Run without sccache

Signed-off-by: Thanh Ha <thanh.ha@linuxfoundation.org>

.ci/aarch64_linux/README.md

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

.ci/aarch64_linux/aarch64_ci_build.sh

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

.ci/aarch64_linux/aarch64_ci_setup.sh

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

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -0,0 +1,333 @@
+#!/usr/bin/env python3
+# encoding: UTF-8
+
+import os
+import shutil
+from subprocess import check_call, check_output
+
+
+def list_dir(path: str) -> list[str]:
+    """'
+    Helper for getting paths for Python
+    """
+    return check_output(["ls", "-1", path]).decode().split("\n")
+
+
+def replace_tag(filename) -> None:
+    with open(filename) as f:
+        lines = f.readlines()
+    for i, line in enumerate(lines):
+        if line.startswith("Tag:"):
+            lines[i] = line.replace("-linux_", "-manylinux_2_28_")
+            print(f"Updated tag from {line} to {lines[i]}")
+            break
+
+    with open(filename, "w") as f:
+        f.writelines(lines)
+
+
+def patch_library_rpath(
+    folder: str,
+    lib_name: str,
+    use_nvidia_pypi_libs: bool = False,
+    desired_cuda: str = "",
+) -> None:
+    """Apply patchelf to set RPATH for a library in torch/lib"""
+    lib_path = f"{folder}/tmp/torch/lib/{lib_name}"
+
+    if use_nvidia_pypi_libs:
+        # For PyPI NVIDIA libraries, construct CUDA RPATH
+        cuda_rpaths = [
+            "$ORIGIN/../../nvidia/cudnn/lib",
+            "$ORIGIN/../../nvidia/nvshmem/lib",
+            "$ORIGIN/../../nvidia/nccl/lib",
+            "$ORIGIN/../../nvidia/cusparselt/lib",
+        ]
+
+        if "130" in desired_cuda:
+            cuda_rpaths.append("$ORIGIN/../../nvidia/cu13/lib")
+        else:
+            cuda_rpaths.extend(
+                [
+                    "$ORIGIN/../../nvidia/cublas/lib",
+                    "$ORIGIN/../../nvidia/cuda_cupti/lib",
+                    "$ORIGIN/../../nvidia/cuda_nvrtc/lib",
+                    "$ORIGIN/../../nvidia/cuda_runtime/lib",
+                    "$ORIGIN/../../nvidia/cufft/lib",
+                    "$ORIGIN/../../nvidia/curand/lib",
+                    "$ORIGIN/../../nvidia/cusolver/lib",
+                    "$ORIGIN/../../nvidia/cusparse/lib",
+                    "$ORIGIN/../../nvidia/nvtx/lib",
+                    "$ORIGIN/../../nvidia/cufile/lib",
+                ]
+            )
+
+        # Add $ORIGIN for local torch libs
+        rpath = ":".join(cuda_rpaths) + ":$ORIGIN"
+    else:
+        # For bundled libraries, just use $ORIGIN
+        rpath = "$ORIGIN"
+
+    if os.path.exists(lib_path):
+        os.system(
+            f"cd {folder}/tmp/torch/lib/; "
+            f"patchelf --set-rpath '{rpath}' --force-rpath {lib_name}"
+        )
+
+
+def copy_and_patch_library(
+    src_path: str,
+    folder: str,
+    use_nvidia_pypi_libs: bool = False,
+    desired_cuda: str = "",
+) -> None:
+    """Copy a library to torch/lib and patch its RPATH"""
+    if os.path.exists(src_path):
+        lib_name = os.path.basename(src_path)
+        shutil.copy2(src_path, f"{folder}/tmp/torch/lib/{lib_name}")
+        patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
+
+
+def package_cuda_wheel(wheel_path, desired_cuda) -> None:
+    """
+    Package the cuda wheel libraries
+    """
+    folder = os.path.dirname(wheel_path)
+    os.mkdir(f"{folder}/tmp")
+    os.system(f"unzip {wheel_path} -d {folder}/tmp")
+    # Delete original wheel since it will be repackaged
+    os.system(f"rm {wheel_path}")
+
+    # Check if we should use PyPI NVIDIA libraries or bundle system libraries
+    use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
+
+    if use_nvidia_pypi_libs:
+        print("Using nvidia libs from pypi - skipping CUDA library bundling")
+        # For PyPI approach, we don't bundle CUDA libraries - they come from PyPI packages
+        # We only need to bundle non-NVIDIA libraries
+        minimal_libs_to_copy = [
+            "/lib64/libgomp.so.1",
+            "/usr/lib64/libgfortran.so.5",
+            "/acl/build/libarm_compute.so",
+            "/acl/build/libarm_compute_graph.so",
+            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_lapack_core.so.0",
+            "/usr/local/lib/libnvpl_blas_core.so.0",
+        ]
+
+        # Copy minimal libraries to unzipped_folder/torch/lib
+        for lib_path in minimal_libs_to_copy:
+            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
+
+        # Patch torch libraries used for searching libraries
+        torch_libs_to_patch = [
+            "libtorch.so",
+            "libtorch_cpu.so",
+            "libtorch_cuda.so",
+            "libtorch_cuda_linalg.so",
+            "libtorch_global_deps.so",
+            "libtorch_python.so",
+            "libtorch_nvshmem.so",
+            "libc10.so",
+            "libc10_cuda.so",
+            "libcaffe2_nvrtc.so",
+            "libshm.so",
+        ]
+        for lib_name in torch_libs_to_patch:
+            patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
+    else:
+        print("Bundling CUDA libraries with wheel")
+        # Original logic for bundling system CUDA libraries
+        # Common libraries for all CUDA versions
+        common_libs = [
+            # Non-NVIDIA system libraries
+            "/lib64/libgomp.so.1",
+            "/usr/lib64/libgfortran.so.5",
+            "/acl/build/libarm_compute.so",
+            "/acl/build/libarm_compute_graph.so",
+            # Common CUDA libraries (same for all versions)
+            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_lapack_core.so.0",
+            "/usr/local/lib/libnvpl_blas_core.so.0",
+            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
+            "/usr/local/cuda/lib64/libcudnn.so.9",
+            "/usr/local/cuda/lib64/libcusparseLt.so.0",
+            "/usr/local/cuda/lib64/libcurand.so.10",
+            "/usr/local/cuda/lib64/libnccl.so.2",
+            "/usr/local/cuda/lib64/libnvshmem_host.so.3",
+            "/usr/local/cuda/lib64/libcudnn_adv.so.9",
+            "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
+            "/usr/local/cuda/lib64/libcudnn_graph.so.9",
+            "/usr/local/cuda/lib64/libcudnn_ops.so.9",
+            "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
+            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
+            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
+            "/usr/local/cuda/lib64/libcufile.so.0",
+            "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            "/usr/local/cuda/lib64/libcusparse.so.12",
+        ]
+
+        # CUDA version-specific libraries
+        if "13" in desired_cuda:
+            minor_version = desired_cuda[-1]
+            version_specific_libs = [
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
+                "/usr/local/cuda/lib64/libcublas.so.13",
+                "/usr/local/cuda/lib64/libcublasLt.so.13",
+                "/usr/local/cuda/lib64/libcudart.so.13",
+                "/usr/local/cuda/lib64/libcufft.so.12",
+                "/usr/local/cuda/lib64/libcusolver.so.12",
+                "/usr/local/cuda/lib64/libnvJitLink.so.13",
+                "/usr/local/cuda/lib64/libnvrtc.so.13",
+                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.{minor_version}",
+            ]
+        elif "12" in desired_cuda:
+            # Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
+            minor_version = desired_cuda[-1]
+            version_specific_libs = [
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
+                "/usr/local/cuda/lib64/libcublas.so.12",
+                "/usr/local/cuda/lib64/libcublasLt.so.12",
+                "/usr/local/cuda/lib64/libcudart.so.12",
+                "/usr/local/cuda/lib64/libcufft.so.11",
+                "/usr/local/cuda/lib64/libcusolver.so.11",
+                "/usr/local/cuda/lib64/libnvJitLink.so.12",
+                "/usr/local/cuda/lib64/libnvrtc.so.12",
+                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
+            ]
+        else:
+            raise ValueError(f"Unsupported CUDA version: {desired_cuda}.")
+
+        # Combine all libraries
+        libs_to_copy = common_libs + version_specific_libs
+
+        # Copy libraries to unzipped_folder/torch/lib
+        for lib_path in libs_to_copy:
+            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
+
+    # Make sure the wheel is tagged with manylinux_2_28
+    for f in os.scandir(f"{folder}/tmp/"):
+        if f.is_dir() and f.name.endswith(".dist-info"):
+            replace_tag(f"{f.path}/WHEEL")
+            break
+
+    os.system(f"wheel pack {folder}/tmp/ -d {folder}")
+    os.system(f"rm -rf {folder}/tmp/")
+
+
+def complete_wheel(folder: str) -> str:
+    """
+    Complete wheel build and put in artifact location
+    """
+    wheel_name = list_dir(f"/{folder}/dist")[0]
+
+    # Please note for cuda we don't run auditwheel since we use custom script to package
+    # the cuda dependencies to the wheel file using update_wheel() method.
+    # However we need to make sure filename reflects the correct Manylinux platform.
+    if "pytorch" in folder and not enable_cuda:
+        print("Repairing Wheel with AuditWheel")
+        check_call(["auditwheel", "repair", f"dist/{wheel_name}"], cwd=folder)
+        repaired_wheel_name = list_dir(f"/{folder}/wheelhouse")[0]
+
+        print(f"Moving {repaired_wheel_name} wheel to /{folder}/dist")
+        os.rename(
+            f"/{folder}/wheelhouse/{repaired_wheel_name}",
+            f"/{folder}/dist/{repaired_wheel_name}",
+        )
+    else:
+        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
+
+    print(f"Copying {repaired_wheel_name} to artifacts")
+    shutil.copy2(
+        f"/{folder}/dist/{repaired_wheel_name}", f"/artifacts/{repaired_wheel_name}"
+    )
+
+    return repaired_wheel_name
+
+
+def parse_arguments():
+    """
+    Parse inline arguments
+    """
+    from argparse import ArgumentParser
+
+    parser = ArgumentParser("AARCH64 wheels python CD")
+    parser.add_argument("--debug", action="store_true")
+    parser.add_argument("--build-only", action="store_true")
+    parser.add_argument("--test-only", type=str)
+    parser.add_argument("--enable-mkldnn", action="store_true")
+    parser.add_argument("--enable-cuda", action="store_true")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    """
+    Entry Point
+    """
+    args = parse_arguments()
+    enable_mkldnn = args.enable_mkldnn
+    enable_cuda = args.enable_cuda
+    branch = check_output(
+        ["git", "rev-parse", "--abbrev-ref", "HEAD"], cwd="/pytorch"
+    ).decode()
+
+    print("Building PyTorch wheel")
+    build_vars = ""
+    # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
+    if enable_cuda:
+        build_vars += "MAX_JOBS=5 "
+
+        # Handle PyPI NVIDIA libraries vs bundled libraries
+        use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
+        if use_nvidia_pypi_libs:
+            print("Configuring build for PyPI NVIDIA libraries")
+            # Configure for dynamic linking (matching x86 logic)
+            build_vars += "ATEN_STATIC_CUDA=0 USE_CUDA_STATIC_LINK=0 USE_CUPTI_SO=1 "
+        else:
+            print("Configuring build for bundled NVIDIA libraries")
+            # Keep existing static linking approach - already configured above
+
+    override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
+    desired_cuda = os.getenv("DESIRED_CUDA")
+    if override_package_version is not None:
+        version = override_package_version
+        build_vars += (
+            f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={version} PYTORCH_BUILD_NUMBER=1 "
+        )
+    elif branch in ["nightly", "main"]:
+        build_date = (
+            check_output(["git", "log", "--pretty=format:%cs", "-1"], cwd="/pytorch")
+            .decode()
+            .replace("-", "")
+        )
+        version = (
+            check_output(["cat", "version.txt"], cwd="/pytorch").decode().strip()[:-2]
+        )
+        if enable_cuda:
+            build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={version}.dev{build_date}+{desired_cuda} PYTORCH_BUILD_NUMBER=1 "
+        else:
+            build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={version}.dev{build_date} PYTORCH_BUILD_NUMBER=1 "
+    elif branch.startswith(("v1.", "v2.")):
+        build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={branch[1 : branch.find('-')]} PYTORCH_BUILD_NUMBER=1 "
+
+    if enable_mkldnn:
+        print("build pytorch with mkldnn+acl backend")
+        build_vars += "USE_MKLDNN=ON USE_MKLDNN_ACL=ON "
+        build_vars += "ACL_ROOT_DIR=/acl "
+        if enable_cuda:
+            build_vars += "BLAS=NVPL "
+        else:
+            build_vars += "BLAS=OpenBLAS OpenBLAS_HOME=/opt/OpenBLAS "
+    else:
+        print("build pytorch without mkldnn backend")
+
+    os.system(f"cd /pytorch; {build_vars} python3 -m build --wheel --no-isolation")
+    if enable_cuda:
+        print("Updating Cuda Dependency")
+        filename = os.listdir("/pytorch/dist/")
+        wheel_path = f"/pytorch/dist/{filename[0]}"
+        package_cuda_wheel(wheel_path, desired_cuda)
+    pytorch_wheel_name = complete_wheel("/pytorch/")
+    print(f"Build Complete. Created {pytorch_wheel_name}..")

