               | - Switching context between devices<br>- Initializing specific device resources<br>- Multi-GPU training loops  |
+| `exchange_device()`       | Atomically swap device and return the previous device            | - Implementing device guards<br>- Temporarily switching device context<br>- RAII-based device management       |
+| `maybe_exchange_device()` | Conditionally exchange device only if the index is valid (-1 OK) | - Safe device switching with optional indices<br>- Guard implementations with nullable device values           |
+
+These functions are building blocks for more complex features like streams, events, and memory management. Make sure to validate inputs and handle errors properly.
+
+## Implementation
+
+This section shows how to implement device management using `set_device` as an example. The implementation requires:
+1. C++ wrappers around the device runtime
+2. Python bindings to expose the C++ functions
+3. User-friendly Python APIs
+
+### C++ Side
+
+Wrap the device runtime's API and add error handling. The `SetDevice` function shows this pattern:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegFunctions.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG SetDevice FUNCTION
+    :end-before: LITERALINCLUDE END: OPENREG SetDevice FUNCTION
+    :linenos:
+```
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegFunctions.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG set_device FUNCTION
+    :end-before: LITERALINCLUDE END: OPENREG set_device FUNCTION
+    :linenos:
+```
+
+### Binding
+
+Expose the C++ functions to Python using pybind11:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/torch_openreg/csrc/Module.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: MODULE SET DEVICE HELPER
+    :end-before: LITERALINCLUDE END: MODULE SET DEVICE HELPER
+    :linenos:
+```
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/torch_openreg/csrc/Module.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG MODULE METHODS
+    :end-before: LITERALINCLUDE END: OPENREG MODULE METHODS
+    :linenos:
+    :emphasize-lines: 5
+```
+
+### Python Side
+
+Wrap the C++ bindings with user-friendly Python functions:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/torch_openreg/openreg/__init__.py
+    :language: python
+    :start-after: LITERALINCLUDE START: PYTHON SET DEVICE FUNCTION
+    :end-before: LITERALINCLUDE END: PYTHON SET DEVICE FUNCTION
+    :linenos:
+```
+
+Here's the complete mapping from C++ to Python:
+
+| C++ Binding Function | C++ Binding API (pybind11)               | Python User API                  | Description                                  |
+| -------------------- | ---------------------------------------- | -------------------------------- | -------------------------------------------- |
+| `_getDeviceCount`    | `torch_openreg._C._get_device_count()`   | `torch.openreg.device_count()`   | Returns the total number of devices          |
+| `_getDevice`         | `torch_openreg._C._get_device()`         | `torch.openreg.current_device()` | Returns the current active device index      |
+| `_setDevice`         | `torch_openreg._C._set_device(idx)`      | `torch.openreg.set_device(idx)`  | Sets the active device                       |
+| `_exchangeDevice`    | `torch_openreg._C._exchange_device(idx)` | N/A (internal use only)          | Atomically swaps device and returns previous |
+
+## Guard
+
+Device guards provide automatic device switching with exception safety. They're similar to lock guards in C++ - they switch device on construction and restore it on destruction.
+
+Implement `DeviceGuardImplInterface` to integrate with PyTorch's guard system:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegGuard.h
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG DEVICE MGMT GUARD IMPL EXAMPLE
+    :end-before: LITERALINCLUDE END: OPENREG DEVICE MGMT GUARD IMPL EXAMPLE
+    :linenos:
+```
+
+**What needs to be implemented:**
+
+1. **exchangeDevice()**: Switch to a new device and return the old one (used by guard constructors)
+2. **getDevice()**: Get the current device
+3. **setDevice()**: Set the active device
+4. **Type checking**: Validate that device type matches the backend
+
+This makes the guard available to PyTorch for the `PrivateUse1` device type. Users can then use standard PyTorch device guards with the custom backend.
+
+[OpenReg Device Management]: https://github.com/pytorch/pytorch/blob/main/test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegFunctions.cpp "OpenReg Device Management"

