update triton commit hash

This reverts commit f49833de54450b03b808a5b9ad774ce14ff2c8a2.

Reverted https://github.com/pytorch/pytorch/pull/167231 on behalf of https://github.com/yangw-dev due to the diff breaks tests internally  ([comment](https://github.com/pytorch/pytorch/pull/167231#issuecomment-3555183647))

.ci/aarch64_linux/README.md

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

.ci/aarch64_linux/aarch64_ci_build.sh

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

.ci/aarch64_linux/aarch64_ci_setup.sh

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

.ci/aarch64_linux/aarch64_wheel_ci_build.py

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

.ci/aarch64_linux/build_aarch64_wheel.py

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

.ci/aarch64_linux/embed_library.py

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

.ci/docker/build.sh

@@ -125,10 +125,10 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks)
     CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=9
+    GCC_VERSION=11
     VISION=yes
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
@@ -146,16 +146,6 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=9
-    VISION=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    TRITON=yes
-    ;;
   pytorch-linux-jammy-py3-clang12-onnx)
     ANACONDA_PYTHON_VERSION=3.10
     CLANG_VERSION=12
@@ -188,7 +178,7 @@ case "$tag" in
     fi
     GCC_VERSION=11
     VISION=yes
-    ROCM_VERSION=7.0
+    ROCM_VERSION=7.1
     NINJA_VERSION=1.9.0
     TRITON=yes
     KATEX=yes

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-bfeb066872bc1e8b2d2bc0a3b295b99dd77206e7
+9946bd3634beb4b0709641fb677fcda247f453f2

.ci/docker/common/install_rocm.sh

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

.ci/docker/common/install_rocm_magma.sh

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

.ci/docker/requirements-ci.txt

@@ -402,3 +402,6 @@ scikit-build==0.18.1
 pyre-extensions==0.0.32
 tabulate==0.9.0
 #Description: These package are needed to build FBGEMM and torchrec on PyTorch CI
+
+Jinja2==3.1.6
+#Description: required for torch.distributed.debug

.ci/lumen_cli/cli/lib/core/vllm/vllm_test.py

@@ -84,7 +84,6 @@ class VllmTestRunner(BaseRunner):
         self.VLLM_TEST_WHLS_REGEX = [
             "xformers/*.whl",
             "vllm/vllm*.whl",
-            "flashinfer-python/flashinfer*.whl",
         ]
 
     def prepare(self):

.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/test.sh

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

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-2d82dc5caa336d179d9b46ac4a0fb8c43d84c5cc
+617079d944b0e72632311c30ae2bbdf1168b901e

.github/ci_commit_pins/xla.txt

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

.github/ci_configs/vllm/Dockerfile

@@ -1,4 +1,4 @@
-ARG CUDA_VERSION=12.8.1
+ARG CUDA_VERSION=12.9.1
 ARG PYTHON_VERSION=3.12
 
 # BUILD_BASE_IMAGE: used to setup python build xformers, and vllm wheels, It can be replaced with a different base image from local machine,
@@ -124,7 +124,7 @@ RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
     git clone https://github.com/facebookresearch/xformers.git
 
     pushd xformers
-    git checkout v0.0.32.post2
+    git checkout v0.0.33.post1
     git submodule update --init --recursive
     python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose
     popd
@@ -256,7 +256,7 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
-# Install build and runtime dependencies, this is needed for flashinfer install
+# Install build and runtime dependencies
 COPY requirements/build.txt requirements/build.txt
 COPY use_existing_torch.py use_existing_torch.py
 RUN python3 use_existing_torch.py
@@ -294,33 +294,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
 
-# Build FlashInfer from source
-ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-
-# TODO(elainewy): remove this once vllm commit is updated, and install flashinfer from pip
-# see https://github.com/pytorch/pytorch/pull/165274#issuecomment-3408531784
-ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
-ARG FLASHINFER_GIT_REF="v0.2.14.post1"
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    git clone --depth 1 --recursive --shallow-submodules \
-        --branch ${FLASHINFER_GIT_REF} \
-        ${FLASHINFER_GIT_REPO} flashinfer \
-    && echo "Building FlashInfer with AOT for arches: ${torch_cuda_arch_list}" \
-    && cd flashinfer \
-    && python3 -m flashinfer.aot \
-    && python3 -m build --no-isolation --wheel --outdir ../wheels/flashinfer \
-    && cd .. \
-    && rm -rf flashinfer
-
-# Install FlashInfer
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system wheels/flashinfer/*.whl --verbose
-
 # Logging to confirm the torch versions
-RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm\|^flashinfer' > build_summary.txt
+RUN pip freeze | grep -E 'torch|xformers|vllm'
+RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm' > build_summary.txt
 ################### VLLM INSTALLED IMAGE ####################
 