.ci/aarch64_linux/build_aarch64_wheel.py

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

.ci/aarch64_linux/embed_library.py

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

.ci/manywheel/build.sh

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

.ci/manywheel/build_common.sh

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

.ci/manywheel/build_cpu.sh

@@ -15,10 +15,6 @@ if [[ -z "$EXTRA_CAFFE2_CMAKE_FLAGS" ]]; then
     EXTRA_CAFFE2_CMAKE_FLAGS=()
 fi
 
-# Detect architecture
-ARCH=$(uname -m)
-echo "Building CPU wheel for architecture: $ARCH"
-
 WHEELHOUSE_DIR="wheelhousecpu"
 LIBTORCH_HOUSE_DIR="libtorch_housecpu"
 if [[ -z "$PYTORCH_FINAL_PACKAGE_DIR" ]]; then
@@ -38,10 +34,8 @@ elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
 elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
     LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
 elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
-    if [[ "$ARCH" == "s390x" ]]; then
+    if [[ "$(uname -m)" == "s390x" ]]; then
         LIBGOMP_PATH="/usr/lib/s390x-linux-gnu/libgomp.so.1"
-    elif [[ "$ARCH" == "aarch64" ]]; then
-        LIBGOMP_PATH="/usr/lib/aarch64-linux-gnu/libgomp.so.1"
     else
         LIBGOMP_PATH="/usr/lib/x86_64-linux-gnu/libgomp.so.1"
     fi
@@ -55,32 +49,6 @@ DEPS_SONAME=(
     "libgomp.so.1"
 )
 
-# Add ARM-specific library dependencies for CPU builds
-if [[ "$ARCH" == "aarch64" ]]; then
-    echo "Adding ARM-specific CPU library dependencies"
-
-    # ARM Compute Library (if available)
-    if [[ -d "/acl/build" ]]; then
-        echo "Adding ARM Compute Library for CPU"
-        DEPS_LIST+=(
-            "/acl/build/libarm_compute.so"
-            "/acl/build/libarm_compute_graph.so"
-        )
-        DEPS_SONAME+=(
-            "libarm_compute.so"
-            "libarm_compute_graph.so"
-        )
-    fi
-
-    # ARM system libraries
-    DEPS_LIST+=(
-        "/usr/lib64/libgfortran.so.5"
-    )
-    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

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

.ci/onnx/common.sh

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

.ci/onnx/test.sh

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

.ci/pytorch/build.sh

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

.ci/pytorch/smoke_test/check_binary_symbols.py

@@ -100,337 +100,6 @@ def check_lib_statically_linked_libstdc_cxx_abi_symbols(lib: str) -> None:
         )
 
 
-def _compile_and_extract_symbols(
-    cpp_content: str, compile_flags: list[str], exclude_list: list[str] | None = None
-) -> list[str]:
-    """
-    Helper to compile a C++ file and extract all symbols.
-
-    Args:
-        cpp_content: C++ source code to compile
-        compile_flags: Compilation flags
-        exclude_list: List of symbol names to exclude. Defaults to ["main"].
-
-    Returns:
-        List of all symbols found in the object file (excluding those in exclude_list).
-    """
-    import subprocess
-    import tempfile
-
-    if exclude_list is None:
-        exclude_list = ["main"]
-
-    with tempfile.TemporaryDirectory() as tmpdir:
-        tmppath = Path(tmpdir)
-        cpp_file = tmppath / "test.cpp"
-        obj_file = tmppath / "test.o"
-
-        cpp_file.write_text(cpp_content)
-
-        result = subprocess.run(
-            compile_flags + [str(cpp_file), "-o", str(obj_file)],
-            capture_output=True,
-            text=True,
-            timeout=60,
-        )
-
-        if result.returncode != 0:
-            raise RuntimeError(f"Compilation failed: {result.stderr}")
-
-        symbols = get_symbols(str(obj_file))
-
-        # Return all symbol names, excluding those in the exclude list
-        return [name for _addr, _stype, name in symbols if name not in exclude_list]
-
-
-def check_stable_only_symbols(install_root: Path) -> None:
-    """
-    Test TORCH_STABLE_ONLY and TORCH_TARGET_VERSION by compiling test code and comparing symbol counts.
-
-    This approach tests:
-    1. WITHOUT macros -> many torch symbols exposed
-    2. WITH TORCH_STABLE_ONLY -> zero torch symbols (all hidden)
-    3. WITH TORCH_TARGET_VERSION -> zero torch symbols (all hidden)
-    4. WITH both macros -> zero torch symbols (all hidden)
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    test_cpp_content = """
-// Main torch C++ API headers
-#include <torch/torch.h>
-#include <torch/all.h>
-
-// ATen tensor library
-#include <ATen/ATen.h>
-
-// Core c10 headers (commonly used)
-#include <c10/core/Device.h>
-#include <c10/core/DeviceType.h>
-#include <c10/core/ScalarType.h>
-#include <c10/core/TensorOptions.h>
-#include <c10/util/Optional.h>
-
-int main() { return 0; }
-"""
-
-    base_compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",  # Compile only, don't link
-    ]
-
-    # Compile WITHOUT any macros
-    symbols_without = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=base_compile_flags,
-    )
-
-    # We expect constexpr symbols, inline functions used by other headers etc.
-    # to produce symbols
-    num_symbols_without = len(symbols_without)
-    print(f"Found {num_symbols_without} symbols without any macros defined")
-    assert num_symbols_without != 0, (
-        "Expected a non-zero number of symbols without any macros"
-    )
-
-    # Compile WITH TORCH_STABLE_ONLY (expect 0 symbols)
-    compile_flags_with_stable_only = base_compile_flags + ["-DTORCH_STABLE_ONLY"]
-
-    symbols_with_stable_only = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=compile_flags_with_stable_only,
-    )
-
-    num_symbols_with_stable_only = len(symbols_with_stable_only)
-    assert num_symbols_with_stable_only == 0, (
-        f"Expected no symbols with TORCH_STABLE_ONLY macro, but found {num_symbols_with_stable_only}"
-    )
-
-    # Compile WITH TORCH_TARGET_VERSION (expect 0 symbols)
-    compile_flags_with_target_version = base_compile_flags + [
-        "-DTORCH_TARGET_VERSION=1"
-    ]
-
-    symbols_with_target_version = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=compile_flags_with_target_version,
-    )
-
-    num_symbols_with_target_version = len(symbols_with_target_version)
-    assert num_symbols_with_target_version == 0, (
-        f"Expected no symbols with TORCH_TARGET_VERSION macro, but found {num_symbols_with_target_version}"
-    )
-
-    # Compile WITH both macros (expect 0 symbols)
-    compile_flags_with_both = base_compile_flags + [
-        "-DTORCH_STABLE_ONLY",
-        "-DTORCH_TARGET_VERSION=1",
-    ]
-
-    symbols_with_both = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=compile_flags_with_both,
-    )
-
-    num_symbols_with_both = len(symbols_with_both)
-    assert num_symbols_with_both == 0, (
-        f"Expected no symbols with both macros, but found {num_symbols_with_both}"
-    )
-
-
-def check_stable_api_symbols(install_root: Path) -> None:
-    """
-    Test that stable API headers still expose symbols with TORCH_STABLE_ONLY.
-    The torch/csrc/stable/c/shim.h header is tested in check_stable_c_shim_symbols
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    stable_dir = include_dir / "torch" / "csrc" / "stable"
-    assert stable_dir.exists(), f"Expected {stable_dir} to be present"
-
-    stable_headers = list(stable_dir.rglob("*.h"))
-    if not stable_headers:
-        raise RuntimeError("Could not find any stable headers")
-
-    includes = []
-    for header in stable_headers:
-        rel_path = header.relative_to(include_dir)
-        includes.append(f"#include <{rel_path.as_posix()}>")
-
-    includes_str = "\n".join(includes)
-    test_stable_content = f"""
-{includes_str}
-int main() {{ return 0; }}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_stable = _compile_and_extract_symbols(
-        cpp_content=test_stable_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_stable = len(symbols_stable)
-    print(f"Found {num_symbols_stable} symbols in torch/csrc/stable")
-    assert num_symbols_stable > 0, (
-        f"Expected stable headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_stable} symbols"
-    )
-
-
-def check_headeronly_symbols(install_root: Path) -> None:
-    """
-    Test that header-only utility headers still expose symbols with TORCH_STABLE_ONLY.
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    # Find all headers in torch/headeronly
-    headeronly_dir = include_dir / "torch" / "headeronly"
-    assert headeronly_dir.exists(), f"Expected {headeronly_dir} to be present"
-    headeronly_headers = list(headeronly_dir.rglob("*.h"))
-    if not headeronly_headers:
-        raise RuntimeError("Could not find any headeronly headers")
-
-    # Filter out platform-specific headers that may not compile everywhere
-    platform_specific_keywords = [
-        "cpu/vec",
-    ]
-
-    filtered_headers = []
-    for header in headeronly_headers:
-        rel_path = header.relative_to(include_dir).as_posix()
-        if not any(
-            keyword in rel_path.lower() for keyword in platform_specific_keywords
-        ):
-            filtered_headers.append(header)
-
-    includes = []
-    for header in filtered_headers:
-        rel_path = header.relative_to(include_dir)
-        includes.append(f"#include <{rel_path.as_posix()}>")
-
-    includes_str = "\n".join(includes)
-    test_headeronly_content = f"""
-{includes_str}
-int main() {{ return 0; }}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_headeronly = _compile_and_extract_symbols(
-        cpp_content=test_headeronly_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_headeronly = len(symbols_headeronly)
-    print(f"Found {num_symbols_headeronly} symbols in torch/headeronly")
-    assert num_symbols_headeronly > 0, (
-        f"Expected headeronly headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_headeronly} symbols"
-    )
-
-
-def check_aoti_shim_symbols(install_root: Path) -> None:
-    """
-    Test that AOTI shim headers still expose symbols with TORCH_STABLE_ONLY.
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    # There are no constexpr symbols etc., so we need to actually use functions
-    # so that some symbols are found.
-    test_shim_content = """
-#include <torch/csrc/inductor/aoti_torch/c/shim.h>
-int main() {
-    int32_t (*fp1)() = &aoti_torch_device_type_cpu;
-    int32_t (*fp2)() = &aoti_torch_dtype_float32;
-    (void)fp1; (void)fp2;
-    return 0;
-}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_shim = _compile_and_extract_symbols(
-        cpp_content=test_shim_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_shim = len(symbols_shim)
-    assert num_symbols_shim > 0, (
-        f"Expected shim headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_shim} symbols"
-    )
-
-
-def check_stable_c_shim_symbols(install_root: Path) -> None:
-    """
-    Test that stable C shim headers still expose symbols with TORCH_STABLE_ONLY.
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    # Check if the stable C shim exists
-    stable_shim = include_dir / "torch" / "csrc" / "stable" / "c" / "shim.h"
-    if not stable_shim.exists():
-        raise RuntimeError("Could not find stable c shim")
-
-    # There are no constexpr symbols etc., so we need to actually use functions
-    # so that some symbols are found.
-    test_stable_shim_content = """
-#include <torch/csrc/stable/c/shim.h>
-int main() {
-    // Reference stable C API functions to create undefined symbols
-    AOTITorchError (*fp1)(const char*, uint32_t*, int32_t*) = &torch_parse_device_string;
-    AOTITorchError (*fp2)(uint32_t*) = &torch_get_num_threads;
-    (void)fp1; (void)fp2;
-    return 0;
-}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_stable_shim = _compile_and_extract_symbols(
-        cpp_content=test_stable_shim_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_stable_shim = len(symbols_stable_shim)
-    assert num_symbols_stable_shim > 0, (
-        f"Expected stable C shim headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_stable_shim} symbols"
-    )
-
-
 def check_lib_symbols_for_abi_correctness(lib: str) -> None:
     print(f"lib: {lib}")
     cxx11_symbols = grep_symbols(lib, LIBTORCH_CXX11_PATTERNS)
@@ -460,13 +129,6 @@ def main() -> None:
     check_lib_symbols_for_abi_correctness(libtorch_cpu_path)
     check_lib_statically_linked_libstdc_cxx_abi_symbols(libtorch_cpu_path)
 
-    # Check symbols when TORCH_STABLE_ONLY is defined
-    check_stable_only_symbols(install_root)
-    check_stable_api_symbols(install_root)
-    check_headeronly_symbols(install_root)
-    check_aoti_shim_symbols(install_root)
-    check_stable_c_shim_symbols(install_root)
-
 
 if __name__ == "__main__":
     main()

.ci/pytorch/test.sh

@@ -389,6 +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
@@ -1243,6 +1250,97 @@ test_custom_script_ops() {
   assert_git_not_dirty
 }
 
+test_libtorch_agnostic_targetting() {
+    echo "Testing libtorch_agnostic runs correctly on TORCH_TARGET_VERSION"
+
+    REPO_DIR=$(pwd)
+    WHEEL_DIR="${REPO_DIR}/test/cpp_extensions/.wheels"
+
+    # Build wheel with current PyTorch (this has TORCH_TARGET_VERSION 2_9_0)
+    echo "Building 2.9 extension wheel with current PyTorch..."
+    pushd test/cpp_extensions/libtorch_agnostic_2_9_extension
+    time python setup.py bdist_wheel
+
+    # Save the wheel
+    mkdir -p "$WHEEL_DIR"
+    cp dist/*.whl "$WHEEL_DIR/"
+    WHEEL_FILE=$(find "$WHEEL_DIR" -maxdepth 1 -name "*.whl" -type f | head -1)
+    echo "Built wheel: $(basename "$WHEEL_FILE")"
+    popd
+
+    # Create venv and install PyTorch 2.9
+    python -m venv venv_pytorch_2_9
+    # shellcheck disable=SC1091
+    . venv_pytorch_2_9/bin/activate
+
+    # Clear PYTHONPATH to avoid using the development PyTorch
+    echo "Clearing PYTHONPATH to use only venv packages..."
+    unset PYTHONPATH
+
+    # Upgrade pip to latest version
+    echo "Upgrading pip to latest version..."
+    pip install --upgrade pip
+    pip --version
+
+    echo "Installing PyTorch 2.9..."
+
+    # Install from release channel only
+    PYTORCH_VERSION="2.9.0"
+
+    # Extract CUDA version from BUILD_ENVIRONMENT (e.g., "cuda12.1" -> "cu121")
+    if [[ "$BUILD_ENVIRONMENT" =~ cuda([0-9]+)\.([0-9]+) ]]; then
+        CUDA_MAJOR="${BASH_REMATCH[1]}"
+        CUDA_MINOR="${BASH_REMATCH[2]}"
+        CUDA_VERSION="cu${CUDA_MAJOR}${CUDA_MINOR}"
+        echo "  Detected CUDA ${CUDA_MAJOR}.${CUDA_MINOR} from BUILD_ENVIRONMENT, using ${CUDA_VERSION}"
+    else
+        # Default to CPU build
+        CUDA_VERSION="cpu"
+        echo "  No CUDA detected in BUILD_ENVIRONMENT, using CPU build"
+    fi
+
+    if pip install torch=="${PYTORCH_VERSION}" --index-url https://download.pytorch.org/whl/${CUDA_VERSION}/; then
+        echo "Installed PyTorch ${PYTORCH_VERSION} from release channel (${CUDA_VERSION})"
+    else
+        echo "  FAILED to install PyTorch 2.9.0 from release channel"
+        echo "  URL: https://download.pytorch.org/whl/${CUDA_VERSION}/"
+        deactivate
+        rm -rf venv_pytorch_2_9
+        return 1
+    fi
+
+    INSTALLED_VERSION=$(python -c "import torch; print(torch.__version__)" 2>/dev/null || echo "unknown")
+    echo "  Installed version: $INSTALLED_VERSION"
+
+    # Install test dependencies
+    echo "Installing test dependencies..."
+    pip install expecttest numpy unittest-xml-reporting
+
+    # Install the pre-built wheel
+    echo ""
+    echo "Installing pre-built 2.9 extension wheel (built with PyTorch 2.10)..."
+    pip install "$WHEEL_FILE"
+    echo "Installed $(basename "$WHEEL_FILE") into PyTorch 2.9 environment"
+
+    # Run tests with PyTorch 2.9 runtime (2.10 tests will be skipped automatically)
+    echo ""
+    echo "Running tests with PyTorch 2.9 runtime (using wheel built on PyTorch 2.10)..."
+    if time python test/cpp_extensions/test_libtorch_agnostic.py -v; then
+        echo ""
+        echo "  Wheel built with current torch and TORCH_TARGET_VERSION 2_9_0 works with PyTorch 2.9 runtime!"
+    else
+        echo "targeting test failed"
+        deactivate
+        rm -rf venv_pytorch_2_9 "$WHEEL_DIR"
+        return 1
+    fi
+
+    deactivate
+    rm -rf venv_pytorch_2_9 "$WHEEL_DIR"
+
+    assert_git_not_dirty
+}
+
 test_jit_hooks() {
   echo "Testing jit hooks in cpp"
   HOOK_BUILD="${CUSTOM_TEST_ARTIFACT_BUILD_DIR}/jit-hook-build"
@@ -1715,6 +1813,8 @@ elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]];
 elif [[ "${TEST_CONFIG}" == *backward* ]]; then
   test_forward_backward_compatibility
   # Do NOT add tests after bc check tests, see its comment.
+elif [[ "${TEST_CONFIG}" == *libtorch_agnostic_targetting* ]]; then
+  test_libtorch_agnostic_targetting
 elif [[ "${TEST_CONFIG}" == *xla* ]]; then
   install_torchvision
   build_xla
@@ -1814,6 +1914,8 @@ 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 @@
-07b6cbde121417a70e4dc871adb6d27030e0ce3f
+ee1a1350eb37804b94334768f328144f058f14e9