docs/source/accelerator/hooks.md

@@ -0,0 +1,164 @@
+# Accelerator Hooks
+
+## Background
+
+OpenReg hooks provide a mechanism for integrating custom accelerator devices into PyTorch's runtime system. OpenReg (Open Registration) is PyTorch's extensibility framework that allows accelerator vendors to register custom device backends without modifying PyTorch core code.
+
+## Design
+
+The following tables list all hooks that accelerator vendors need to implement when integrating a new device backend. These hooks are categorized into two priority levels:
+
+- **High Priority Hooks**: Core APIs that PyTorch runtime directly depends on. Accelerator vendors are recommended to implement all high priority hooks to ensure full PyTorch compatibility and enable basic device functionality.
+
+- **Low Priority Hooks**: Device management and utility APIs that PyTorch does not directly depend on. These hooks enhance user experience and multi-device support but are *optional*. Accelerator vendors can choose to implement them based on their specific requirements and use cases.
+
+### High Priority Hooks
+
+| Hook Method                        | Description                                               | Application Scenario                                                             |
+| ---------------------------------- | --------------------------------------------------------- | -------------------------------------------------------------------------------- |
+| `init()`                           | Initializes the accelerator runtime and device contexts   | Set up necessary state when PyTorch first accesses the device                    |
+| `hasPrimaryContext(DeviceIndex)`   | Checks if a primary context exists for the device         | Determine whether device initialization has occurred                             |
+| `getDefaultGenerator(DeviceIndex)` | Returns the default random number generator for a device  | Access the device's primary RNG for reproducible random operations               |
+| `getNewGenerator(DeviceIndex)`     | Creates a new independent random number generator         | Create isolated RNG instances for parallel operations                            |
+| `getDeviceFromPtr(void*)`          | Determines which device a memory pointer belongs to       | Identify the accelerator device associated with a memory allocation              |
+| `getPinnedMemoryAllocator()`       | Returns an allocator for pinned (page-locked) host memory | Allocate host memory that can be efficiently transferred to/from the accelerator |
+| `isPinnedPtr(void*)`               | Checks if a pointer points to pinned memory               | Validate memory types before performing operations                               |
+
+### Low Priority Hooks
+
+| Hook Method                        | Description                                                                  | Application Scenario                                                 |
+| ---------------------------------- | ---------------------------------------------------------------------------- | -------------------------------------------------------------------- |
+| `isBuilt()`                        | Returns whether the accelerator backend is built/compiled into the extension | Check whether the accelerator library is available at compile time   |
+| `isAvailable()`                    | Returns whether the accelerator hardware is available at runtime             | Verify whether accelerator devices can be detected and initialized   |
+| `deviceCount()`                    | Returns the number of available accelerator devices                          | Enumerate all available accelerator devices for device selection     |
+| `setCurrentDevice(DeviceIndex)`    | Sets the active device for the current thread                                | Switch the current thread's context to a specific accelerator device |
+| `getCurrentDevice()`               | Returns the currently active device index                                    | Query which accelerator device is active in the current thread       |
+| `exchangeDevice(DeviceIndex)`      | Atomically exchanges the current device and returns the previous one         | Temporarily switch devices and restore the previous device afterward |
+| `maybeExchangeDevice(DeviceIndex)` | Conditionally exchanges device only if the index is valid                    | Safely attempt device switching with validation                      |
+
+## Implementation
+
+We can just take `getDefaultGenerator` as an implementation example:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegHooks.h
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG HOOK EXAMPLES
+    :end-before: LITERALINCLUDE END: OPENREG HOOK EXAMPLES
+    :linenos:
+```
+