 
@@ -331,4 +307,3 @@ FROM scratch as export-wheels
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
 COPY --from=vllm-base /workspace/build_summary.txt /wheels/build_summary.txt
-COPY --from=vllm-base /workspace/wheels/flashinfer /wheels/flashinfer-python

.github/labeler.yml

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

.github/scripts/generate_binary_build_matrix.py

@@ -50,6 +50,7 @@ CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "12.9-aarch64", "13.0-aar
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
     "12.6": (
+        "cuda-bindings==12.9.4; platform_system == 'Linux' | "
         "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | "
         "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | "
         "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | "
@@ -67,6 +68,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
     ),
     "12.8": (
+        "cuda-bindings==12.9.4; platform_system == 'Linux' | "
         "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | "
         "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | "
         "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | "
@@ -84,6 +86,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
     ),
     "12.9": (
+        "cuda-bindings==12.9.4; platform_system == 'Linux' | "
         "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | "
         "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | "
         "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | "
@@ -101,6 +104,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
     ),
     "13.0": (
+        "cuda-bindings==13.0.3; platform_system == 'Linux' | "
         "nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | "
         "nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | "
         "nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | "

.github/scripts/prepare_vllm_wheels.sh

@@ -88,7 +88,7 @@ repackage_wheel() {
 ${PYTHON_EXECUTABLE} -mpip install wheel==0.45.1
 
 pushd externals/vllm/wheels
-for package in xformers flashinfer-python vllm; do
+for package in xformers vllm; do
   repackage_wheel $package
 done
 popd

.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/attention_op_microbenchmark.yml

@@ -23,7 +23,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     with:
       runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '8.0 9.0'
       test-matrix: |
@@ -39,7 +39,7 @@ jobs:
     needs: attn-microbenchmark-build
     with:
       timeout-minutes: 500
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image: ${{ needs.attn-microbenchmark-build.outputs.docker-image }}
       test-matrix: ${{ needs.attn-microbenchmark-build.outputs.test-matrix }}
     secrets: inherit
@@ -51,7 +51,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     with:
       runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      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'
       test-matrix: |
@@ -66,7 +66,7 @@ jobs:
     needs: opmicrobenchmark-build-b200
     with:
       timeout-minutes: 500
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-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

.github/workflows/docker-builds.yml

@@ -52,8 +52,7 @@ jobs:
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11,
           pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9,
+          pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
           pytorch-linux-jammy-py3.10-clang12,
           pytorch-linux-jammy-py3.11-clang12,
@@ -75,7 +74,8 @@ jobs:
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter,
-          pytorch-linux-jammy-py3-clang12-executorch,
+          # TODO: Re-enable me when docker pin update happens
+          # pytorch-linux-jammy-py3-clang12-executorch,
           pytorch-linux-jammy-py3.12-triton-cpu,
           pytorch-linux-noble-riscv64-py3.12-gcc14
         ]

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

@@ -6,10 +6,9 @@ on:
     branches: [main, release]
     types:
       - completed
-  workflow_dispatch:
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name }}
+  group: ${{ github.workflow }}-${{ github.event.workflow_run.head_branch }}
   cancel-in-progress: true
 
 permissions:
@@ -50,9 +49,10 @@ jobs:
       matrix:
         runner: [linux.rocm.gfx942.docker-cache]
         docker-image: [
-          "${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3 }}",
-          "${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-noble-rocm-n-py3 }}",
-          "${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3-benchmarks }}"
+          "${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3 }}"
+          #"${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3 }}",
+          #"${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-noble-rocm-n-py3 }}",
+          #"${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3-benchmarks }}"
         ]
     runs-on: "${{ matrix.runner }}"
     steps:

.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml

@@ -132,7 +132,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -178,7 +178,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -224,7 +224,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -270,7 +270,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -381,7 +381,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -427,7 +427,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -473,7 +473,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -519,7 +519,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -630,7 +630,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -676,7 +676,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -722,7 +722,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -768,7 +768,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -879,7 +879,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -925,7 +925,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -971,7 +971,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1017,7 +1017,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1128,7 +1128,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1174,7 +1174,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1220,7 +1220,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1266,7 +1266,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1377,7 +1377,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1423,7 +1423,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1469,7 +1469,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1515,7 +1515,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1626,7 +1626,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1672,7 +1672,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1718,7 +1718,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1764,7 +1764,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/generated-linux-binary-manywheel-nightly.yml

@@ -127,7 +127,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_6-test:  # Testing
@@ -193,7 +193,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_8-test:  # Testing
@@ -259,7 +259,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_9-test:  # Testing
@@ -325,7 +325,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda13_0-test:  # Testing
@@ -793,7 +793,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_6-test:  # Testing
@@ -859,7 +859,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_8-test:  # Testing
@@ -925,7 +925,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_9-test:  # Testing
@@ -991,7 +991,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda13_0-test:  # Testing
@@ -1459,7 +1459,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_6-test:  # Testing
@@ -1525,7 +1525,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing
@@ -1591,7 +1591,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_9-test:  # Testing
@@ -1657,7 +1657,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda13_0-test:  # Testing
@@ -2125,7 +2125,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_6-test:  # Testing
@@ -2191,7 +2191,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_8-test:  # Testing
@@ -2257,7 +2257,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_9-test:  # Testing
@@ -2323,7 +2323,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda13_0-test:  # Testing
@@ -2791,7 +2791,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_6-test:  # Testing
@@ -2857,7 +2857,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_8-test:  # Testing
@@ -2923,7 +2923,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_9-test:  # Testing
@@ -2989,7 +2989,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda13_0-test:  # Testing
@@ -3457,7 +3457,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_6-test:  # Testing
@@ -3523,7 +3523,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_8-test:  # Testing
@@ -3589,7 +3589,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_9-test:  # Testing
@@ -3655,7 +3655,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda13_0-test:  # Testing
@@ -4123,7 +4123,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_6-test:  # Testing
@@ -4189,7 +4189,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_8-test:  # Testing
@@ -4255,7 +4255,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==12.9.4; platform_system == 'Linux' | nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_9-test:  # Testing
@@ -4321,7 +4321,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: cuda-bindings==13.0.3; platform_system == 'Linux' | nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda13_0-test:  # Testing

.github/workflows/inductor-micro-benchmark.yml

@@ -30,14 +30,14 @@ jobs:
       opt_out_experiments: lf
 
   build:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-build.yml
     needs:
       - get-default-label-prefix
     with:
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.0'
       test-matrix: |
         { include: [
@@ -46,11 +46,11 @@ jobs:
     secrets: inherit
 
   test:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
       timeout-minutes: 720

.github/workflows/inductor-perf-compare.yml

@@ -27,14 +27,14 @@ jobs:
       opt_out_experiments: lf
 
   build:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-build.yml
     needs:
       - get-default-label-prefix
     with:
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.0'
       test-matrix: |
         { include: [
@@ -47,11 +47,11 @@ jobs:
     secrets: inherit
 
   test:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
       # disable monitor in perf tests for more investigation

.github/workflows/inductor-perf-test-b200.yml

@@ -80,7 +80,7 @@ jobs:
       opt_out_experiments: lf
 
   build:
-    name: cuda12.8-py3.10-gcc9-sm100
+    name: cuda12.8-py3.10-gcc11-sm100
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
@@ -90,8 +90,8 @@ jobs:
       # from trunk. Also use a memory-intensive runner here because memory is
       # usually the bottleneck
       runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '10.0'
       test-matrix: |
         { include: [
@@ -104,12 +104,12 @@ jobs:
     secrets: inherit
 
   test-periodically:
-    name: cuda12.8-py3.10-gcc9-sm100
+    name: cuda12.8-py3.10-gcc11-sm100
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     if: github.event.schedule == '0 7 * * 1-6'
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
@@ -121,12 +121,12 @@ jobs:
     secrets: inherit
 
   test-weekly:
-    name: cuda12.8-py3.10-gcc9-sm100
+    name: cuda12.8-py3.10-gcc11-sm100
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     if: github.event.schedule == '0 7 * * 0'
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-maxautotune-true-freeze_autotune_cudagraphs-true-cudagraphs_low_precision-true
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
@@ -138,11 +138,11 @@ jobs:
     secrets: inherit
 
   test:
-    name: cuda12.8-py3.10-gcc9-sm100
+    name: cuda12.8-py3.10-gcc11-sm100
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
       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.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}