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-acccf86477759b2d3500f1ae1be065f7b1e409ec
+2d82dc5caa336d179d9b46ac4a0fb8c43d84c5cc

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-e4d25697f9dc5eedaf8f0a5bf085c62c5455a53a
+94631807d22c09723dd006f7be5beb649d5f88d0

.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/workflows/_binary-build-linux.yml

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

.github/workflows/_linux-build.yml

@@ -330,8 +330,6 @@ jobs:
             -e PR_NUMBER \
             -e SHA1 \
             -e BRANCH \
-            -e SCCACHE_BUCKET \
-            -e SCCACHE_REGION \
             -e XLA_CUDA \
             -e XLA_CLANG_CACHE_S3_BUCKET_NAME \
             -e SKIP_SCCACHE_INITIALIZATION=1 \

.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/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/inductor-perf-test-nightly-xpu.yml

@@ -85,7 +85,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-noble-xpu-n-py3.10
       docker-image-name: ci-image:pytorch-linux-noble-xpu-n-py3-inductor-benchmarks
-      runner: linux.c7i.12xlarge
+      runner: linux.c7i.2xlarge
       test-matrix: |
         { include: [
           { config: "inductor_huggingface_perf_xpu", shard: 1, num_shards: 5, runner: "linux.idc.xpu" },
@@ -108,41 +108,41 @@ jobs:
         ]}
     secrets: inherit
 
-  xpu-n-py3_10-inductor-benchmark-test-nightly:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name != 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-benchmark
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-noble-xpu-n-py3.10
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-false-cppwrapper-true-aotinductor-true-freezing_cudagraphs-false-cudagraphs_low_precision-false
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      # Disable monitor in perf tests for more investigation
-      disable-monitor: true
-      monitor-log-interval: 10
-      monitor-data-collect-interval: 2
-    secrets: inherit
+  # xpu-n-py3_10-inductor-benchmark-test-nightly:
+  #   permissions:
+  #     id-token: write
+  #     contents: read
+  #   if: github.event_name != 'workflow_dispatch'
+  #   name: xpu-n-py3.10-inductor-benchmark
+  #   uses: ./.github/workflows/_xpu-test.yml
+  #   needs: xpu-n-py3_10-inductor-benchmark-build
+  #   with:
+  #     build-environment: linux-noble-xpu-n-py3.10
+  #     dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-false-cppwrapper-true-aotinductor-true-freezing_cudagraphs-false-cudagraphs_low_precision-false
+  #     docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
+  #     timeout-minutes: 720
+  #     # Disable monitor in perf tests for more investigation
+  #     disable-monitor: true
+  #     monitor-log-interval: 10
+  #     monitor-data-collect-interval: 2
+  #   secrets: inherit
 
-  xpu-n-py3_10-inductor-benchmark-test:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name == 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-test
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-noble-xpu-n-py3.10
-      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit
+  # xpu-n-py3_10-inductor-benchmark-test:
+  #   permissions:
+  #     id-token: write
+  #     contents: read
+  #   if: github.event_name == 'workflow_dispatch'
+  #   name: xpu-n-py3.10-inductor-test
+  #   uses: ./.github/workflows/_xpu-test.yml
+  #   needs: xpu-n-py3_10-inductor-benchmark-build
+  #   with:
+  #     build-environment: linux-noble-xpu-n-py3.10
+  #     dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
+  #     docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
+  #     timeout-minutes: 720
+  #     disable-monitor: false
+  #     monitor-log-interval: 15
+  #     monitor-data-collect-interval: 4
+  #   secrets: inherit

.github/workflows/operator_microbenchmark.yml

@@ -24,7 +24,7 @@ jobs:
     name: opmicrobenchmark-build
     uses: ./.github/workflows/_linux-build.yml
     with:
-      runner: linux.12xlarge.memory
+      runner: linux.c7i.2xlarge
       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'
@@ -35,16 +35,16 @@ jobs:
         ]}
     secrets: inherit
 
-  opmicrobenchmark-test:
-    name: opmicrobenchmark-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs: opmicrobenchmark-build
-    with:
-      timeout-minutes: 500
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image: ${{ needs.opmicrobenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.opmicrobenchmark-build.outputs.test-matrix }}
-    secrets: inherit
+  # opmicrobenchmark-test:
+  #   name: opmicrobenchmark-test
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs: opmicrobenchmark-build
+  #   with:
+  #     timeout-minutes: 500
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+  #     docker-image: ${{ needs.opmicrobenchmark-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.opmicrobenchmark-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   # B200 runner
   opmicrobenchmark-build-b200:
@@ -62,17 +62,17 @@ jobs:
         ]}
     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
+  # opmicrobenchmark-test-b200:
+  #   name: opmicrobenchmark-test-b200
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs: opmicrobenchmark-build-b200
+  #   with:
+  #     timeout-minutes: 500
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+  #     docker-image: ${{ needs.opmicrobenchmark-build-b200.outputs.docker-image }}
+  #     test-matrix: ${{ needs.opmicrobenchmark-build-b200.outputs.test-matrix }}
+  #     aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+  #   secrets: inherit
 
   # ROCM MI300 runner
   opmicrobenchmark-build-rocm:
@@ -88,13 +88,13 @@ jobs:
         ]}
     secrets: inherit
 
-  opmicrobenchmark-test-rocm:
-    name: opmicrobenchmark-test-rocm
-    uses: ./.github/workflows/_rocm-test.yml
-    needs: opmicrobenchmark-build-rocm
-    with:
-      timeout-minutes: 500
-      build-environment: linux-jammy-rocm-py3_10
-      docker-image: ${{ needs.opmicrobenchmark-build-rocm.outputs.docker-image }}
-      test-matrix: ${{ needs.opmicrobenchmark-build-rocm.outputs.test-matrix }}
-    secrets: inherit
+  # opmicrobenchmark-test-rocm:
+  #   name: opmicrobenchmark-test-rocm
+  #   uses: ./.github/workflows/_rocm-test.yml
+  #   needs: opmicrobenchmark-build-rocm
+  #   with:
+  #     timeout-minutes: 500
+  #     build-environment: linux-jammy-rocm-py3_10
+  #     docker-image: ${{ needs.opmicrobenchmark-build-rocm.outputs.docker-image }}
+  #     test-matrix: ${{ needs.opmicrobenchmark-build-rocm.outputs.test-matrix }}
+  #   secrets: inherit

.github/workflows/pull.yml

@@ -70,20 +70,21 @@ jobs:
           { config: "distributed", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
           { config: "distributed", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
           { config: "numpy_2_x", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "libtorch_agnostic_targetting", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
         ]}
     secrets: inherit
 
-  linux-jammy-py3_10-gcc11-test:
-    name: linux-jammy-py3.10-gcc11
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-py3_10-gcc11-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-py3.10-gcc11
-      docker-image: ${{ needs.linux-jammy-py3_10-gcc11-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_10-gcc11-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-py3_10-gcc11-test:
+  #   name: linux-jammy-py3.10-gcc11
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-py3_10-gcc11-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-py3.10-gcc11
+  #     docker-image: ${{ needs.linux-jammy-py3_10-gcc11-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3_10-gcc11-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-docs:
     name: linux-docs
@@ -127,7 +128,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      runner: linux.2xlarge.memory
+      runner: linux.c7i.2xlarge
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
@@ -144,18 +145,18 @@ jobs:
       sync-tag: asan-build
     secrets: inherit
 
-  linux-jammy-py3_10-clang18-asan-test:
-    name: linux-jammy-py3.10-clang18-asan
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-py3_10-clang18-asan-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-py3.10-clang18-asan
-      docker-image: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.test-matrix }}
-      sync-tag: asan-test
-    secrets: inherit
+  # linux-jammy-py3_10-clang18-asan-test:
+  #   name: linux-jammy-py3.10-clang18-asan
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-py3_10-clang18-asan-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-py3.10-clang18-asan
+  #     docker-image: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.test-matrix }}
+  #     sync-tag: asan-test
+  #   secrets: inherit
 
   linux-jammy-py3_10-clang12-onnx-build:
     name: linux-jammy-py3.10-clang12-onnx
@@ -172,17 +173,17 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-py3_10-clang12-onnx-test:
-    name: linux-jammy-py3.10-clang12-onnx
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-py3_10-clang12-onnx-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-py3.10-clang12-onnx
-      docker-image: ${{ needs.linux-jammy-py3_10-clang12-onnx-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_10-clang12-onnx-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-py3_10-clang12-onnx-test:
+  #   name: linux-jammy-py3.10-clang12-onnx
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-py3_10-clang12-onnx-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-py3.10-clang12-onnx
+  #     docker-image: ${{ needs.linux-jammy-py3_10-clang12-onnx-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3_10-clang12-onnx-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-jammy-py3_10-clang12-build:
     name: linux-jammy-py3.10-clang12
@@ -208,17 +209,17 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-py3_10-clang12-test:
-    name: linux-jammy-py3.10-clang12
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-py3_10-clang12-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-py3.10-clang12
-      docker-image: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-py3_10-clang12-test:
+  #   name: linux-jammy-py3.10-clang12
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-py3_10-clang12-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-py3.10-clang12
+  #     docker-image: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-jammy-py3_13-clang12-build:
     name: linux-jammy-py3.13-clang12
@@ -244,15 +245,15 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-py3_13-clang12-test:
-    name: linux-jammy-py3.13-clang12
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-py3_13-clang12-build
-    with:
-      build-environment: linux-jammy-py3.13-clang12
-      docker-image: ${{ needs.linux-jammy-py3_13-clang12-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_13-clang12-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-py3_13-clang12-test:
+  #   name: linux-jammy-py3.13-clang12
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs: linux-jammy-py3_13-clang12-build
+  #   with:
+  #     build-environment: linux-jammy-py3.13-clang12
+  #     docker-image: ${{ needs.linux-jammy-py3_13-clang12-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3_13-clang12-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-jammy-cuda12_8-cudnn9-py3_10-clang12-build:
     name: linux-jammy-cuda12.8-cudnn9-py3.10-clang12
@@ -268,20 +269,20 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-cpu-py3_10-gcc11-bazel-test:
-    name: linux-jammy-cpu-py3.10-gcc11-bazel-test
-    uses: ./.github/workflows/_bazel-build-test.yml
-    needs: get-label-type
-    with:
-      runner: "${{ needs.get-label-type.outputs.label-type }}linux.large"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-bazel-test
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-version: cpu
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-        ]}
-    secrets: inherit
+  # linux-jammy-cpu-py3_10-gcc11-bazel-test:
+  #   name: linux-jammy-cpu-py3.10-gcc11-bazel-test
+  #   uses: ./.github/workflows/_bazel-build-test.yml
+  #   needs: get-label-type
+  #   with:
+  #     runner: "${{ needs.get-label-type.outputs.label-type }}linux.large"
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc11-bazel-test
+  #     docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
+  #     cuda-version: cpu
+  #     test-matrix: |
+  #       { include: [
+  #         { config: "default", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+  #       ]}
+  #   secrets: inherit
 
   linux-jammy-py3_10-gcc11-mobile-lightweight-dispatch-build:
     name: linux-jammy-py3.10-gcc11-mobile-lightweight-dispatch-build
@@ -332,15 +333,15 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
-    name: cuda12.8-py3.10-gcc9-sm75
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm75
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
+  #   name: cuda12.8-py3.10-gcc9-sm75
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
+  #   with:
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm75
+  #     docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-noble-xpu-n-py3_10-build:
     name: linux-noble-xpu-n-py3.10

.github/workflows/riscv64.yml

@@ -20,5 +20,5 @@ jobs:
     with:
       build-environment: linux-noble-riscv64-py3.12-gcc14
       docker-image-name: pytorch-linux-noble-riscv64-py3.12-gcc14
-      runner: linux.2xlarge
+      runner: linux.c7i.2xlarge
     secrets: inherit

.github/workflows/slow.yml

@@ -49,68 +49,68 @@ jobs:
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
 
-  linux-jammy-cuda12_8-py3_10-gcc11-sm86-build:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm86
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: 8.6
-      test-matrix: |
-        { include: [
-          { config: "slow", shard: 1, num_shards: 3, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "slow", shard: 2, num_shards: 3, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "slow", shard: 3, num_shards: 3, runner: "linux.g5.4xlarge.nvidia.gpu" },
-        ]}
-    secrets: inherit
+  # linux-jammy-cuda12_8-py3_10-gcc11-sm86-build:
+  #   name: linux-jammy-cuda12.8-py3.10-gcc11-sm86
+  #   uses: ./.github/workflows/_linux-build.yml
+  #   needs: get-label-type
+  #   with:
+  #     runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
+  #     docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
+  #     cuda-arch-list: 8.6
+  #     test-matrix: |
+  #       { include: [
+  #         { config: "slow", shard: 1, num_shards: 3, runner: "linux.g5.4xlarge.nvidia.gpu" },
+  #         { config: "slow", shard: 2, num_shards: 3, runner: "linux.g5.4xlarge.nvidia.gpu" },
+  #         { config: "slow", shard: 3, num_shards: 3, runner: "linux.g5.4xlarge.nvidia.gpu" },
+  #       ]}
+  #   secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc11-sm86-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm86
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-sm86-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm86-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm86-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-cuda12_8-py3_10-gcc11-sm86-test:
+  #   name: linux-jammy-cuda12.8-py3.10-gcc11-sm86
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-cuda12_8-py3_10-gcc11-sm86-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
+  #     docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm86-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm86-build.outputs.test-matrix }}
+  #   secrets: inherit
 