+In this implementation:
+
+1. **Override the base interface**: The `getDefaultGenerator` method overrides the virtual method from `at::PrivateUse1HooksInterface`.
+
+2. **Delegate to device-specific implementation**: It calls `getDefaultOpenRegGenerator(device_index)`, which manages a per-device generator instance.
+
+3. **Return device-specific generator**: The returned `at::Generator` wraps an `OpenRegGeneratorImpl` that implements device-specific random number generation.
+
+This pattern applies to all hooks: override the interface method, validate inputs, delegate to your device-specific API, and return results in PyTorch's expected format.
+
+## Integration Example
+
+The following sections demonstrate how PyTorch integrates with accelerator hooks when accessing the default random number generator. The example traces the complete flow from user-facing Python code down to the device-specific implementation.
+
+### Layer 1: User Code
+
+User code initiates the operation by calling `manual_seed` to set the random seed for reproducible results:
+
+```python
+import torch
+torch.openreg.manual_seed(42)
+```
+
+### Layer 2: Extension Python API
+
+The Python API layer handles device management and calls into the C++ extension (defined in [`torch_openreg/openreg/random.py`][random.py]):
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/torch_openreg/openreg/random.py
+    :language: python
+    :start-after: LITERALINCLUDE START: OPENREG MANUAL SEED
+    :end-before: LITERALINCLUDE END: OPENREG MANUAL SEED
+    :linenos:
+```
+
+The `manual_seed` function gets the current device index and calls `torch_openreg._C._get_default_generator(idx)` to obtain the device-specific generator, then sets the seed on it.
+
+### Layer 3: Python/C++ Bridge
+
+The C++ extension exposes `_getDefaultGenerator` to Python, which bridges to PyTorch's core runtime:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/torch_openreg/csrc/Module.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG GET DEFAULT GENERATOR
+    :end-before: LITERALINCLUDE END: OPENREG GET DEFAULT GENERATOR
+    :linenos:
+    :emphasize-lines: 10-11
+```
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/torch_openreg/csrc/Module.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG MODULE METHODS
+    :end-before: LITERALINCLUDE END: OPENREG MODULE METHODS
+    :linenos:
+    :emphasize-lines: 3
+```
+
+This function unpacks the device index from Python, creates a `PrivateUse1` device object, and calls `at::globalContext().defaultGenerator()`. PyTorch's context then dispatches to the registered hooks.
+
+### Layer 4: PyTorch Core Context
+
+PyTorch's Context class dispatches to the appropriate accelerator hooks ([`aten/src/ATen/Context.h`][Context.h]):
+
+```{eval-rst}
+.. literalinclude:: ../../../aten/src/ATen/Context.h
+    :language: c++
+    :lines: 60-103
+    :linenos:
+    :emphasize-lines: 8-9, 24-25
+```
+
+This layered architecture enables PyTorch to remain device-agnostic while delegating hardware-specific operations to accelerator implementations. The hooks are registered once at module load time:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegHooks.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG HOOK REGISTER
+    :end-before: LITERALINCLUDE END: OPENREG HOOK REGISTER
+    :linenos:
+    :emphasize-lines: 4
+```
+
+### Layer 5: Accelerator Hooks
+
+The hooks interface provides the abstraction that PyTorch uses to delegate to device-specific implementations:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegHooks.h
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG HOOK EXAMPLES
+    :end-before: LITERALINCLUDE END: OPENREG HOOK EXAMPLES
+    :linenos:
+```
+
+The `getDefaultGenerator` hook method overrides the base interface and delegates to `getDefaultOpenRegGenerator`, which manages the actual generator instances.
+
+### Layer 6: Device-Specific Implementation
+
+The device-specific implementation manages per-device generator instances:
+
+```{eval-rst}
+.. literalinclude:: ../../../test/cpp_extensions/open_registration_extension/torch_openreg/csrc/runtime/OpenRegGenerator.cpp
+    :language: c++
+    :start-after: LITERALINCLUDE START: OPENREG GET DEFAULT GENERATOR IMPL
+    :end-before: LITERALINCLUDE END: OPENREG GET DEFAULT GENERATOR IMPL
+    :linenos:
+```
+
+This function maintains a static vector of generators (one per device), initializes them on first access, validates the device index, and returns the appropriate generator instance.
+
+[random.py]: https://github.com/pytorch/pytorch/tree/main/test/cpp_extensions/open_registration_extension/torch_openreg/torch_openreg/openreg/random.py#L48-L53 "random.py"
+[Context.h]: https://github.com/pytorch/pytorch/tree/main/aten/src/ATen/Context.h#L61-L102 "Context.h"