.github/workflows/inductor-perf-test-nightly-h100.yml

@@ -95,8 +95,8 @@ jobs:
       # from trunk. Also use a memory-intensive runner here because memory is
       # usually the bottleneck
       runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '9.0'
       test-matrix: |
         { include: [
@@ -132,7 +132,7 @@ jobs:
     needs: build
     if: github.event.schedule == '15 0 * * 1-6'
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
@@ -149,7 +149,7 @@ jobs:
     needs: build
     if: github.event.schedule == '0 7 * * 0'
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-maxautotune-true-freeze_autotune_cudagraphs-true-cudagraphs_low_precision-true
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
@@ -168,7 +168,7 @@ jobs:
     # needs one round of benchmark
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'pull_request' }}
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90
       dashboard-tag: training-${{ inputs.training || 'true' }}-inference-${{ inputs.inference || 'true' }}-default-${{ inputs.default || 'true' }}-dynamic-${{ inputs.dynamic || 'true' }}-cudagraphs-${{ inputs.cudagraphs || 'true' }}-cppwrapper-${{ inputs.cppwrapper || 'false' }}-aotinductor-${{ inputs.aotinductor || 'false' }}-maxautotune-${{ inputs.maxautotune || 'false' }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs || 'false' }}-cudagraphs_low_precision-${{ inputs.cudagraphs || 'false' }}
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}

.github/workflows/inductor-perf-test-nightly.yml

@@ -80,15 +80,15 @@ jobs:
       opt_out_experiments: lf
 
   build:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       # Every bit to make perf run faster helps
       runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.0'
       test-matrix: |
         { include: [
@@ -117,12 +117,12 @@ jobs:
     secrets: inherit
 
   test-nightly:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     if: github.event.schedule == '0 7 * * 1-6'
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
@@ -133,12 +133,12 @@ jobs:
     secrets: inherit
 
   test-weekly:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     if: github.event.schedule == '0 7 * * 0'
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-maxautotune-true-freeze_autotune_cudagraphs-true-cudagraphs_low_precision-true
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
@@ -150,12 +150,12 @@ jobs:
     secrets: inherit
 
   test:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     if: github.event_name == 'workflow_dispatch'
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       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.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}

.github/workflows/inductor-periodic.yml

@@ -37,8 +37,8 @@ jobs:
     needs: get-default-label-prefix
     with:
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.0;8.6'
       test-matrix: |
         { include: [
@@ -76,7 +76,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: periodic-dynamo-benchmarks-build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
       docker-image: ${{ needs.periodic-dynamo-benchmarks-build.outputs.docker-image }}
       test-matrix: ${{ needs.periodic-dynamo-benchmarks-build.outputs.test-matrix }}
     secrets: inherit
@@ -138,8 +138,8 @@ jobs:
       - get-default-label-prefix
     with:
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.0'
       test-matrix: |
         { include: [
@@ -153,7 +153,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-smoke-build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image: ${{ needs.inductor-smoke-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-smoke-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/inductor-unittest.yml

@@ -33,8 +33,8 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.6'
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
@@ -52,7 +52,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/inductor.yml

@@ -49,8 +49,8 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.6'
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
@@ -69,7 +69,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm86
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/operator_microbenchmark.yml

@@ -25,7 +25,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     with:
       runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '8.0 9.0'
       test-matrix: |
@@ -41,7 +41,7 @@ jobs:
     needs: opmicrobenchmark-build
     with:
       timeout-minutes: 500
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image: ${{ needs.opmicrobenchmark-build.outputs.docker-image }}
       test-matrix: ${{ needs.opmicrobenchmark-build.outputs.test-matrix }}
     secrets: inherit
@@ -53,7 +53,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     with:
       runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      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'
       test-matrix: |
@@ -68,7 +68,7 @@ jobs:
     needs: opmicrobenchmark-build-b200
     with:
       timeout-minutes: 500
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-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