-  linux-jammy-py3_10-clang12-build:
-    name: linux-jammy-py3.10-clang12
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-clang12
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.10-clang12
-      test-matrix: |
-        { include: [
-          { config: "slow", shard: 1, num_shards: 2, runner: "linux.2xlarge" },
-          { config: "slow", shard: 2, num_shards: 2, runner: "linux.2xlarge" },
-        ]}
-    secrets: inherit
+  # linux-jammy-py3_10-clang12-build:
+  #   name: linux-jammy-py3.10-clang12
+  #   uses: ./.github/workflows/_linux-build.yml
+  #   needs: get-label-type
+  #   with:
+  #     runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+  #     build-environment: linux-jammy-py3.10-clang12
+  #     docker-image-name: ci-image:pytorch-linux-jammy-py3.10-clang12
+  #     test-matrix: |
+  #       { include: [
+  #         { config: "slow", shard: 1, num_shards: 2, runner: "linux.2xlarge" },
+  #         { config: "slow", shard: 2, num_shards: 2, runner: "linux.2xlarge" },
+  #       ]}
+  #   secrets: inherit
 
-  linux-jammy-py3_10-clang12-test:
-    name: linux-jammy-py3.10-clang12
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-py3_10-clang12-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-py3.10-clang12
-      docker-image: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-py3_10-clang12-test:
+  #   name: linux-jammy-py3.10-clang12
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-py3_10-clang12-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-py3.10-clang12
+  #     docker-image: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-jammy-py3_10-clang18-asan-build:
     name: linux-jammy-py3.10-clang18-asan
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      runner: linux.2xlarge.memory
+      runner: linux.c7i.2xlarge
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
@@ -123,15 +123,15 @@ jobs:
       sync-tag: asan-build
     secrets: inherit
 
-  linux-jammy-py3_10-clang18-asan-test:
-    name: linux-jammy-py3.10-clang18-asan
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-py3_10-clang18-asan-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-py3.10-clang18-asan
-      docker-image: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.test-matrix }}
-      sync-tag: asan-test
-    secrets: inherit
+  # linux-jammy-py3_10-clang18-asan-test:
+  #   name: linux-jammy-py3.10-clang18-asan
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-py3_10-clang18-asan-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-py3.10-clang18-asan
+  #     docker-image: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3_10-clang18-asan-build.outputs.test-matrix }}
+  #     sync-tag: asan-test
+  #   secrets: inherit

.github/workflows/test-b200.yml

@@ -54,7 +54,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: linux.12xlarge.memory
+      runner: linux.r7i.4xlarge
       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'
@@ -65,14 +65,14 @@ jobs:
       # config: "smoke_b200" maps to test_python_smoke_b200() in .ci/pytorch/test.sh
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc11-sm100-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-sm100-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
-      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
+  # linux-jammy-cuda12_8-py3_10-gcc11-sm100-test:
+  #   name: linux-jammy-cuda12.8-py3.10-gcc11-sm100
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-cuda12_8-py3_10-gcc11-sm100-build
+  #   with:
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
+  #     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

.github/workflows/trunk.yml

@@ -56,7 +56,7 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       build-generates-artifacts: false
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: "linux.c7i.4xlarge"
+      runner: "linux.c7i.2xlarge"
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 1 },
@@ -83,21 +83,22 @@ jobs:
           { config: "distributed", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.12xlarge.nvidia.gpu" },
           { config: "distributed", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.12xlarge.nvidia.gpu" },
           { config: "pr_time_benchmarks", shard: 1, num_shards: 1, runner: "linux.g4dn.metal.nvidia.gpu" },
+          { config: "libtorch_agnostic_targetting", shard: 1, num_shards: 1, runner: "linux.g4dn.metal.nvidia.gpu" },
         ]}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc11-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc11
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-build
-      - target-determination
-    with:
-      timeout-minutes: 360
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-cuda12_8-py3_10-gcc11-test:
+  #   name: linux-jammy-cuda12.8-py3.10-gcc11
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-cuda12_8-py3_10-gcc11-build
+  #     - target-determination
+  #   with:
+  #     timeout-minutes: 360
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc11
+  #     docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.test-matrix }}
+  #   secrets: inherit
 
 
   # no-ops builds test USE_PER_OPERATOR_HEADERS=0 where ATen/ops is not generated
@@ -136,19 +137,19 @@ jobs:
         ]}
     secrets: inherit
 
-  macos-py3-arm64-test:
-    name: macos-py3-arm64
-    uses: ./.github/workflows/_mac-test.yml
-    needs:
-      - macos-py3-arm64-build
-      - target-determination
-    with:
-      build-environment: macos-py3-arm64
-      # Same as the build job
-      python-version: 3.12.7
-      test-matrix: ${{ needs.macos-py3-arm64-build.outputs.test-matrix }}
-      disable-monitor: false
-    secrets: inherit
+  # macos-py3-arm64-test:
+  #   name: macos-py3-arm64
+  #   uses: ./.github/workflows/_mac-test.yml
+  #   needs:
+  #     - macos-py3-arm64-build
+  #     - target-determination
+  #   with:
+  #     build-environment: macos-py3-arm64
+  #     # Same as the build job
+  #     python-version: 3.12.7
+  #     test-matrix: ${{ needs.macos-py3-arm64-build.outputs.test-matrix }}
+  #     disable-monitor: false
+  #   secrets: inherit
 
   win-vs2022-cpu-py3-build:
     name: win-vs2022-cpu-py3
@@ -167,18 +168,18 @@ jobs:
         ]}
     secrets: inherit
 
-  win-vs2022-cpu-py3-test:
-    name: win-vs2022-cpu-py3
-    uses: ./.github/workflows/_win-test.yml
-    needs:
-      - win-vs2022-cpu-py3-build
-      - target-determination
-    with:
-      build-environment: win-vs2022-cpu-py3
-      cuda-version: cpu
-      test-matrix: ${{ needs.win-vs2022-cpu-py3-build.outputs.test-matrix }}
-      disable-monitor: false
-    secrets: inherit
+  # win-vs2022-cpu-py3-test:
+  #   name: win-vs2022-cpu-py3
+  #   uses: ./.github/workflows/_win-test.yml
+  #   needs:
+  #     - win-vs2022-cpu-py3-build
+  #     - target-determination
+  #   with:
+  #     build-environment: win-vs2022-cpu-py3
+  #     cuda-version: cpu
+  #     test-matrix: ${{ needs.win-vs2022-cpu-py3-build.outputs.test-matrix }}
+  #     disable-monitor: false
+  #   secrets: inherit
 
   win-vs2022-cuda12_8-py3-build:
     name: win-vs2022-cuda12.8-py3
@@ -208,22 +209,22 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-rocm-py3_10-test:
-    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor distributed/test_c10d_common distributed/test_c10d_nccl"
-    secrets: inherit
+  # linux-jammy-rocm-py3_10-test:
+  #   if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
+  #   permissions:
+  #     id-token: write
+  #     contents: read
+  #   name: linux-jammy-rocm-py3.10
+  #   uses: ./.github/workflows/_rocm-test.yml
+  #   needs:
+  #     - linux-jammy-rocm-py3_10-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-rocm-py3.10
+  #     docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+  #     tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor distributed/test_c10d_common distributed/test_c10d_nccl"
+  #   secrets: inherit
 
   inductor-build:
     name: inductor-build
@@ -236,21 +237,21 @@ jobs:
     secrets: inherit
 
   # Test cross-compiled models with Windows libs extracted from wheel
-  cross-compile-linux-test:
-    name: cross-compile-linux-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-build
-      - get-label-type
-      - win-vs2022-cuda12_8-py3-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
-      test-matrix: |
-        { include: [
-          { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
-        ]}
-    secrets: inherit
+  # cross-compile-linux-test:
+  #   name: cross-compile-linux-test
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - linux-jammy-cuda12_8-py3_10-gcc11-build
+  #     - get-label-type
+  #     - win-vs2022-cuda12_8-py3-build
+  #   with:
+  #     build-environment: linux-jammy-cuda12.8-py3.10-gcc11
+  #     docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
+  #     test-matrix: |
+  #       { include: [
+  #         { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
+  #       ]}
+  #   secrets: inherit
 
   verify-cachebench-cpu-build:
     name: verify-cachebench-cpu-build
@@ -266,17 +267,17 @@ jobs:
         ]}
     secrets: inherit
 
-  verify-cachebench-cpu-test:
-    name: verify-cachebench-cpu-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - verify-cachebench-cpu-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-py3.10-gcc11
-      docker-image: ${{ needs.verify-cachebench-cpu-build.outputs.docker-image }}
-      test-matrix: ${{ needs.verify-cachebench-cpu-build.outputs.test-matrix }}
-    secrets: inherit
+  # verify-cachebench-cpu-test:
+  #   name: verify-cachebench-cpu-test
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs:
+  #     - verify-cachebench-cpu-build
+  #     - target-determination
+  #   with:
+  #     build-environment: linux-jammy-py3.10-gcc11
+  #     docker-image: ${{ needs.verify-cachebench-cpu-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.verify-cachebench-cpu-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-jammy-py3-clang12-executorch-build:
     name: linux-jammy-py3-clang12-executorch
@@ -292,15 +293,15 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-py3-clang12-executorch-test:
-    name: linux-jammy-py3-clang12-executorch
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-py3-clang12-executorch-build
-    with:
-      build-environment: linux-jammy-py3-clang12-executorch
-      docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-jammy-py3-clang12-executorch-test:
+  #   name: linux-jammy-py3-clang12-executorch
+  #   uses: ./.github/workflows/_linux-test.yml
+  #   needs: linux-jammy-py3-clang12-executorch-build
+  #   with:
+  #     build-environment: linux-jammy-py3-clang12-executorch
+  #     docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
+  #   secrets: inherit
 
   linux-jammy-py3_10-gcc11-full-debug-build-only:
     name: linux-jammy-py3.10-gcc11-full-debug-build-only
@@ -308,7 +309,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: linux.2xlarge.memory
+      runner: linux.c7i.2xlarge
       build-environment: linux-jammy-py3.10-gcc11-full-debug-build-only
       docker-image-name: ci-image:pytorch-linux-jammy-py3.10-gcc11
     secrets: inherit

.github/workflows/vllm.yml

@@ -43,7 +43,7 @@ jobs:
       build-environment: linux-jammy-cuda12.8-py3.12-gcc11
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm
       cuda-arch-list: '8.0 8.9 9.0'
-      runner: linux.24xlarge.memory
+      runner: linux.c7i.2xlarge
       test-matrix: |
         { include: [
           { config: "vllm_basic_correctness_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
@@ -65,14 +65,14 @@ jobs:
         ]}
     secrets: inherit
 
-  vllm-test-sm89:
-      name: ci-vllm-test
-      uses: ./.github/workflows/_linux-test.yml
-      needs: [
-        torch-build,
-      ]
-      with:
-        build-environment: linux-jammy-cuda12.8-py3.12-gcc11
-        docker-image: ${{ needs.torch-build.outputs.docker-image }}
-        test-matrix: ${{ needs.torch-build.outputs.test-matrix }}
-      secrets: inherit
+  # vllm-test-sm89:
+  #     name: ci-vllm-test
+  #     uses: ./.github/workflows/_linux-test.yml
+  #     needs: [
+  #       torch-build,
+  #     ]
+  #     with:
+  #       build-environment: linux-jammy-cuda12.8-py3.12-gcc11
+  #       docker-image: ${{ needs.torch-build.outputs.docker-image }}
+  #       test-matrix: ${{ needs.torch-build.outputs.test-matrix }}
+  #     secrets: inherit

.github/workflows/xpu.yml

@@ -35,7 +35,7 @@ jobs:
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
       build-environment: linux-jammy-xpu-n-1-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-1-py3
-      runner: linux.c7i.12xlarge
+      runner: linux.c7i.2xlarge
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
@@ -56,7 +56,7 @@ jobs:
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
       build-environment: linux-noble-xpu-n-py3.10
       docker-image-name: ci-image:pytorch-linux-noble-xpu-n-py3
-      runner: linux.c7i.12xlarge
+      runner: linux.c7i.2xlarge
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 12, runner: "linux.idc.xpu" },
@@ -74,39 +74,39 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-noble-xpu-n-py3_10-test:
-    name: linux-noble-xpu-n-py3.10
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: linux-noble-xpu-n-py3_10-build
-    permissions:
-      id-token: write
-      contents: read
-    with:
-      build-environment: linux-noble-xpu-n-py3.10
-      docker-image: ${{ needs.linux-noble-xpu-n-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-noble-xpu-n-py3_10-build.outputs.test-matrix }}
-    secrets: inherit
+  # linux-noble-xpu-n-py3_10-test:
+  #   name: linux-noble-xpu-n-py3.10
+  #   uses: ./.github/workflows/_xpu-test.yml
+  #   needs: linux-noble-xpu-n-py3_10-build
+  #   permissions:
+  #     id-token: write
+  #     contents: read
+  #   with:
+  #     build-environment: linux-noble-xpu-n-py3.10
+  #     docker-image: ${{ needs.linux-noble-xpu-n-py3_10-build.outputs.docker-image }}
+  #     test-matrix: ${{ needs.linux-noble-xpu-n-py3_10-build.outputs.test-matrix }}
+  #   secrets: inherit
 
-  windows-xpu-n-1-build:
-    if: github.repository_owner == 'pytorch'
-    name: win-vs2022-xpu-n-1-py3
-    uses: ./.github/workflows/_win-build.yml
-    with:
-      build-environment: win-vs2022-xpu-n-1-py3
-      cuda-version: cpu
-      use-xpu: true
-      xpu-version: '2025.1'
-      vc-year: '2022'
-    secrets: inherit
+  # windows-xpu-n-1-build:
+  #   if: github.repository_owner == 'pytorch'
+  #   name: win-vs2022-xpu-n-1-py3
+  #   uses: ./.github/workflows/_win-build.yml
+  #   with:
+  #     build-environment: win-vs2022-xpu-n-1-py3
+  #     cuda-version: cpu
+  #     use-xpu: true
+  #     xpu-version: '2025.1'
+  #     vc-year: '2022'
+  #   secrets: inherit
 