docs/source/accelerator/index.md

@@ -42,6 +42,8 @@ Next, we will delve into each chapter of this guide. Each chapter focuses on a k
 :glob:
 :maxdepth: 1
 
+device
+hooks
 autoload
 operators
 amp

docs/source/nn.attention.rst

@@ -14,6 +14,10 @@ Utils
 
     sdpa_kernel
     SDPBackend
+    register_flash_attention_impl
+    activate_flash_attention_impl
+    list_flash_attention_impls
+    current_flash_attention_impl
 
 Submodules
 ----------

docs/source/notes/cuda.rst

@@ -254,7 +254,7 @@ To toggle the reduced precision reduction flags in C++, one can do
 
 .. _fp16accumulation:
 
-Full FP16 Accmumulation in FP16 GEMMs
+Full FP16 Accumulation in FP16 GEMMs
 -------------------------------------
 
 Certain GPUs have increased performance when doing _all_ FP16 GEMM accumulation

docs/source/torch.compiler_api.md

@@ -30,5 +30,6 @@ For a quick overview of `torch.compiler`, see {ref}`torch.compiler_overview`.
      skip_guard_on_all_nn_modules_unsafe
      keep_tensor_guards_unsafe
      skip_guard_on_globals_unsafe
+     skip_all_guards_unsafe
      nested_compile_region
 ```

functorch/op_analysis/gen_data.py

@@ -24,15 +24,11 @@ def gen_data(special_op_lists, analysis_name):
     all_ops = get_ops_for_key(None)
     composite_ops = get_ops_for_key("CompositeImplicitAutograd")
     noncomposite_ops = all_ops - composite_ops
+    with open("../../aten/src/ATen/native/native_functions.yaml") as f:
+        ops = yaml.load(f.read(), Loader=yaml.CLoader)
 
-    ops = yaml.load(
-        open("../../aten/src/ATen/native/native_functions.yaml").read(),
-        Loader=yaml.CLoader,
-    )
-
-    annotated_ops = {
-        a.strip(): b.strip() for a, b in list(csv.reader(open("annotated_ops")))
-    }
+    with open("annotated_ops") as f:
+        annotated_ops = {a.strip(): b.strip() for a, b in csv.reader(f)}
 
     uniq_ops = []
     uniq_names = set()

pyproject.toml

@@ -376,3 +376,19 @@ keep-runtime-typing = true
 
 [tool.codespell]
 ignore-words = "tools/linter/dictionary.txt"
+
+[tool.spin]
+package = 'torch'
+
+[tool.spin.commands]
+"Build" = [
+  ".spin/cmds.py:lint",
+  ".spin/cmds.py:fixlint",
+  ".spin/cmds.py:quicklint",
+  ".spin/cmds.py:quickfix",
+]
+"Regenerate" = [
+  ".spin/cmds.py:regenerate_version",
+  ".spin/cmds.py:regenerate_type_stubs",
+  ".spin/cmds.py:regenerate_clangtidy_files",
+]

pyrefly.toml

@@ -32,7 +32,7 @@ project-excludes = [
   "torch/utils/tensorboard/summary.py",
   # formatting issues, will turn on after adjusting where suppressions can be
   # in import statements
-  "tools/flight_recorder/components/types.py",
+  "torch/distributed/flight_recorder/components/types.py",
   "torch/linalg/__init__.py",
   "torch/package/importer.py",
   "torch/package/_package_pickler.py",

requirements.txt

@@ -14,6 +14,7 @@ lintrunner ; platform_machine != "s390x" and platform_machine != "riscv64"
 networkx>=2.5.1
 optree>=0.13.0
 psutil
+spin
 sympy>=1.13.3
 typing-extensions>=4.13.2
 wheel

scripts/onnx/install-develop.sh

@@ -10,7 +10,7 @@ tp2_dir="$top_dir/third_party"
 pip install ninja
 
 # Install onnx
-pip install --no-use-pep517 -e "$tp2_dir/onnx"
+pip install -e "$tp2_dir/onnx"
 
 # Install caffe2 and pytorch
 pip install -r "$top_dir/caffe2/requirements.txt"

setup.py

@@ -1585,7 +1585,7 @@ def configure_extension_build() -> tuple[
     if cmake_cache_vars["USE_DISTRIBUTED"]:
         # Only enable fr_trace command if distributed is enabled
         entry_points["console_scripts"].append(
-            "torchfrtrace = tools.flight_recorder.fr_trace:main",
+            "torchfrtrace = torch.distributed.flight_recorder.fr_trace:main",
         )
     return ext_modules, cmdclass, packages, entry_points, extra_install_requires
 

test/conftest.py

@@ -308,12 +308,16 @@ class StepcurrentPlugin:
         self.report_status = ""
         assert config.cache is not None
         self.cache: pytest.Cache = config.cache
-        self.directory = f"{STEPCURRENT_CACHE_DIR}/{config.getoption('stepcurrent')}"
-        self.lastrun: Optional[str] = self.cache.get(self.directory, None)
+        directory = f"{STEPCURRENT_CACHE_DIR}/{config.getoption('stepcurrent')}"
+        self.lastrun_location = f"{directory}/lastrun"
+        self.lastrun: Optional[str] = self.cache.get(self.lastrun_location, None)
         self.initial_val = self.lastrun
         self.skip: bool = config.getoption("stepcurrent_skip")
         self.run_single: bool = config.getoption("run_single")
 
+        self.made_failing_xml_location = f"{directory}/made_failing_xml"
+        self.cache.set(self.made_failing_xml_location, False)
+
     def pytest_collection_modifyitems(self, config: Config, items: list[Any]) -> None:
         if not self.lastrun:
             self.report_status = "Cannot find last run test, not skipping"
@@ -349,8 +353,10 @@ class StepcurrentPlugin:
 
     def pytest_runtest_protocol(self, item, nextitem) -> None:
         self.lastrun = item.nodeid
-        self.cache.set(self.directory, self.lastrun)
+        self.cache.set(self.lastrun_location, self.lastrun)
 
     def pytest_sessionfinish(self, session, exitstatus):
         if exitstatus == 0:
-            self.cache.set(self.directory, self.initial_val)
+            self.cache.set(self.lastrun_location, self.initial_val)