.github/workflows/periodic.yml

@@ -90,6 +90,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc11
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
+      cuda-arch-list: 8.6
       test-matrix: |
         { include: [
           { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
@@ -97,7 +98,9 @@ jobs:
           { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
           { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
           { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
+          { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
+          { config: "multigpu", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.12xlarge.nvidia.gpu", owners: ["oncall:distributed"] },
+          { config: "multigpu", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.12xlarge.nvidia.gpu", owners: ["oncall:distributed"] },
         ]}
     secrets: inherit
 
@@ -113,40 +116,14 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.test-matrix }}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc9-build:
-    name: linux-jammy-cuda12.8-py3.10-gcc9
+  linux-jammy-cuda12_8-py3_10-gcc11-debug-build:
+    name: linux-jammy-cuda12.8-py3.10-gcc11-debug
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9
-      cuda-arch-list: 8.6
-      test-matrix: |
-        { include: [
-          { config: "multigpu", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.12xlarge.nvidia.gpu", owners: ["oncall:distributed"] },
-          { config: "multigpu", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.12xlarge.nvidia.gpu", owners: ["oncall:distributed"] },
-        ]}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc9-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc9
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-build.outputs.test-matrix }}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc9-debug-build:
-    name: linux-jammy-cuda12.8-py3.10-gcc9-debug
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-debug
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-debug
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: 8.9
       test-matrix: |
         { include: [
@@ -160,16 +137,16 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc9-debug-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc9-debug
+  linux-jammy-cuda12_8-py3_10-gcc11-debug-test:
+    name: linux-jammy-cuda12.8-py3.10-gcc11-debug
     uses: ./.github/workflows/_linux-test.yml
     needs:
-      - linux-jammy-cuda12_8-py3_10-gcc9-debug-build
+      - linux-jammy-cuda12_8-py3_10-gcc11-debug-build
       - target-determination
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-debug
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-debug-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-debug-build.outputs.test-matrix }}
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-debug
+      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-debug-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-debug-build.outputs.test-matrix }}
     secrets: inherit
 
   linux-jammy-cuda13_0-py3_10-gcc11-build:

.github/workflows/pull.yml

@@ -70,6 +70,7 @@ jobs:
           { config: "distributed", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
           { config: "distributed", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
           { config: "numpy_2_x", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "libtorch_agnostic_targetting", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
         ]}
     secrets: inherit
 
@@ -317,14 +318,14 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-build:
-    name: cuda12.8-py3.10-gcc9-sm75
+  linux-jammy-cuda12_8-py3_10-gcc11-inductor-build:
+    name: cuda12.8-py3.10-gcc11-sm75
     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-gcc9-sm75
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm75
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '7.5'
       test-matrix: |
         { include: [
@@ -332,14 +333,14 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
-    name: cuda12.8-py3.10-gcc9-sm75
+  linux-jammy-cuda12_8-py3_10-gcc11-inductor-test:
+    name: cuda12.8-py3.10-gcc11-sm75
     uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-inductor-build
+    needs: linux-jammy-cuda12_8-py3_10-gcc11-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 }}
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm75
+      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-inductor-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-inductor-build.outputs.test-matrix }}
     secrets: inherit
 
   linux-noble-xpu-n-py3_10-build:

.github/workflows/torchbench.yml

@@ -26,14 +26,14 @@ jobs:
       curr_ref_type: ${{ github.ref_type }}
 
   build:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-build.yml
     needs:
       - get-default-label-prefix
     with:
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.0'
       test-matrix: |
         { include: [
@@ -42,11 +42,11 @@ jobs:
     secrets: inherit
 
   test:
-    name: cuda12.8-py3.10-gcc9-sm80
+    name: cuda12.8-py3.10-gcc11-sm80
     uses: ./.github/workflows/_linux-test.yml
     needs: build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm80
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/trunk.yml

@@ -83,6 +83,7 @@ jobs:
           { config: "distributed", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.12xlarge.nvidia.gpu" },
           { config: "distributed", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.12xlarge.nvidia.gpu" },
           { config: "pr_time_benchmarks", shard: 1, num_shards: 1, runner: "linux.g4dn.metal.nvidia.gpu" },
+          { config: "libtorch_agnostic_targetting", shard: 1, num_shards: 1, runner: "linux.g4dn.metal.nvidia.gpu" },
         ]}
     secrets: inherit
 