-  windows-xpu-n-build:
-    if: github.repository_owner == 'pytorch'
-    name: win-vs2022-xpu-n-py3
-    uses: ./.github/workflows/_win-build.yml
-    with:
-      build-environment: win-vs2022-xpu-n-py3
-      cuda-version: cpu
-      use-xpu: true
-      xpu-version: '2025.2'
-      vc-year: '2022'
-    secrets: inherit
+  # windows-xpu-n-build:
+  #   if: github.repository_owner == 'pytorch'
+  #   name: win-vs2022-xpu-n-py3
+  #   uses: ./.github/workflows/_win-build.yml
+  #   with:
+  #     build-environment: win-vs2022-xpu-n-py3
+  #     cuda-version: cpu
+  #     use-xpu: true
+  #     xpu-version: '2025.2'
+  #     vc-year: '2022'
+  #   secrets: inherit

aten/src/ATen/LegacyBatchedTensorImpl.h

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

aten/src/ATen/TensorIndexing.cpp

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

aten/src/ATen/TensorNames.cpp

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

aten/src/ATen/TensorUtils.cpp

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

aten/src/ATen/Version.cpp

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

aten/src/ATen/code_template.h

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

aten/src/ATen/core/Dimname.cpp

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

aten/src/ATen/core/Range.cpp

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

aten/src/ATen/core/Tensor.cpp

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

aten/src/ATen/core/TensorAccessor.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <torch/headeronly/core/TensorAccessor.h>
 #include <c10/macros/Macros.h>
 #include <c10/util/ArrayRef.h>
 #include <c10/util/Deprecated.h>
@@ -11,252 +12,37 @@
 
 namespace at {
 
-// The PtrTraits argument to the TensorAccessor/GenericPackedTensorAccessor
-// is used to enable the __restrict__ keyword/modifier for the data
-// passed to cuda.
-template <typename T>
-struct DefaultPtrTraits {
-  typedef T* PtrType;
-};
-
+using torch::headeronly::DefaultPtrTraits;
 #if defined(__CUDACC__) || defined(__HIPCC__)
-template <typename T>
-struct RestrictPtrTraits {
-  typedef T* __restrict__ PtrType;
-};
+  using torch::headeronly::RestrictPtrTraits;
 #endif
 
-// TensorAccessorBase and TensorAccessor are used for both CPU and CUDA tensors.
-// For CUDA tensors it is used in device code (only). This means that we restrict ourselves
-// to functions and types available there (e.g. IntArrayRef isn't).
-
-// The PtrTraits argument is only relevant to cuda to support `__restrict__` pointers.
 template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
-class TensorAccessorBase {
-public:
-  typedef typename PtrTraits<T>::PtrType PtrType;
+using TensorAccessorBase = torch::headeronly::detail::TensorAccessorBase<c10::IntArrayRef, T, N, PtrTraits, index_t>;
 
-  C10_HOST_DEVICE TensorAccessorBase(
-      PtrType data_,
-      const index_t* sizes_,
-      const index_t* strides_)
-      : data_(data_), sizes_(sizes_), strides_(strides_) {}
-  C10_HOST IntArrayRef sizes() const {
-    return IntArrayRef(sizes_,N);
-  }
-  C10_HOST IntArrayRef strides() const {
-    return IntArrayRef(strides_,N);
-  }
-  C10_HOST_DEVICE index_t stride(index_t i) const {
-    return strides_[i];
-  }
-  C10_HOST_DEVICE index_t size(index_t i) const {
-    return sizes_[i];
-  }
-  C10_HOST_DEVICE PtrType data() {
-    return data_;
-  }
-  C10_HOST_DEVICE const PtrType data() const {
-    return data_;
-  }
-protected:
-  PtrType data_;
-  const index_t* sizes_;
-  const index_t* strides_;
-};
-
-// The `TensorAccessor` is typically instantiated for CPU `Tensor`s using
-// `Tensor.accessor<T, N>()`.
-// For CUDA `Tensor`s, `GenericPackedTensorAccessor` is used on the host and only
-// indexing on the device uses `TensorAccessor`s.
 template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
-class TensorAccessor : public TensorAccessorBase<T,N,PtrTraits,index_t> {
-public:
-  typedef typename PtrTraits<T>::PtrType PtrType;
+using TensorAccessor = torch::headeronly::detail::TensorAccessor<c10::IntArrayRef, T, N, PtrTraits, index_t>;
 
-  C10_HOST_DEVICE TensorAccessor(
-      PtrType data_,
-      const index_t* sizes_,
-      const index_t* strides_)
-      : TensorAccessorBase<T, N, PtrTraits, index_t>(data_,sizes_,strides_) {}
+namespace detail {
 
-  C10_HOST_DEVICE TensorAccessor<T, N - 1, PtrTraits, index_t> operator[](index_t i) {
-    return TensorAccessor<T,N-1,PtrTraits,index_t>(this->data_ + this->strides_[0]*i,this->sizes_+1,this->strides_+1);
-  }
-
-  C10_HOST_DEVICE const TensorAccessor<T, N-1, PtrTraits, index_t> operator[](index_t i) const {
-    return TensorAccessor<T,N-1,PtrTraits,index_t>(this->data_ + this->strides_[0]*i,this->sizes_+1,this->strides_+1);
-  }
-};
-
-template<typename T, template <typename U> class PtrTraits, typename index_t>
-class TensorAccessor<T,1,PtrTraits,index_t> : public TensorAccessorBase<T,1,PtrTraits,index_t> {
-public:
-  typedef typename PtrTraits<T>::PtrType PtrType;
-
-  C10_HOST_DEVICE TensorAccessor(
-      PtrType data_,
-      const index_t* sizes_,
-      const index_t* strides_)
-      : TensorAccessorBase<T, 1, PtrTraits, index_t>(data_,sizes_,strides_) {}
-  C10_HOST_DEVICE T & operator[](index_t i) {
-    // NOLINTNEXTLINE(clang-analyzer-core.NullDereference)
-    return this->data_[this->strides_[0]*i];
-  }
-  C10_HOST_DEVICE const T & operator[](index_t i) const {
-    return this->data_[this->strides_[0]*i];
-  }
-};
-
-
-// GenericPackedTensorAccessorBase and GenericPackedTensorAccessor are used on for CUDA `Tensor`s on the host
-// and as
-// In contrast to `TensorAccessor`s, they copy the strides and sizes on instantiation (on the host)
-// in order to transfer them on the device when calling kernels.
-// On the device, indexing of multidimensional tensors gives to `TensorAccessor`s.
-// Use RestrictPtrTraits as PtrTraits if you want the tensor's data pointer to be marked as __restrict__.
-// Instantiation from data, sizes, strides is only needed on the host and std::copy isn't available
-// on the device, so those functions are host only.
-template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
-class GenericPackedTensorAccessorBase {
-public:
-  typedef typename PtrTraits<T>::PtrType PtrType;
-  C10_HOST GenericPackedTensorAccessorBase(
-      PtrType data_,
-      const index_t* sizes_,
-      const index_t* strides_)
-      : data_(data_) {
-    std::copy(sizes_, sizes_ + N, std::begin(this->sizes_));
-    std::copy(strides_, strides_ + N, std::begin(this->strides_));
-  }
-
-  // if index_t is not int64_t, we want to have an int64_t constructor
-  template <typename source_index_t, class = std::enable_if_t<std::is_same_v<source_index_t, int64_t>>>
-  C10_HOST GenericPackedTensorAccessorBase(
-      PtrType data_,
-      const source_index_t* sizes_,
-      const source_index_t* strides_)
-      : data_(data_) {
-    for (const auto i : c10::irange(N)) {
-      this->sizes_[i] = sizes_[i];
-      this->strides_[i] = strides_[i];
-    }
-  }
-
-  C10_HOST_DEVICE index_t stride(index_t i) const {
-    return strides_[i];
-  }
-  C10_HOST_DEVICE index_t size(index_t i) const {
-    return sizes_[i];
-  }
-  C10_HOST_DEVICE PtrType data() {
-    return data_;
-  }
-  C10_HOST_DEVICE const PtrType data() const {
-    return data_;
-  }
-protected:
-  PtrType data_;
-  // NOLINTNEXTLINE(*c-arrays*)
-  index_t sizes_[N];
-  // NOLINTNEXTLINE(*c-arrays*)
-  index_t strides_[N];
-  C10_HOST void bounds_check_(index_t i) const {
-    TORCH_CHECK_INDEX(
+template <size_t N, typename index_t>
+struct IndexBoundsCheck {
+    IndexBoundsCheck(index_t i) {
+      TORCH_CHECK_INDEX(
         0 <= i && i < index_t{N},
         "Index ",
         i,
         " is not within bounds of a tensor of dimension ",
         N);
-  }
+    }
 };
+}  // namespace detail
 
 template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
-class GenericPackedTensorAccessor : public GenericPackedTensorAccessorBase<T,N,PtrTraits,index_t> {
-public:
-  typedef typename PtrTraits<T>::PtrType PtrType;
-
-  C10_HOST GenericPackedTensorAccessor(
-      PtrType data_,
-      const index_t* sizes_,
-      const index_t* strides_)
-      : GenericPackedTensorAccessorBase<T, N, PtrTraits, index_t>(data_, sizes_, strides_) {}
-
-  // if index_t is not int64_t, we want to have an int64_t constructor
-  template <typename source_index_t, class = std::enable_if_t<std::is_same_v<source_index_t, int64_t>>>
-  C10_HOST GenericPackedTensorAccessor(
-      PtrType data_,
-      const source_index_t* sizes_,
-      const source_index_t* strides_)
-      : GenericPackedTensorAccessorBase<T, N, PtrTraits, index_t>(data_, sizes_, strides_) {}
-
-  C10_DEVICE TensorAccessor<T, N - 1, PtrTraits, index_t> operator[](index_t i) {
-    index_t* new_sizes = this->sizes_ + 1;
-    index_t* new_strides = this->strides_ + 1;
-    return TensorAccessor<T,N-1,PtrTraits,index_t>(this->data_ + this->strides_[0]*i, new_sizes, new_strides);
-  }
-
-  C10_DEVICE const TensorAccessor<T, N - 1, PtrTraits, index_t> operator[](index_t i) const {
-    const index_t* new_sizes = this->sizes_ + 1;
-    const index_t* new_strides = this->strides_ + 1;
-    return TensorAccessor<T,N-1,PtrTraits,index_t>(this->data_ + this->strides_[0]*i, new_sizes, new_strides);
-  }
-
-  /// Returns a PackedTensorAccessor of the same dimension after transposing the
-  /// two dimensions given. Does not actually move elements; transposition is
-  /// made by permuting the size/stride arrays. If the dimensions are not valid,
-  /// asserts.
-  C10_HOST GenericPackedTensorAccessor<T, N, PtrTraits, index_t> transpose(
-      index_t dim1,
-      index_t dim2) const {
-    this->bounds_check_(dim1);
-    this->bounds_check_(dim2);
-    GenericPackedTensorAccessor<T, N, PtrTraits, index_t> result(
-        this->data_, this->sizes_, this->strides_);
-    std::swap(result.strides_[dim1], result.strides_[dim2]);
-    std::swap(result.sizes_[dim1], result.sizes_[dim2]);
-    return result;
-  }
-};
-
-template<typename T, template <typename U> class PtrTraits, typename index_t>
-class GenericPackedTensorAccessor<T,1,PtrTraits,index_t> : public GenericPackedTensorAccessorBase<T,1,PtrTraits,index_t> {
-public:
-  typedef typename PtrTraits<T>::PtrType PtrType;
-  C10_HOST GenericPackedTensorAccessor(
-      PtrType data_,
-      const index_t* sizes_,
-      const index_t* strides_)
-      : GenericPackedTensorAccessorBase<T, 1, PtrTraits, index_t>(data_, sizes_, strides_) {}
-
-  // if index_t is not int64_t, we want to have an int64_t constructor
-  template <typename source_index_t, class = std::enable_if_t<std::is_same_v<source_index_t, int64_t>>>
-  C10_HOST GenericPackedTensorAccessor(
-      PtrType data_,
-      const source_index_t* sizes_,
-      const source_index_t* strides_)
-      : GenericPackedTensorAccessorBase<T, 1, PtrTraits, index_t>(data_, sizes_, strides_) {}
-
-  C10_DEVICE T & operator[](index_t i) {
-    return this->data_[this->strides_[0] * i];
-  }
-  C10_DEVICE const T& operator[](index_t i) const {
-    return this->data_[this->strides_[0]*i];
-  }
-
-  // Same as in the general N-dimensional case, but note that in the
-  // 1-dimensional case the returned PackedTensorAccessor will always be an
-  // identical copy of the original
-  C10_HOST GenericPackedTensorAccessor<T, 1, PtrTraits, index_t> transpose(
-      index_t dim1,
-      index_t dim2) const {
-    this->bounds_check_(dim1);
-    this->bounds_check_(dim2);
-    return GenericPackedTensorAccessor<T, 1, PtrTraits, index_t>(
-        this->data_, this->sizes_, this->strides_);
-  }
-};
+using GenericPackedTensorAccessorBase = torch::headeronly::detail::GenericPackedTensorAccessorBase<detail::IndexBoundsCheck<N, index_t>, T, N, PtrTraits, index_t>;
 
+template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
+using GenericPackedTensorAccessor = torch::headeronly::detail::GenericPackedTensorAccessor<TensorAccessor<T, N-1, PtrTraits, index_t>, detail::IndexBoundsCheck<N, index_t>, T, N, PtrTraits, index_t>;
 
 // Can't put this directly into the macro function args because of commas
 #define AT_X GenericPackedTensorAccessor<T, N, PtrTraits, index_t>

aten/src/ATen/core/TensorBase.h

@@ -245,6 +245,9 @@ class TORCH_API TensorBase {
   size_t weak_use_count() const noexcept {
     return impl_.weak_use_count();
   }
+  bool is_uniquely_owned() const noexcept {
+    return impl_.is_uniquely_owned();
+  }
 
   std::string toString() const;
 

aten/src/ATen/core/Vitals.cpp

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

aten/src/ATen/core/alias_info.h

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

aten/src/ATen/core/blob.h

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

aten/src/ATen/core/class_type.cpp

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

aten/src/ATen/core/class_type.h

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

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

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

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

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

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

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

aten/src/ATen/core/function_schema.cpp

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

aten/src/ATen/core/function_schema.h

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

aten/src/ATen/core/ivalue.cpp

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

aten/src/ATen/core/jit_type.h

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

aten/src/ATen/core/operator_name.cpp

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

aten/src/ATen/core/tensor_type.cpp

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

aten/src/ATen/core/type.cpp

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

aten/src/ATen/core/union_type.cpp

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

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

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

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

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

aten/src/ATen/cuda/CUDAContextLight.h

@@ -3,6 +3,7 @@
 
 #include <cstdint>
 #include <map>
+#include <shared_mutex>
 
 #include <cuda_runtime_api.h>
 #include <cusparse.h>
@@ -88,8 +89,13 @@ TORCH_CUDA_CPP_API cublasHandle_t getCurrentCUDABlasHandle();
 TORCH_CUDA_CPP_API cublasLtHandle_t getCurrentCUDABlasLtHandle();
 