+        if exitstatus != 0:
+            self.cache.set(self.made_failing_xml_location, True)

test/cpp/aoti_abi_check/test_scalartype.cpp

@@ -13,6 +13,17 @@ TEST(TestScalarType, ScalarTypeToCPPTypeT) {
 #undef DEFINE_CHECK
 }
 
+TEST(TestScalarType, CppTypeToScalarType) {
+  using torch::headeronly::CppTypeToScalarType;
+  using torch::headeronly::ScalarType;
+
+#define DEFINE_CHECK(TYPE, SCALARTYPE) \
+  EXPECT_EQ(CppTypeToScalarType<TYPE>::value, ScalarType::SCALARTYPE);
+
+  AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_CHECK);
+#undef DEFINE_CHECK
+}
+
 #define DEFINE_CHECK(TYPE, SCALARTYPE)                                       \
   {                                                                          \
     EXPECT_EQ(                                                               \

test/cpp_extensions/libtorch_agnostic_extension/libtorch_agnostic/csrc/kernel.cpp

@@ -38,7 +38,7 @@ using torch::stable::Tensor;
 Tensor sgd_out_of_place(
     const Tensor param,
     const Tensor grad,
-    const float weight_decay,
+    const double weight_decay,
     const double lr,
     const bool maximize) {
   STD_TORCH_CHECK(param.dim() == 1, "param must be 1D");
@@ -57,7 +57,7 @@ Tensor sgd_out_of_place(
     reinterpret_cast<float*>(param.data_ptr()),
     reinterpret_cast<float*>(grad.data_ptr()),
     reinterpret_cast<float*>(out.data_ptr()),
-    weight_decay,
+    float(weight_decay),
     lr,
     maximize,
     param.numel()
@@ -66,44 +66,29 @@ Tensor sgd_out_of_place(
   return out;
 }
 
-void boxed_sgd_out_of_place(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = sgd_out_of_place(
-    torch::stable::detail::to<Tensor>(stack[0]),
-    torch::stable::detail::to<Tensor>(stack[1]),
-    float(torch::stable::detail::to<double>(stack[2])),
-    torch::stable::detail::to<double>(stack[3]),
-    torch::stable::detail::to<bool>(stack[4]));
-
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY(libtorch_agnostic, m) {
   m.def("sgd_out_of_place(Tensor param, Tensor grad, float weight_decay, float lr, bool maximize) -> Tensor");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CPU, m) {
-  m.impl("sgd_out_of_place", &boxed_sgd_out_of_place);
+  m.impl("sgd_out_of_place", TORCH_BOX(&sgd_out_of_place));
 }
 
 Tensor identity(Tensor t) {
   return t;
 }
 
-void boxed_identity(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = identity(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("identity(Tensor t) -> Tensor");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CUDA, m) {
-  m.impl("identity", &boxed_identity);
+  m.impl("identity", TORCH_BOX(&identity));
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CPU, m) {
-  m.impl("identity", &boxed_identity);
+  m.impl("identity", TORCH_BOX(&identity));
 }
 
 Tensor my_abs(Tensor t) {
@@ -114,17 +99,12 @@ Tensor my_abs(Tensor t) {
   return torch::stable::detail::to<Tensor>(stack[0]);
 }
 
-void boxed_my_abs(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor tensor_res = my_abs(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(tensor_res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("my_abs(Tensor t) -> Tensor");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("my_abs", &boxed_my_abs);
+  m.impl("my_abs", TORCH_BOX(&my_abs));
 }
 
 Tensor my_ones_like(Tensor t, StableIValue device) {
@@ -145,17 +125,12 @@ Tensor my_ones_like(Tensor t, StableIValue device) {
   return torch::stable::detail::to<Tensor>(stack[0]);
 }
 
-void boxed_my_ones_like(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = my_ones_like(torch::stable::detail::to<Tensor>(stack[0]), stack[1]);
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("my_ones_like(Tensor t, Device d) -> Tensor");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("my_ones_like", &boxed_my_ones_like);
+  m.impl("my_ones_like", TORCH_BOX(&my_ones_like));
 }
 
 std::tuple<Tensor, Tensor, bool> exp_neg_is_leaf(Tensor t1, Tensor t2, Tensor t3) {
@@ -177,19 +152,12 @@ std::tuple<Tensor, Tensor, bool> exp_neg_is_leaf(Tensor t1, Tensor t2, Tensor t3
     torch::stable::detail::to<bool>(stack_is_leaf[0]));
 }
 
-void boxed_exp_neg_is_leaf(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto tuple = exp_neg_is_leaf(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<Tensor>(stack[1]), torch::stable::detail::to<Tensor>(stack[2]));
-  stack[0] = torch::stable::detail::from(std::get<0>(tuple));
-  stack[1] = torch::stable::detail::from(std::get<1>(tuple));
-  stack[2] = torch::stable::detail::from(std::get<2>(tuple));
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("exp_neg_is_leaf(Tensor t1, Tensor t2, Tensor t3) -> (Tensor, Tensor, bool)");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("exp_neg_is_leaf", &boxed_exp_neg_is_leaf);
+  m.impl("exp_neg_is_leaf", TORCH_BOX(&exp_neg_is_leaf));
 }
 
 Tensor neg_exp(Tensor t) {
@@ -200,17 +168,12 @@ Tensor neg_exp(Tensor t) {
   return torch::stable::detail::to<Tensor>(stack[0]);
 }
 
-void boxed_neg_exp(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = neg_exp(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("neg_exp(Tensor t) -> Tensor");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("neg_exp", &boxed_neg_exp);
+  m.impl("neg_exp", TORCH_BOX(&neg_exp));
 }
 
 Tensor divide_neg_exp(Tensor t) {