@@ -230,8 +231,8 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      build-environment: linux-jammy-cuda12.8-py3.12-gcc9-sm80
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
+      build-environment: linux-jammy-cuda12.8-py3.12-gcc11-sm80
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11-inductor-benchmarks
       cuda-arch-list: '8.0'
     secrets: inherit
 
@@ -282,6 +283,7 @@ jobs:
     name: linux-jammy-py3-clang12-executorch
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
+    if: false # Has been broken for a while
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3-clang12-executorch

AGENTS.md

@@ -10,6 +10,7 @@
 - Do NOT run pre-commit, it is not setup
 - To run lint, run 'lintrunner -a' (which will autoapply changes)
 - Do NOT attempt to install dependencies, you do not have Internet access
+- Do NOT create summary files unless explicitly asked
 - When you are ready to make a PR, do exactly these steps:
   - git stash -u
   - git reset --hard $(cat /tmp/orig_work.txt) # NB: reset to the LOCAL branch, do NOT fetch

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/type.cpp

@@ -680,7 +680,7 @@ TORCH_API bool elementTypeCanBeInferredFromMembers(const TypePtr& elem_type) {
     return false;
   }
   if (elem_type->kind() == AnyType::Kind) {
-    // List of Any can contains heterogenous types
+    // List of Any can contains heterogeneous types
     return false;
   }
   return true;

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/CUDAEvent.h

@@ -238,11 +238,18 @@ private:
   }
 
   void moveHelper(CUDAEvent&& other) {
-    std::swap(flags_, other.flags_);
-    std::swap(is_created_, other.is_created_);
-    std::swap(was_recorded_, other.was_recorded_);
-    std::swap(device_index_, other.device_index_);
-    std::swap(event_, other.event_);
+    // Transfer ownership of all state from other to this
+    flags_ = other.flags_;
+    is_created_ = other.is_created_;
+    was_recorded_ = other.was_recorded_;
+    external_ = other.external_;
+    device_index_ = other.device_index_;
+    event_ = other.event_;
+
+    // Reset other to a valid empty state to prevent double-free
+    // The moved-from object must not attempt to destroy the event
+    other.is_created_ = false;
+    other.event_ = cudaEvent_t{};
   }
 };
 

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/native/LinearAlgebra.cpp

@@ -2909,13 +2909,26 @@ Tensor linalg_matrix_norm(
   // Check A, dim, and dtype
   _linalg_matrix_norm_checks(A, dim_, opt_dtype, /*low_precision*/abs_ord != 2.);
 
-  auto max_min = [ord, keepdim](const Tensor& A, int64_t dim) { return ord > 0 ? A.amax(dim, keepdim) : A.amin(dim, keepdim); };
+  auto max_min_wrapper = [ord, keepdim](const Tensor &A, int64_t dim) {
+    if (A.size(dim) == 0 && ord > 0) {
+      auto new_shape(DimVector(A.sizes()));
+      auto dim_ = maybe_wrap_dim(dim, A.dim());
+      if (keepdim) {
+        new_shape[dim_] = 1;
+      } else {
+        new_shape.erase(std::begin(new_shape) + dim_);
+      }
+      return at::zeros(new_shape, A.options());
+    } else {
+      return ord > 0 ? A.amax(dim, keepdim) : A.amin(dim, keepdim);
+    }
+  };
   if (abs_ord == 2.) {
     // Move dims to the end
     auto permutation = create_dim_backshift_permutation(dim_[0], dim_[1], A.dim());
 
     auto A_ = opt_dtype.has_value() ? A.to(*opt_dtype) : A;
-    auto result = max_min(at::linalg_svdvals(A_.permute(permutation)), -1);
+    auto result = max_min_wrapper(at::linalg_svdvals(A_.permute(permutation)), -1);
     if (keepdim) {
       auto permutation_reverse = create_reverse_permutation(std::move(permutation));
       result = result.unsqueeze(-1).permute(permutation_reverse);
@@ -2932,7 +2945,7 @@ Tensor linalg_matrix_norm(
     if (!keepdim && (dim_[0] < dim_[1])) {
       dim_[1]--;
     }
-    return max_min(at::linalg_vector_norm(A, 1., {dim_[0]}, keepdim, opt_dtype), dim_[1]);
+    return max_min_wrapper(at::linalg_vector_norm(A, 1., {dim_[0]}, keepdim, opt_dtype), dim_[1]);
   }
 }
 