 TORCH_CUDA_CPP_API void clearCublasWorkspaces();
-TORCH_CUDA_CPP_API std::map<std::tuple<void *, void *>, at::DataPtr>& cublas_handle_stream_to_workspace();
-TORCH_CUDA_CPP_API std::map<std::tuple<void *, void *>, at::DataPtr>& cublaslt_handle_stream_to_workspace();
+struct WorkspaceMapWithMutex {
+  std::map<std::tuple<void*, void*>, at::DataPtr> map;
+  std::shared_mutex mutex;
+};
+
+TORCH_CUDA_CPP_API WorkspaceMapWithMutex& cublas_handle_stream_to_workspace();
+TORCH_CUDA_CPP_API WorkspaceMapWithMutex& cublaslt_handle_stream_to_workspace();
 TORCH_CUDA_CPP_API size_t getChosenWorkspaceSize();
 TORCH_CUDA_CPP_API size_t getCUDABlasLtWorkspaceSize();
 TORCH_CUDA_CPP_API void* getCUDABlasLtWorkspace();

aten/src/ATen/cuda/CUDAGraph.cpp

@@ -175,17 +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/CublasHandlePool.cpp

@@ -99,7 +99,7 @@ void destroyCublasHandle(cublasHandle_t handle) {
 //   - Comments of @soumith copied from cuDNN handle pool implementation
 #ifdef NO_CUDNN_DESTROY_HANDLE
 #else
-    cublasDestroy(handle);
+  cublasDestroy(handle);
 #endif
 }
 
@@ -107,19 +107,27 @@ using CuBlasPoolType = DeviceThreadHandlePool<cublasHandle_t, createCublasHandle
 
 } // namespace
 
-std::map<std::tuple<void *, void *>, at::DataPtr>& cublas_handle_stream_to_workspace() {
-  static auto& instance = *new std::map<std::tuple<void *, void *>, at::DataPtr>;
+WorkspaceMapWithMutex& cublas_handle_stream_to_workspace() {
+  static auto& instance = *new WorkspaceMapWithMutex;
   return instance;
 }
 
-std::map<std::tuple<void *, void *>, at::DataPtr>& cublaslt_handle_stream_to_workspace() {
-  static auto& instance = *new std::map<std::tuple<void *, void *>, at::DataPtr>;
+WorkspaceMapWithMutex& cublaslt_handle_stream_to_workspace() {
+  static auto& instance = *new WorkspaceMapWithMutex;
   return instance;
 }
 
 void clearCublasWorkspaces() {
-  cublas_handle_stream_to_workspace().clear();
-  cublaslt_handle_stream_to_workspace().clear();
+  {
+    auto& workspace = cublas_handle_stream_to_workspace();
+    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
+    workspace.map.clear();
+  }
+  {
+    auto& workspace = cublaslt_handle_stream_to_workspace();
+    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
+    workspace.map.clear();
+  }
 }
 
 size_t parseChosenWorkspaceSize() {
@@ -233,6 +241,38 @@ at::DataPtr getNewCUDABlasLtWorkspace() {
   return c10::cuda::CUDACachingAllocator::get()->allocate(getCUDABlasLtWorkspaceSize());
 }
 
+void setWorkspaceForHandle(cublasHandle_t handle, c10::cuda::CUDAStream stream) {
+  cudaStream_t _stream = stream;
+  auto key = std::make_tuple(static_cast<void *>(handle), static_cast<void *>(_stream));
+
+  auto& workspace = cublas_handle_stream_to_workspace();
+
+  size_t workspace_size = getChosenWorkspaceSize();
+
+  // Fast path: check if workspace already exists
+  {
+    std::shared_lock<std::shared_mutex> lock(workspace.mutex);
+    auto workspace_it = workspace.map.find(key);
+    if (workspace_it != workspace.map.end()) {
+      TORCH_CUDABLAS_CHECK(cublasSetWorkspace(
+          handle, workspace_it->second.get(), workspace_size));
+      return;
+    }
+  }
+
+  // Slow path: allocate workspace outside the lock
+  auto new_workspace = getNewWorkspace();
+
+  // Insert with lock (double-check in case another thread inserted while we
+  // were allocating)
+  {
+    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
+    auto workspace_it = workspace.map.try_emplace(key, std::move(new_workspace)).first;
+    TORCH_CUDABLAS_CHECK(
+        cublasSetWorkspace(handle, workspace_it->second.get(), workspace_size));
+  }
+}
+
 void* getCUDABlasLtWorkspace() {
 #ifndef USE_ROCM
   static bool unified = c10::utils::check_env(TORCH_CUBLASLT_UNIFIED_WORKSPACE) == true;
@@ -241,8 +281,10 @@ void* getCUDABlasLtWorkspace() {
     auto stream = c10::cuda::getCurrentCUDAStream();
     cudaStream_t _stream = stream;
     auto key = std::make_tuple(static_cast<void *>(handle), static_cast<void *>(_stream));
-    auto workspace_it = at::cuda::cublas_handle_stream_to_workspace().find(key);
-    TORCH_INTERNAL_ASSERT(workspace_it != at::cuda::cublas_handle_stream_to_workspace().end());
+    auto& workspace = at::cuda::cublas_handle_stream_to_workspace();
+    std::shared_lock<std::shared_mutex> lock(workspace.mutex);
+    auto workspace_it = workspace.map.find(key);
+    TORCH_INTERNAL_ASSERT(workspace_it != workspace.map.end());
     return workspace_it->second.mutable_get();
   }
 #endif
@@ -250,11 +292,29 @@ void* getCUDABlasLtWorkspace() {
   auto stream = c10::cuda::getCurrentCUDAStream();
   cudaStream_t _stream = stream;
   auto key = std::make_tuple(static_cast<void *>(handle), static_cast<void *>(_stream));
-  auto workspace_it = cublaslt_handle_stream_to_workspace().find(key);
-  if (workspace_it == cublaslt_handle_stream_to_workspace().end()) {
-    workspace_it = cublaslt_handle_stream_to_workspace().insert(workspace_it, {key, getNewCUDABlasLtWorkspace()});
+
+  auto& workspace = cublaslt_handle_stream_to_workspace();
+
+  // Fast path: check if workspace already exists
+  {
+    std::shared_lock<std::shared_mutex> lock(workspace.mutex);
+    auto workspace_it = workspace.map.find(key);
+    if (workspace_it != workspace.map.end()) {
+      return workspace_it->second.mutable_get();
+    }
+  }
+
+  // Slow path: allocate workspace outside the lock
+  auto new_workspace = getNewCUDABlasLtWorkspace();
+
+  // Insert with lock (double-check in case another thread inserted while we
+  // were allocating)
+  {
+    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
+    auto workspace_it =
+          workspace.map.try_emplace(key, std::move(new_workspace)).first;
+    return workspace_it->second.mutable_get();
   }
-  return workspace_it->second.mutable_get();
 }
 
 cublasHandle_t getCurrentCUDABlasHandle() {
@@ -298,13 +358,8 @@ cublasHandle_t getCurrentCUDABlasHandle() {
   // will allocate memory dynamically (even if they're cheap) outside
   // PyTorch's CUDA caching allocator. It's possible that CCA used up
   // all the memory and cublas's cudaMallocAsync will return OOM
-  cudaStream_t _stream = stream;
-  auto key = std::make_tuple(static_cast<void *>(handle), static_cast<void *>(_stream));
-  auto workspace_it = cublas_handle_stream_to_workspace().find(key);
-  if (workspace_it == cublas_handle_stream_to_workspace().end()) {
-    workspace_it = cublas_handle_stream_to_workspace().insert(workspace_it, {key, getNewWorkspace()});
-  }
-  TORCH_CUDABLAS_CHECK(cublasSetWorkspace(handle, workspace_it->second.get(), getChosenWorkspaceSize()));
+  setWorkspaceForHandle(handle, stream);
+
 #if !defined(USE_ROCM)
   // On CUDA >= 11, and architecture >= Ampere, cuBLAS can use TF32 to speedup
   // FP32 data type calculations based on the value of the allow_tf32 flag.

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

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

aten/src/ATen/cuda/jiterator.cu

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

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

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

aten/src/ATen/cudnn/Descriptors.cpp

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

aten/src/ATen/functorch/DynamicLayer.cpp

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

aten/src/ATen/functorch/TensorWrapper.cpp

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

aten/src/ATen/miopen/Descriptors.cpp

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

aten/src/ATen/mps/MPSProfiler.h

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

aten/src/ATen/mps/MPSProfiler.mm

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

aten/src/ATen/native/ConvUtils.h

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

aten/src/ATen/native/Convolution.cpp

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

aten/src/ATen/native/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()
+  // and tensor1_larger iff tensor2.dim() > tensor1.dim(9
   const auto t1 = tensor1_larger ? MaybeOwned<Tensor>::borrowed(tensor1)
                                  : MaybeOwned<Tensor>::owned(tensor2.mT());
   const int64_t dim_t1 = t1->dim();
@@ -1948,11 +1948,20 @@ static bool should_fold(const Tensor& tensor1, const Tensor& tensor2, bool has_o
     return false;
   }
 
-  // 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) {
+  // 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
     return true;
   }
 

aten/src/ATen/native/SpectralOps.cpp

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

aten/src/ATen/native/TensorAdvancedIndexing.cpp

@@ -1087,7 +1087,8 @@ TORCH_IMPL_FUNC(index_copy_out)
     result.copy_(self);
 
   // See Note [Enabling Deterministic Operations]
-  if (result.is_cuda() && globalContext().deterministicAlgorithms()) {
+  if ((result.is_cuda() || result.is_xpu()) &&
+      globalContext().deterministicAlgorithms()) {
     torch::List<std::optional<Tensor>> indices;
     indices.resize(dim + 1);
     indices.set(dim, index);

aten/src/ATen/native/TensorCompare.cpp

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

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

@@ -296,7 +296,7 @@ template <typename scalar_t, typename res_scalar_t = scalar_t>
 bool launchGemmAndBiasCublasLt(
     // args contains result which is modified
     cublasCommonArgs& args,
-    const Tensor& self,
+    const std::optional<Tensor>& self,
     const Scalar& alpha,
     Activation activation = Activation::None
 ) {
@@ -304,12 +304,8 @@ bool launchGemmAndBiasCublasLt(
   // or when it can be squeezed to 1D.
   // self_ptr == nullptr implies ignore bias epilogue
   // and use standard gemm-like API.
-  const auto* self_ptr = [&]() -> auto {
-    if (self.dim() == 1 || self.squeeze().dim() == 1) {
-      return self.const_data_ptr<scalar_t>();
-    }
-    return static_cast<const scalar_t*>(nullptr);
-  }();
+  const auto* self_ptr = self.has_value() ? self.value().const_data_ptr<scalar_t>() : static_cast<const scalar_t*>(nullptr);
+
 
   const auto tuning_ctx = at::cuda::tunable::getTuningContext();
   if (tuning_ctx->IsTunableOpEnabled()) {
@@ -392,35 +388,30 @@ Tensor& addmm_out_cuda_impl(Tensor& result, const Tensor& self, const Tensor& ma
   bool disable_addmm_cuda_lt = persistent_disable_addmm_cuda_lt || disable_addmm_cuda_lt_override;
   #ifdef USE_ROCM
   // Conditioned on the device index, which is not persistent
-  disable_addmm_cuda_lt = isGloballyDisabledAddmmCudaLt(self.device()) || disable_addmm_cuda_lt;
+  disable_addmm_cuda_lt = disable_addmm_cuda_lt || isGloballyDisabledAddmmCudaLt(self.device());
   #endif
   // Condition on the input
-  disable_addmm_cuda_lt = !isInputCompliesAddmmCudaLt(result, self, mat1, mat2, beta, alpha, activation) || disable_addmm_cuda_lt;
-  // }
+  disable_addmm_cuda_lt = disable_addmm_cuda_lt || !isInputCompliesAddmmCudaLt(result, self, mat1, mat2, beta, alpha, activation);
 
   at::ScalarType scalar_type = mat1.scalar_type();
   bool is_float_output_with_half_input = (scalar_type == at::ScalarType::Half || scalar_type == at::ScalarType::BFloat16) && result.scalar_type() == at::ScalarType::Float;
 
+  #ifdef USE_ROCM
+  disable_addmm_cuda_lt = disable_addmm_cuda_lt || is_float_output_with_half_input;
+  #endif
+
+  bool use_bias_ptr_lt = (self.dim() == 1) && !disable_addmm_cuda_lt;
+  // for float output with half input cublasLT with bias produces wrong results
+  use_bias_ptr_lt &= !is_float_output_with_half_input;
+
   // Handle result/self shapes
   if (!result.is_same(self)) {
     at::native::resize_output(result, {mat1.sizes()[0], mat2.sizes()[1]});
 
-    // We use bias ptr in the Lt path only when bias is 1D
-    const auto use_bias_ptr_lt = (self.dim() == 1) && !disable_addmm_cuda_lt;
-    const auto self_maybe_expanded = [&]() -> c10::MaybeOwned<Tensor> {
-      if (!use_bias_ptr_lt) {
-        // We do expand self even before
-        // check for beta != 0.0 to make sure that
-        // test_sparse_csr.py::TestSparseCSRCUDA::test_addmm_errors_*
-        // runs green.
-        return expand_size(self, result.sizes(), "addmm");
-      }
-      return c10::MaybeOwned<Tensor>::borrowed(self);
-    }();
-    // We do not copy bias only when we need the bias ptr
+      // We do not copy bias only when we need the bias ptr
     if (beta.toComplexDouble() != 0.0 && !use_bias_ptr_lt) {
       // NOTE: self should broadcast over result
-      at::native::copy_(result, *self_maybe_expanded);
+      at::native::copy_(result, *expand_size(self, result.sizes(), "addmm"));
     }
   }
 
@@ -468,7 +459,7 @@ Tensor& addmm_out_cuda_impl(Tensor& result, const Tensor& self, const Tensor& ma
         scalar_type,
         "addmm_cuda_lt",
         [&] {
-          lt_success = launchGemmAndBiasCublasLt<scalar_t, float>(args, self, alpha, activation);
+          lt_success = launchGemmAndBiasCublasLt<scalar_t, float>(args, use_bias_ptr_lt ? std::make_optional(self) : std::nullopt, alpha, activation);
         }
       );
       #endif
@@ -480,7 +471,7 @@ Tensor& addmm_out_cuda_impl(Tensor& result, const Tensor& self, const Tensor& ma
         scalar_type,
         "addmm_cuda_lt",
         [&] {
-          lt_success = launchGemmAndBiasCublasLt<scalar_t>(args, self, alpha, activation);
+          lt_success = launchGemmAndBiasCublasLt<scalar_t>(args, use_bias_ptr_lt ? std::make_optional(self) : std::nullopt, alpha, activation);
         }
       );
     } // end is_float_output_with_half_input
@@ -936,7 +927,7 @@ Tensor _int_mm_cuda(const Tensor& self, const Tensor& mat2) {
   return _int_mm_out_cuda(self, mat2, result);
 }
 