@@ -229,108 +192,53 @@ Tensor divide_neg_exp(Tensor t) {
   return torch::stable::detail::to<Tensor>(stack_div[0]);
 }
 
-void boxed_divide_neg_exp(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = divide_neg_exp(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("divide_neg_exp(Tensor t) -> Tensor");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("divide_neg_exp", &boxed_divide_neg_exp);
+  m.impl("divide_neg_exp", TORCH_BOX(&divide_neg_exp));
 }
 
 bool is_contiguous(Tensor t) {
   return t.is_contiguous();
 }
 
-void boxed_is_contiguous(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  bool res = is_contiguous(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("is_contiguous(Tensor t) -> bool");
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("is_contiguous", &boxed_is_contiguous);
+  m.impl("is_contiguous", TORCH_BOX(&is_contiguous));
 }
 
 Tensor my_transpose(Tensor t, int64_t dim0, int64_t dim1) {
   return transpose(t, dim0, dim1);
 }
 
-void boxed_my_transpose(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_transpose(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<int64_t>(stack[1]), torch::stable::detail::to<int64_t>(stack[2]));
-
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_empty_like(Tensor t) {
   return empty_like(t);
 }
 
-void boxed_empty_like(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_empty_like(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 bool my_is_cpu(Tensor t) {
   return t.is_cpu();
 }
 
-
-void boxed_my_is_cpu(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_is_cpu(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor fill_infinity(Tensor t) {
   auto value = std::numeric_limits<float>::infinity();
   return fill_(t, value);
 }
 
-void boxed_fill_infinity(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  auto res = fill_infinity(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_pad(Tensor t) {
   std::string mode = "constant";
   double value = 0.0;
   return pad(t, {1, 2, 2, 1}, mode, value);
 }
 
-void boxed_my_pad(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  auto res = my_pad(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_narrow(Tensor t, int64_t dim, int64_t start, int64_t length) {
   return narrow(t, dim, start, length);
 }
 
-void boxed_my_narrow(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  auto res = my_narrow(
-      torch::stable::detail::to<Tensor>(stack[0]),
-      torch::stable::detail::to<int64_t>(stack[1]),
-      torch::stable::detail::to<int64_t>(stack[2]),
-      torch::stable::detail::to<int64_t>(stack[3]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_new_empty_dtype_variant(Tensor t) {
   // Still using a std::vector below even though people can just pass in an
   // initializer list (which will be implicitly converted to an HeaderOnlyArrayRef)
@@ -342,40 +250,19 @@ Tensor my_new_empty_dtype_variant(Tensor t) {
   return new_empty(t, sizes, dtype);
 }
 
-void boxed_my_new_empty_dtype_variant(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_new_empty_dtype_variant(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_new_zeros_dtype_variant(Tensor t) {
   auto dtype = std::make_optional(at::ScalarType::Float);
   return new_zeros(t, {2, 5}, dtype);
 }
 
-void boxed_my_new_zeros_dtype_variant(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_new_zeros_dtype_variant(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_copy_(Tensor dst, Tensor src, bool non_blocking) {
   return copy_(dst, src, non_blocking);
 }
 
-void boxed_my_copy_(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor tensor_res = my_copy_(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<Tensor>(stack[1]), torch::stable::detail::to<bool>(stack[2]));
-  stack[0] = torch::stable::detail::from(tensor_res);
-}
-
 Tensor my_clone(Tensor t) {
   return clone(t);
 }
 
-void boxed_my_clone(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor tensor_res = my_clone(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(tensor_res);
-}
-
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("my_transpose(Tensor t, int dim0, int dim1) -> Tensor");
   m.def("my_empty_like(Tensor t) -> Tensor");
@@ -389,57 +276,39 @@ STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("my_transpose", &boxed_my_transpose);
-  m.impl("my_empty_like", &boxed_empty_like);
-  m.impl("fill_infinity", &boxed_fill_infinity);
-  m.impl("my_is_cpu", &boxed_my_is_cpu);
-  m.impl("my_new_empty_dtype_variant", &boxed_my_new_empty_dtype_variant);
-  m.impl("my_new_zeros_dtype_variant", &boxed_my_new_zeros_dtype_variant);
-  m.impl("my_copy_", &boxed_my_copy_);
-  m.impl("my_clone", &boxed_my_clone);
+  m.impl("my_transpose", TORCH_BOX(&my_transpose));
+  m.impl("my_empty_like", TORCH_BOX(&my_empty_like));
+  m.impl("fill_infinity", TORCH_BOX(&fill_infinity));
+  m.impl("my_is_cpu", TORCH_BOX(&my_is_cpu));
+  m.impl("my_new_empty_dtype_variant", TORCH_BOX(&my_new_empty_dtype_variant));
+  m.impl("my_new_zeros_dtype_variant", TORCH_BOX(&my_new_zeros_dtype_variant));
+  m.impl("my_copy_", TORCH_BOX(&my_copy_));
+  m.impl("my_clone", TORCH_BOX(&my_clone));
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeImplicitAutograd, m) {
-  m.impl("my_pad", &boxed_my_pad);
-  m.impl("my_narrow", &boxed_my_narrow);
+  m.impl("my_pad", TORCH_BOX(&my_pad));
+  m.impl("my_narrow", TORCH_BOX(&my_narrow));
 }
 
 Tensor my_zero_(Tensor t) {
   return zero_(t);
 }
 
-void boxed_my_zero_(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_zero_(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_amax(Tensor t) {
   return amax(t, 0, false);
 }
 
-void boxed_my_amax(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_amax(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 Tensor my_amax_vec(Tensor t) {
   return amax(t, {0,1}, false);
 }
 
-void boxed_my_amax_vec(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_amax_vec(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("my_zero_(Tensor(a!) t) -> Tensor(a!)");
   m.def("my_amax(Tensor a) -> Tensor");
   m.def("my_amax_vec(Tensor a) -> Tensor");
   m.def("my_is_cpu(Tensor t) -> bool");
-}
-
-STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CPU, m) {
-  m.impl("my_zero_", &boxed_my_zero_);
+   m.def("test_default_constructor(bool undefined) -> bool");
 }
 
 bool test_default_constructor(bool defined) {
@@ -461,22 +330,12 @@ bool test_default_constructor(bool defined) {
   return out.defined();
 }
 
-void boxed_test_default_constructor(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  bool res = test_default_constructor(torch::stable::detail::to<bool>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
-STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
-  m.def("test_default_constructor(bool undefined) -> bool");
-}
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("test_default_constructor", &boxed_test_default_constructor);
-  m.impl("my_amax", &boxed_my_amax);
-  m.impl("my_amax_vec", &boxed_my_amax_vec);
+  m.impl("my_zero_", TORCH_BOX(&my_zero_));
+  m.impl("my_amax", TORCH_BOX(&my_amax));
+  m.impl("my_amax_vec", TORCH_BOX(&my_amax_vec));
+  m.impl("test_default_constructor", TORCH_BOX(&test_default_constructor));
 }
 
 std::vector<Tensor> my__foreach_mul(torch::headeronly::HeaderOnlyArrayRef<Tensor> self, torch::headeronly::HeaderOnlyArrayRef<Tensor> other) {
@@ -485,23 +344,11 @@ std::vector<Tensor> my__foreach_mul(torch::headeronly::HeaderOnlyArrayRef<Tensor
   return torch::stable::detail::to<std::vector<Tensor>>(stack[0]);
 }
 
-void boxed_my__foreach_mul(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  // Why is the following NOT torch::stable::detail::to<HeaderOnlyArrayRef<Tensor>>(stack[0])? Because calling `to`
-  // on a StableIValue means that the result is owning its underlying data now! HeaderOnlyArrayRef
-  // is not owning, so it cannot safely steward the result of the torch::stable::detail::to<>.
-  auto res = my__foreach_mul(torch::stable::detail::to<std::vector<Tensor>>(stack[0]), torch::stable::detail::to<std::vector<Tensor>>(stack[1]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 void my__foreach_mul_(torch::headeronly::HeaderOnlyArrayRef<Tensor> self, torch::headeronly::HeaderOnlyArrayRef<Tensor> other) {
   std::array<StableIValue, 2> stack = {torch::stable::detail::from(self), torch::stable::detail::from(other)};
   aoti_torch_call_dispatcher("aten::_foreach_mul_", "List", stack.data());
 }
 
-void boxed_my__foreach_mul_(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  my__foreach_mul_(torch::stable::detail::to<std::vector<Tensor>>(stack[0]), torch::stable::detail::to<std::vector<Tensor>>(stack[1]));
-}
-
 std::vector<Tensor> make_tensor_clones_and_call_foreach(Tensor t1, Tensor t2) {
   // This function tests that my__foreach_mul can take in std::initializer_lists
   // in addition to std::vectors.
@@ -512,11 +359,6 @@ std::vector<Tensor> make_tensor_clones_and_call_foreach(Tensor t1, Tensor t2) {
   return my__foreach_mul({t1_1, t2_1}, {t1_2, t2_2});
 }
 
-void boxed_make_tensor_clones_and_call_foreach(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = make_tensor_clones_and_call_foreach(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<Tensor>(stack[1]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("my__foreach_mul(Tensor[] self, Tensor[] other) -> Tensor[]");
   m.def("my__foreach_mul_(Tensor(a!)[] self, Tensor[] other) -> ()");
@@ -524,9 +366,9 @@ STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("my__foreach_mul", &boxed_my__foreach_mul);
-  m.impl("my__foreach_mul_", &boxed_my__foreach_mul_);
-  m.impl("make_tensor_clones_and_call_foreach", &boxed_make_tensor_clones_and_call_foreach);
+  m.impl("my__foreach_mul", TORCH_BOX(&my__foreach_mul));
+  m.impl("my__foreach_mul_", TORCH_BOX(&my__foreach_mul_));
+  m.impl("make_tensor_clones_and_call_foreach", TORCH_BOX(&make_tensor_clones_and_call_foreach));
 }
 
 // Test functions for torch::stable::Tensor device method
@@ -690,14 +532,6 @@ int64_t test_device_guard(int64_t device_index) {
   return currentDevice;
 }
 
-void boxed_test_device_guard(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  int res = test_device_guard(static_cast<int64_t>(torch::stable::detail::to<int64_t>(stack[0])));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 int64_t test_device_guard_set_index() {
   using torch::stable::accelerator::DeviceGuard;
 
@@ -709,14 +543,6 @@ int64_t test_device_guard_set_index() {
   return currentDevice;
 }
 
-void boxed_test_device_guard_set_index(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  int64_t res = test_device_guard_set_index();