-static void baddbmm_bmm_out_dtype_checks(const Tensor& batch1, const Tensor& batch2, const Scalar& beta, const Scalar& alpha, const at::ScalarType out_dtype, bool is_bmm, const std::optional<Tensor>& self_baddbmm = std::nullopt) {
+static void baddbmm_bmm_out_dtype_checks(const Tensor& batch1, const Tensor& batch2, const Scalar& beta, const Scalar& alpha, const at::ScalarType out_dtype, const std::optional<Tensor>& self_baddbmm = std::nullopt) {
   // ref ATen/native/LinearAlgebra.cpp common_checks_baddbmm_bmm
   TORCH_CHECK(batch1.dim() == 3, "batch1 must be a 3D tensor");
   TORCH_CHECK(batch2.dim() == 3, "batch2 must be a 3D tensor");
@@ -960,7 +951,7 @@ static void baddbmm_bmm_out_dtype_checks(const Tensor& batch1, const Tensor& bat
     (out_dtype == at::ScalarType::Float && (batch1.scalar_type() == at::ScalarType::Half || batch1.scalar_type() == at::ScalarType::BFloat16)),
     "out_dtype must be the same as input dtype or fp32 for fp16/bf16 inputs");
 
-  if (!is_bmm && self_baddbmm.has_value()) {
+  if (self_baddbmm.has_value()) {
     const auto& self = self_baddbmm.value();
     TORCH_CHECK(self.dim() == 3, "self must be a 3D tensor");
     TORCH_CHECK(self.sizes() == output_size, "self must have the same shape as the output");
@@ -968,15 +959,12 @@ static void baddbmm_bmm_out_dtype_checks(const Tensor& batch1, const Tensor& bat
 }
 
 Tensor _bmm_dtype_cuda(const Tensor& batch1, const Tensor& batch2, const at::ScalarType out_dtype) {
-  IntArrayRef batch1_sizes = batch1.sizes();
-  IntArrayRef batch2_sizes = batch2.sizes();
-
-  Tensor out = at::empty({batch1_sizes[0], batch1_sizes[1], batch2_sizes[2]}, batch1.options().dtype(out_dtype));
+  Tensor out = at::empty({batch1.size(0), batch1.size(1), batch2.size(2)}, batch1.options().dtype(out_dtype));
   return _bmm_out_dtype_cuda(batch1, batch2, out_dtype, out);
 }
 
 Tensor& _bmm_out_dtype_cuda(const Tensor& batch1, const Tensor& batch2, const at::ScalarType out_dtype, Tensor &out) {
-  baddbmm_bmm_out_dtype_checks(batch1, batch2, 0.0, 1.0, out_dtype, true);
+  baddbmm_bmm_out_dtype_checks(batch1, batch2, 0.0, 1.0, out_dtype);
   Scalar beta(0.0);
   Scalar alpha(1.0);
   {
@@ -988,14 +976,16 @@ Tensor& _bmm_out_dtype_cuda(const Tensor& batch1, const Tensor& batch2, const at
 }
 
 Tensor _baddbmm_dtype_cuda(const Tensor& self, const Tensor& batch1, const Tensor& batch2, const at::ScalarType out_dtype, const Scalar& beta, const Scalar& alpha) {
-  // We need to copy the tensor
-  Tensor out = self.clone().to(self.options().dtype(out_dtype));
-
-  return _baddbmm_out_dtype_cuda(out, batch1, batch2, out_dtype, beta, alpha, out);
+  TORCH_CHECK(self.scalar_type() == out_dtype || self.scalar_type() == batch1.dtype(),
+  "self dtype must match either out_dtype or batch1 dtype");
+  Tensor out = at::empty({batch1.size(0), batch1.size(1), batch2.size(2)}, batch1.options().dtype(out_dtype));
+  return _baddbmm_out_dtype_cuda(self, batch1, batch2, out_dtype, beta, alpha, out);
 }
 
 Tensor& _baddbmm_out_dtype_cuda(const Tensor& self, const Tensor& batch1, const Tensor& batch2, const at::ScalarType out_dtype, const Scalar& beta, const Scalar& alpha, Tensor &out) {
-  baddbmm_bmm_out_dtype_checks(batch1, batch2, beta, alpha, out_dtype, false, self);
+  baddbmm_bmm_out_dtype_checks(batch1, batch2, beta, alpha, out_dtype, out);
+  // We need to copy the tensor
+  out.copy_(self);
   {
     NoNamesGuard guard;
     baddbmm_out_cuda_impl(out, out, batch1, batch2, beta, alpha);
@@ -1030,24 +1020,27 @@ Tensor& _mm_dtype_out_cuda(const Tensor& self, const Tensor& mat2, const at::Sca
 }
 
 Tensor _addmm_dtype_cuda(const Tensor& self, const Tensor& mat1, const Tensor& mat2, const at::ScalarType out_dtype, const Scalar& beta, const Scalar& alpha) {
-  Tensor result = at::empty(self.sizes(), self.options().dtype(out_dtype));
+  TORCH_CHECK(mat1.dim() == 2, "mat1 must be a matrix, got ", mat1.dim(), "-D tensor");
+  TORCH_CHECK(mat2.dim() == 2, "mat2 must be a matrix, got ", mat2.dim(), "-D tensor");
+  Tensor result = at::empty({mat1.size(0), mat2.size(1)}, self.options().dtype(out_dtype));
   return _addmm_dtype_out_cuda(self, mat1, mat2, out_dtype, beta, alpha, result);
 }
 
 Tensor& _addmm_dtype_out_cuda(const Tensor& self, const Tensor& mat1, const Tensor& mat2, const at::ScalarType out_dtype, const Scalar& beta, const Scalar& alpha, Tensor &out) {
-  TORCH_CHECK(self.scalar_type() == mat2.scalar_type(), "self and mat2 must have the same dtype, but got ", self.scalar_type(), " and ", mat2.scalar_type());
-  TORCH_CHECK(mat1.scalar_type() == mat2.scalar_type(), "mat1 and mat2 must have the same dtype, but got ", mat1.scalar_type(), " and ", mat2.scalar_type());
+// repeat dimensionality checks for direct calls to `out` overload
   TORCH_CHECK(mat1.dim() == 2, "mat1 must be a matrix, got ", mat1.dim(), "-D tensor");
   TORCH_CHECK(mat2.dim() == 2, "mat2 must be a matrix, got ", mat2.dim(), "-D tensor");
   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], ")");
+  TORCH_CHECK(mat1.scalar_type() == mat2.scalar_type(), "mat1 and mat2 must have the same dtype, but got ", mat1.scalar_type(), " and ", mat2.scalar_type());
+  TORCH_CHECK(out_dtype == mat1.scalar_type() ||
+  (out_dtype == at::ScalarType::Float && (mat1.scalar_type() == at::ScalarType::Half || mat1.scalar_type() == at::ScalarType::BFloat16)),
+  "out_dtype must be the same as input dtype or fp32 for fp16/bf16 inputs");
 
   TORCH_CHECK(out_dtype == out.scalar_type(), "out_dtype must be the same as the dtype of the provided out tensor");
-  TORCH_CHECK(out_dtype == self.scalar_type() ||
-    (out_dtype == at::ScalarType::Float && (self.scalar_type() == at::ScalarType::Half || self.scalar_type() == at::ScalarType::BFloat16)),
-    "out_dtype must be the same as input dtype or fp32 for fp16/bf16 inputs");
-  TORCH_CHECK(out_dtype == out.scalar_type(), "out_dtype must be the same as the dtype of the provided out tensor");
+  TORCH_CHECK(out_dtype == self.scalar_type() || self.scalar_type() == mat1.scalar_type(),
+    "self dtype must match either out_dtype or mat1 dtype");
 
   addmm_out_cuda_impl(out, self, mat1, mat2, beta, alpha);
 

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

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

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

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

@@ -2,18 +2,250 @@
 #include <ATen/Dispatch.h>
 #include <ATen/native/DispatchStub.h>
 #include <ATen/native/cuda/Loops.cuh>
+#include <ATen/native/cuda/JitLoops.cuh>
+#include <ATen/native/cuda/jit_utils.h>
+#include <ATen/native/cuda/ScanUtils.cuh>
 #include <ATen/native/TensorIterator.h>
 #include <ATen/native/BinaryOps.h>
 #include <ATen/OpMathType.h>
 #include <c10/util/MathConstants.h>
+#include <c10/util/complex.h>
+
+#include <cmath>
+#include <limits>
 
 // NOTE: CUDA on Windows requires that the enclosing function
 // of a __device__ lambda not have internal linkage.
 
 namespace at::native {
 
+// custom min and max to be used in logaddexp for  complex arguments
+template <typename scalar_t, bool min>
+__host__ __device__ c10::complex<scalar_t> _logaddexp_minmax(const c10::complex<scalar_t>& x, const c10::complex<scalar_t>& y) {
+  scalar_t xr = std::real(x);
+  scalar_t yr = std::real(y);
+  if (::isnan(yr) || (::isnan(std::imag(y)))) {
+    return y;
+  } else if (::isnan(xr) || (::isnan(std::imag(x)))) {
+    return x;
+  } else if (min) { // min
+    return (xr < yr) ? x : y;
+  } else { // max
+    return (xr >= yr) ? x : y;
+  }
+}
+
+template <typename scalar_t>
+__host__ __device__ scalar_t _log_add_exp_helper(const scalar_t& x, const scalar_t& y) {
+  // Reference : https://www.tensorflow.org/api_docs/python/tf/math/cumulative_logsumexp
+  // Using the original expression: `at::_isnan(y) ? y : std::min(x, y)` causes an error in ROCM
+  const auto isnan_x = at::_isnan(x);
+  const auto isnan_y = at::_isnan(y);
+  scalar_t min = isnan_y ? y : (isnan_x ? x : std::min(x, y));
+  scalar_t max = isnan_y ? y : (isnan_x ? x : std::max(x, y));
+  if (min != max || ::isfinite(min)) {
+    // nan will be propagated here
+    return ::log1p(std::exp(min - max)) + max;
+  } else {
+    // special case to correctly handle infinite cases
+    return x;
+  }
+}
+
+template <typename scalar_t>
+__host__ __device__ c10::complex<scalar_t> _fast_build_exp(const c10::complex<scalar_t>& x) {
+  // complex exponential function, but implemented manually to get fast compilation time
+  // this function only handles the case where the x is finite (not inf nor nan)
+  const auto xreal = std::real(x);
+  const auto ximag = std::imag(x);
+  const auto exp_x_abs = std::exp(xreal);
+  auto exp_x_real = exp_x_abs * std::cos(ximag);
+  auto exp_x_imag = exp_x_abs * std::sin(ximag);
+  return {exp_x_real, exp_x_imag};
+}
+
+template <typename scalar_t>
+__host__ __device__ c10::complex<scalar_t> _fast_build_exp_inf(const c10::complex<scalar_t>& x) {
+  // complex exponential function, but implemented manually to get fast compilation time
+  // this function only handles the case where the real part of x is infinite
+  const auto ximag = std::imag(x);
+  constexpr auto exp_x_abs = std::numeric_limits<scalar_t>::infinity();
+  if (!::isfinite(ximag)) {  // add this to make consitent with std::exp(x+yi)
+    return {exp_x_abs, std::numeric_limits<scalar_t>::quiet_NaN()};
+  }
+  const auto sin = std::sin(ximag);
+  const auto cos = std::cos(ximag);
+  // special case if the angle is exactly the multiple of pi/2
+  auto exp_x_real = (cos == 0) ? (scalar_t)0.0 : exp_x_abs * cos;
+  auto exp_x_imag = (sin == 0) ? (scalar_t)0.0 : exp_x_abs * sin;
+  return {exp_x_real, exp_x_imag};
+}
+
+template <typename scalar_t>
+__host__ __device__ c10::complex<scalar_t> _log_add_exp_helper(const c10::complex<scalar_t>& x, const c10::complex<scalar_t>& y) {
+  c10::complex<scalar_t> min = _logaddexp_minmax<scalar_t, /*min=*/true>(x, y);
+  c10::complex<scalar_t> max = _logaddexp_minmax<scalar_t, /*min=*/false>(x, y);
+  scalar_t min_real = std::real(min);
+  scalar_t max_real = std::real(max);
+
+  if (::isnan(min_real) || ::isnan(std::imag(min))) {
+    // handling the "infectious" NaNs
+    return {std::numeric_limits<scalar_t>::quiet_NaN(), std::numeric_limits<scalar_t>::quiet_NaN()};
+  }
+  else if ((!::isfinite(min_real)) && (min_real == max_real)) {
+    if (min_real < 0) {
+      // handle the -inf case, the imaginary part here does not really matter as the exp(value)
+      // will be around 0.0 and the angle (i.e. the imaginary part) cannot be determined.
+      // It does not matter if we're taking the exp of this value
+      return min;
+    } else {
+      // handle the +inf case, we don't need the special precision for log1p for small values
+      // and to avoid producing nan in case of real(max) == real(min) == +inf
+      const auto exp_min = _fast_build_exp_inf(min);
+      const auto exp_max = _fast_build_exp_inf(max);
+      return ::log1p(exp_min + exp_max - 1);  // log1p(x - 1) builds faster than log
+    }
+  } else {
+    const auto minmax = min - max;
+    c10::complex<scalar_t> exp_minmax;
+    if (!::isfinite(minmax.real())) {
+        exp_minmax = minmax.real() < 0 ? c10::complex<scalar_t>{0.0, 0.0} : _fast_build_exp_inf(minmax);
+    } else {
+        exp_minmax = _fast_build_exp(minmax);
+    }
+    return ::log1p(exp_minmax) + max;
+  }
+}
+
+// Complex logaddexp jiterator string
+const auto logaddexp_complex_string = jiterator_stringify(
+    template<typename T>
+    std::complex<T> log1p(const std::complex<T>& z)
+    {
+      using complex_t = std::complex<T>;
+      T x = z.real();
+      T y = z.imag();
+      T zabs = abs(z);
+      T theta = atan2(y, x + T(1));
+      if (zabs < 0.5) {
+          T r = x * (T(2) + x) + y * y;
+          if (r == 0) { // handle underflow
+              return complex_t(x, theta);
+          }
+          return complex_t(T(0.5) * std::log1p(r), theta);
+      } else {
+          T z0 = std::hypot(x + 1, y);
+          return complex_t(log(z0), theta);
+      }
+    }
+
+    // separated _logaddexp_minmax into 2 different functions for jiterator_string
+    template <typename T>
+    std::complex<T> logaddexp_min(const std::complex<T>& x, const std::complex<T>& y) {
+        T xr = x.real();
+        T yr = y.real();
+        if (isnan(yr) || isnan(y.imag())) {
+            return y;
+        } else if (isnan(xr) || isnan(x.imag())) {
+            return x;
+        } else {
+            return (xr < yr) ? x : y;
+        }
+    }
+
+    template <typename T>
+    std::complex<T> logaddexp_max(const std::complex<T>& x, const std::complex<T>& y) {
+        T xr = x.real();
+        T yr = y.real();
+        if (isnan(yr) || isnan(y.imag())) {
+            return y;
+        } else if (isnan(xr) || isnan(x.imag())) {
+            return x;
+        } else {
+            return (xr >= yr) ? x : y;
+        }
+    }
+
+    template <typename T>
+    std::complex<T> fast_build_exp(const std::complex<T>& x) {
+        const auto xreal = x.real();
+        const auto ximag = x.imag();
+        const auto exp_x_abs = exp(xreal);
+        auto exp_x_real = exp_x_abs * cos(ximag);
+        auto exp_x_imag = exp_x_abs * sin(ximag);
+        return std::complex<T>(exp_x_real, exp_x_imag);
+    }
+
+    template <typename T>
+    std::complex<T> fast_build_exp_inf(const std::complex<T>& x) {
+        using complex_t = std::complex<T>;
+        const auto ximag = x.imag();
+        const T exp_x_abs = INFINITY;
+        if (!isfinite(ximag)) {
+            return complex_t(exp_x_abs, NAN);
+        }
+        const auto sin_val = sin(ximag);
+        const auto cos_val = cos(ximag);
+        auto exp_x_real = (cos_val == T(0)) ? T(0) : exp_x_abs * cos_val;
+        auto exp_x_imag = (sin_val == T(0)) ? T(0) : exp_x_abs * sin_val;
+        return complex_t(exp_x_real, exp_x_imag);
+    }
+
+    template <typename complex_t>
+    complex_t logaddexp_complex(complex_t x, complex_t y) {
+        using T = typename complex_t::value_type;
+        complex_t min_val = logaddexp_min(x, y);
+        complex_t max_val = logaddexp_max(x, y);
+        T min_real = min_val.real();
+        T max_real = max_val.real();
+
+        if (isnan(min_real) || isnan(min_val.imag())) {
+            return complex_t(NAN, NAN);
+        }
+        else if ((!isfinite(min_real)) && (min_real == max_real)) {
+            if (min_real < T(0)) {
+                return min_val;
+            } else {
+                const auto exp_min = fast_build_exp_inf<T>(min_val);
+                const auto exp_max = fast_build_exp_inf<T>(max_val);
+                return log1p(exp_min + exp_max - complex_t(1, 0));
+            }
+        } else {
+            const auto minmax = min_val - max_val;
+            complex_t exp_minmax;
+            if (!isfinite(minmax.real())) {
+                exp_minmax = (minmax.real() < T(0)) ? complex_t(0, 0) : fast_build_exp_inf<T>(minmax);
+            } else {
+                exp_minmax = fast_build_exp<T>(minmax);
+            }
+            return log1p(exp_minmax) + max_val;
+        }
+    }
+);
+
+constexpr char logaddexp_complex_name[] = "logaddexp_complex";
 void logaddexp_kernel_cuda(TensorIteratorBase& iter) {
-  AT_DISPATCH_FLOATING_TYPES_AND2(
+  if (at::isComplexType(iter.dtype())) {
+#if AT_USE_JITERATOR()
+    AT_DISPATCH_COMPLEX_TYPES_AND(at::ScalarType::ComplexHalf, iter.dtype(), "logaddexp_cuda", [&]() {
+      jitted_gpu_kernel<
+          /*name=*/logaddexp_complex_name,
+          /*return_dtype=*/scalar_t,
+          /*common_dtype=*/scalar_t,
+          /*arity=*/2>(iter, logaddexp_complex_string);
+    });
+#else
+    AT_DISPATCH_COMPLEX_TYPES_AND(at::ScalarType::ComplexHalf, iter.dtype(), "logaddexp_cuda", [&]() {
+      using opmath_t = at::opmath_type<scalar_t>;
+      gpu_kernel(iter, [] GPU_LAMBDA (scalar_t a_, scalar_t b_) -> scalar_t {
+        const auto a = static_cast<opmath_t>(a_);
+        const auto b = static_cast<opmath_t>(b_);
+        return static_cast<scalar_t>(_log_add_exp_helper(a, b));
+      });
+    });
+#endif
+  } else {
+    AT_DISPATCH_FLOATING_TYPES_AND2(
       ScalarType::BFloat16, ScalarType::Half,
       iter.dtype(), "logaddexp_cuda",
       [&]() {
@@ -29,6 +261,7 @@ void logaddexp_kernel_cuda(TensorIteratorBase& iter) {
           }
         });
       });
+  }
 }
 
 void logaddexp2_kernel_cuda(TensorIteratorBase& iter) {

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

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

aten/src/ATen/native/cuda/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;
@@ -1096,6 +1101,19 @@ _scaled_mxfp8_mxfp8(
   return _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, false /* use_fast_accum */, out);
 }
 
+void
+_check_mxfp4_support() {
+#ifndef USE_ROCM
+  auto dprops = at::cuda::getCurrentDeviceProperties();
+  // Only on B200 GPUs
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    // B200 = 10.0, B300 = 10.3
+    dprops->major == 10,
+    "MXFP4 scaling only supported in CUDA for B200/B300"
+  );
+#endif
+}
+
 
 Tensor&
 _scaled_mxfp4_mxfp4(
@@ -1108,6 +1126,7 @@ _scaled_mxfp4_mxfp4(
 #if defined(_WIN32) || (!defined(USE_ROCM) && !defined(USE_FBGEMM_GENAI))
   TORCH_CHECK_NOT_IMPLEMENTED(false, "MXFP4 scaling supported on ROCM and CUDA+FBGEMM_GENAI only");
 #else
+  _check_mxfp4_support();
   // Restrictions:
   // A, B are FP4, scales are e8m0, A: shape K//32, B: K, N//32
   TORCH_CHECK_VALUE(mat_a.scalar_type() == at::kFloat4_e2m1fn_x2 && mat_b.scalar_type() == at::kFloat4_e2m1fn_x2, "mat_a and mat_b must be fp4 types, got: ",

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -91,25 +91,30 @@ static auto& lib = mps::MetalShaderLibrary::getBundledLibrary();
 #include <ATen/native/mps/Repeat_metallib.h>
 #endif
 
-template <typename index_t>
-void computeRepeatIndices(const index_t* repeat_ptr,
-                          const int64_t* cumsum_ptr,
-                          index_t* result_ptr,
-                          int64_t size,
-                          int64_t result_size) {
-  id<MTLBuffer> repeatBuffer = reinterpret_cast<id<MTLBuffer>>(repeat_ptr);
-  id<MTLBuffer> cumsumBuffer = reinterpret_cast<id<MTLBuffer>>(cumsum_ptr);
-  id<MTLBuffer> resultBuffer = reinterpret_cast<id<MTLBuffer>>(result_ptr);
-  TORCH_CHECK(repeatBuffer && cumsumBuffer && resultBuffer);
-
+Tensor repeat_interleave_mps(const Tensor& repeat, std::optional<int64_t> output_size) {
+  TORCH_CHECK(repeat.dim() == 1, "repeat_interleave only accept 1D vector as repeat");
   std::string scalar_type;
-  if constexpr (std::is_same_v<index_t, int32_t>) {
+  if (repeat.scalar_type() == kInt) {
     scalar_type = "int32_t";
-  } else if constexpr (std::is_same_v<index_t, int64_t>) {
+  } else if (repeat.scalar_type() == kLong) {
     scalar_type = "int64_t";
   } else {
-    TORCH_CHECK(false, "repeat_interleave: unsupported indexing data type");
+    TORCH_CHECK(false, "repeats has to be Long or Int tensor");
   }
+  if (repeat.size(0) == 0) {
+    return at::empty_like(repeat, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
+  }
+  Tensor repeat_ = repeat.contiguous();
+  Tensor cumsum = repeat.cumsum(0);
+  int64_t total = 0;
+  if (output_size.has_value()) {
+    total = output_size.value();
+  } else {
+    total = cumsum[-1].item<int64_t>();
+    TORCH_CHECK((repeat >= 0).all().item<uint8_t>(), "repeats can not be negative");
+  }
+
+  auto result = at::empty({total}, repeat.options());
 
   MPSStream* mpsStream = getCurrentMPSStream();
   dispatch_sync(mpsStream->queue(), ^() {
@@ -121,20 +126,13 @@ void computeRepeatIndices(const index_t* repeat_ptr,
       getMPSProfiler().beginProfileKernel(pipelineState, "repeat_interleave:" + scalar_type, false);
 
       [computeEncoder setComputePipelineState:pipelineState];
-      mps::mtl_setArgs(computeEncoder, repeatBuffer, cumsumBuffer, resultBuffer, size);
-      mps::mtl_dispatch1DJob(computeEncoder, pipelineState, size);
+      mps::mtl_setArgs(computeEncoder, repeat_, cumsum, result, repeat.size(0));
+      mps::mtl_dispatch1DJob(computeEncoder, pipelineState, repeat.size(0));
 
       getMPSProfiler().endProfileKernel(pipelineState);
     }
   });
-}
-
-Tensor repeat_interleave_mps(const Tensor& repeat, std::optional<int64_t> output_size) {
-  Tensor output;
-  AT_DISPATCH_INDEX_TYPES(repeat.scalar_type(), "repeat_interleave_mps", [&]() {
-    output = repeat_interleave_common<index_t, computeRepeatIndices<index_t>>(repeat, output_size);
-  });
-  return output;
+  return result;
 }
 
 } // namespace at::native

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

@@ -5,6 +5,7 @@
 #include <ATen/native/Resize.h>
 #include <ATen/native/TensorCompare.h>
 #include <ATen/native/mps/OperationUtils.h>
+#include <algorithm>
 
 #ifndef AT_PER_OPERATOR_HEADERS
 #include <ATen/Functions.h>
@@ -89,13 +90,21 @@ static void check_min_max_dims(const OptionalTensorRef clamp_opt, const Tensor&
     auto clamp_shape = clamp_opt->sizes();
     auto input_shape = input_t.sizes();
 
-    TORCH_CHECK(num_clamp_dims <= num_input_dims,
-                op_name + ": clamp tensor number of dims must not be greater than that of input tensor")
+    if (num_clamp_dims > num_input_dims) {
+      auto leading_dims = num_clamp_dims - num_input_dims;
+      for (int64_t i = 0; i < leading_dims; ++i) {
+        TORCH_CHECK(clamp_shape[i] == 1,
+                    op_name + ": clamp tensor leading shape must be 1 to broadcast with input tensor");
+      }
+    }
 
-    for (int i = 0; i < num_clamp_dims; i++)
+    auto clamp_idx = num_clamp_dims - 1;
+    auto input_idx = num_input_dims - 1;
+    auto common_dims = std::min(num_clamp_dims, num_input_dims);
+    for (int64_t i = 0; i < common_dims; ++i)
       // One of the indices is allowed to be 1; will be handled by broadcast
-      TORCH_CHECK(clamp_shape[num_clamp_dims - 1 - i] == input_shape[num_input_dims - 1 - i] ||
-                      clamp_shape[num_clamp_dims - 1 - i] == 1 || input_shape[num_input_dims - 1 - i] == 1,
+      TORCH_CHECK(clamp_shape[clamp_idx - i] == input_shape[input_idx - i] || clamp_shape[clamp_idx - i] == 1 ||
+                      input_shape[input_idx - i] == 1,
                   op_name + ": clamp tensor trailing shape must match input tensor")
   }
 }
@@ -136,9 +145,6 @@ static void clamp_tensor_out_mps(const Tensor& input_t,
 
   auto result_type = output_t.scalar_type();
 
-  IntArrayRef new_min_shape;
-  IntArrayRef new_max_shape;
-
   auto num_min_dims = min_opt->dim();
   auto num_max_dims = max_opt->dim();
   auto num_input_dims = input_t.dim();
@@ -146,24 +152,32 @@ static void clamp_tensor_out_mps(const Tensor& input_t,
   std::vector<int64_t> new_min_arr(num_input_dims);
   std::vector<int64_t> new_max_arr(num_input_dims);
 
-  if (has_min && num_min_dims < num_input_dims) {
-    fill_new_shape(num_input_dims, num_min_dims, new_min_arr.data(), min_opt->sizes());
-    new_min_shape = IntArrayRef(new_min_arr);
-  }
-
-  if (has_max && num_max_dims < num_input_dims) {
-    fill_new_shape(num_input_dims, num_max_dims, new_max_arr.data(), max_opt->sizes());
-    new_max_shape = IntArrayRef(new_max_arr);
-  }
-
   Tensor min_opt_tensor;
   Tensor max_opt_tensor;
 
+  auto reshape_clamp_tensor = [&](const OptionalTensorRef clamp_tensor_ref,
+                                  int64_t num_clamp_dims,
+                                  std::vector<int64_t>& new_shape_storage) -> Tensor {
+    IntArrayRef clamp_shape = clamp_tensor_ref->sizes();
+    bool requires_view = false;
+
+    if (num_clamp_dims > num_input_dims) {
+      clamp_shape = clamp_shape.slice(num_clamp_dims - num_input_dims);
+      requires_view = true;
+    } else if (num_clamp_dims < num_input_dims) {
+      fill_new_shape(num_input_dims, num_clamp_dims, new_shape_storage.data(), clamp_shape);
+      clamp_shape = IntArrayRef(new_shape_storage);
+      requires_view = true;
+    }
+
+    return requires_view ? (*clamp_tensor_ref).view(clamp_shape) : *clamp_tensor_ref;
+  };
+
   if (has_min) {
-    min_opt_tensor = (num_min_dims < num_input_dims) ? (*min_opt).view(new_min_shape) : *min_opt;
+    min_opt_tensor = reshape_clamp_tensor(min_opt, num_min_dims, new_min_arr);
   }
   if (has_max) {
-    max_opt_tensor = (num_max_dims < num_input_dims) ? (*max_opt).view(new_max_shape) : *max_opt;
+    max_opt_tensor = reshape_clamp_tensor(max_opt, num_max_dims, new_max_arr);
   }
 
   @autoreleasepool {
@@ -244,8 +258,8 @@ 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

@@ -4225,7 +4225,7 @@
     MTIA: mm_out_mtia
     MPS: mm_out_mps
     XPU: mm_out_xpu
-    SparseCPU, SparseCUDA: _sparse_mm_out
+    SparseCPU, SparseCUDA, SparseMPS: _sparse_mm_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: _sparse_csr_mm_out
 
 - func: mm.dtype(Tensor self, Tensor mat2, ScalarType out_dtype) -> Tensor

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

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

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

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