-  stack[0] = torch::stable::detail::from(res);
-}
-
 int64_t test_stream(int32_t device_index) {
   STD_TORCH_CHECK(
       device_index >= std::numeric_limits<int32_t>::min() &&
@@ -726,26 +552,10 @@ int64_t test_stream(int32_t device_index) {
   return torch::stable::accelerator::getCurrentStream(device_index).id();
 }
 
-void boxed_test_stream(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  int64_t res = test_stream(static_cast<int64_t>(torch::stable::detail::to<int64_t>(stack[0])));
-  stack[0] = torch::stable::detail::from(res);
-}
-
 int64_t test_get_current_device_index() {
   return torch::stable::accelerator::getCurrentDeviceIndex();
 }
 
-void boxed_test_get_current_device_index(
-    StableIValue* stack,
-    uint64_t num_args,
-    uint64_t num_outputs) {
-  int64_t res = test_get_current_device_index();
-  stack[0] = torch::stable::detail::from(res);
-}
-
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
   m.def("test_device_guard(int device_index) -> int");
   m.def("test_device_guard_set_index() -> int");
@@ -754,10 +564,10 @@ STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
 }
 
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("test_device_guard", &boxed_test_device_guard);
-  m.impl("test_device_guard_set_index", &boxed_test_device_guard_set_index);
-  m.impl("test_stream", &boxed_test_stream);
-  m.impl("test_get_current_device_index", &boxed_test_get_current_device_index);
+  m.impl("test_device_guard", TORCH_BOX(&test_device_guard));
+  m.impl("test_device_guard_set_index", TORCH_BOX(&test_device_guard_set_index));
+  m.impl("test_stream", TORCH_BOX(&test_stream));
+  m.impl("test_get_current_device_index", TORCH_BOX(&test_get_current_device_index));
 }
 
 #endif // LAE_USE_CUDA
@@ -824,3 +634,38 @@ STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
   m.impl("test_parallel_for", &boxed_test_parallel_for);
   m.impl("test_get_num_threads", &boxed_test_get_num_threads);
 }
+
+Tensor my_empty(
+    torch::headeronly::HeaderOnlyArrayRef<int64_t> size,
+    std::optional<torch::headeronly::ScalarType> dtype,
+    std::optional<torch::stable::Device> device,
+    std::optional<bool> pin_memory) {
+  return empty(size, dtype, device, pin_memory);
+}
+
+Tensor my_flatten(Tensor t, int64_t start_dim, int64_t end_dim) {
+  return flatten(t, start_dim, end_dim);
+}
+
+Tensor my_reshape(Tensor t, torch::headeronly::HeaderOnlyArrayRef<int64_t> shape) {
+  return reshape(t, shape);
+}
+
+Tensor my_view(Tensor t, torch::headeronly::HeaderOnlyArrayRef<int64_t> size) {
+  return view(t, size);
+}
+
+STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
+  m.def(
+      "my_empty(int[] size, ScalarType? dtype=None, Device? device=None, bool? pin_memory=None) -> Tensor");
+  m.def("my_flatten(Tensor t, int start_dim=0, int end_dim=-1) -> Tensor");
+  m.def("my_reshape(Tensor t, int[] shape) -> Tensor");
+  m.def("my_view(Tensor t, int[] size) -> Tensor");
+}
+
+STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
+  m.impl("my_empty", TORCH_BOX(&my_empty));
+  m.impl("my_flatten", TORCH_BOX(&my_flatten));
+  m.impl("my_reshape", TORCH_BOX(&my_reshape));
+  m.impl("my_view", TORCH_BOX(&my_view));
+}

test/cpp_extensions/libtorch_agnostic_extension/libtorch_agnostic/ops.py

@@ -487,3 +487,58 @@ def test_get_num_threads() -> int:
     Returns: int - the number of threads for the parallel backend
     """
     return torch.ops.libtorch_agnostic.test_get_num_threads.default()
+
+
+def my_empty(size, dtype=None, device=None, pin_memory=None) -> Tensor:
+    """
+    Creates an empty tensor with the specified size, dtype, device, and pin_memory.
+
+    Args:
+        size: list[int] - size of the tensor to create
+        dtype: ScalarType or None - data type of the tensor
+        device: Device or None - device on which to create the tensor
+        pin_memory: bool or None - whether to use pinned memory
+
+    Returns: Tensor - an uninitialized tensor with the specified properties
+    """
+    return torch.ops.libtorch_agnostic.my_empty.default(size, dtype, device, pin_memory)
+
+
+def my_flatten(t, start_dim=0, end_dim=-1) -> Tensor:
+    """
+    Flattens the input tensor from start_dim to end_dim into a single dimension.
+
+    Args:
+        t: Tensor - tensor to flatten
+        start_dim: int - first dimension to flatten (default: 0)
+        end_dim: int - last dimension to flatten (default: -1)
+
+    Returns: Tensor - flattened tensor
+    """
+    return torch.ops.libtorch_agnostic.my_flatten.default(t, start_dim, end_dim)
+
+
+def my_reshape(t, shape) -> Tensor:
+    """
+    Returns a tensor with the same data but different shape.
+
+    Args:
+        t: Tensor - tensor to reshape
+        shape: list[int] - new shape for the tensor
+
+    Returns: Tensor - reshaped tensor
+    """
+    return torch.ops.libtorch_agnostic.my_reshape.default(t, shape)
+
+
+def my_view(t, size) -> Tensor:
+    """
+    Returns a new tensor with the same data as the input tensor but of a different shape.
+
+    Args:
+        t: Tensor - tensor to view
+        size: list[int] - new size for the tensor
+
+    Returns: Tensor - tensor with new view
+    """
+    return torch.ops.libtorch_agnostic.my_view.default(t, size)