aten/src/ATen/native/cpu/BinaryOpsKernel.cpp

@@ -813,8 +813,43 @@ void smooth_l1_kernel(TensorIteratorBase& iter, double beta) {
 }
 
 void huber_kernel(TensorIterator& iter, double delta) {
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-      kBFloat16, kHalf, iter.dtype(), "huber_cpu", [&]() {
+  // Special-case kHalf: compute in float for numerical stability
+  if (iter.dtype() == kHalf) {
+    const float delta_val(static_cast<float>(delta));
+    const Vectorized<float> delta_vec(static_cast<float>(delta));
+    const Vectorized<float> point_five_vec(static_cast<float>(0.5));
+    cpu_kernel_vec(
+      iter,
+      // scalar lambda: convert half -> float, compute in float, cast back to half
+      [&delta_val] (at::Half a, at::Half b) -> at::Half {
+        float af = static_cast<float>(a);
+        float bf = static_cast<float>(b);
+        float z = std::abs(af - bf);
+        float out = z < delta_val
+          ? 0.5f * z * z
+          : delta_val * (z - 0.5f * delta_val);
+        return static_cast<at::Half>(out);
+      },
+      [&delta_vec, &point_five_vec] (Vectorized<Half> a, Vectorized<Half> b) {
+        auto [a0, a1] = convert_half_float(a);
+        auto [b0, b1] = convert_half_float(b);
+        auto z = (a0 - b0).abs();
+        a0 = Vectorized<float>::blendv(
+          point_five_vec * z * z,
+          delta_vec * (z - point_five_vec * delta_vec),
+          z >= delta_vec);
+        z = (a1 - b1).abs();
+        a1 = Vectorized<float>::blendv(
+          point_five_vec * z * z,
+          delta_vec * (z - point_five_vec * delta_vec),
+          z >= delta_vec);
+        return convert_float_half(a0, a1);
+      }
+    );
+    return;
+  }
+  else {
+   AT_DISPATCH_FLOATING_TYPES_AND(kBFloat16, iter.dtype(), "huber_cpu", [&]() {
         using Vec = Vectorized<scalar_t>;
         const scalar_t delta_val(delta);
         const Vec delta_val_vec(delta_val);
@@ -835,6 +870,7 @@ void huber_kernel(TensorIterator& iter, double delta) {
                   z >= delta_val_vec);
             });
       });
+  }
 }
 
 void sigmoid_backward_kernel(TensorIteratorBase& iter) {

aten/src/ATen/native/cpu/moments_utils.h

@@ -46,8 +46,11 @@ C10_ALWAYS_INLINE void AddMomentsVec(
   const T c = n == 0 ? static_cast<T>(0) : static_cast<T>(m0_add) / static_cast<T>(n);
   const Vec c_vec(c);
   const Vec delta = m1_add - m1;
-  m1 += c_vec * delta;
-  m2 += m2_add + delta * delta * c_vec * Vec(static_cast<T>(m0));
+  const Vec m2_tmp = m2 + m2_add;
+  const Vec c_vec_delta = c_vec * delta;
+  const Vec m0_delta = delta * Vec(static_cast<T>(m0));
+  m1 = m1 + c_vec_delta;
+  m2 = fmadd(m0_delta, c_vec_delta, m2_tmp);
   m0 = n;
 }
 
@@ -65,9 +68,11 @@ UpdateMomentsVec(
   Vec m2_vec(0);
   for (const auto j : c10::irange(m0)) {
     const Vec x_vec = Vec::loadu(X_ptr + j * Vec::size());
+    const Vec tmpVec = c_vecs[j];
     const Vec delta_vec = x_vec - m1_vec;
-    m1_vec += delta_vec * c_vecs[j];
-    m2_vec += delta_vec * (x_vec - m1_vec);
+    m1_vec = fmadd(tmpVec, delta_vec, m1_vec);
+    const Vec tmpVec2 = x_vec - m1_vec;
+    m2_vec = fmadd(delta_vec, tmpVec2, m2_vec);
   }
   AddMomentsVec(m0, m1_vec, m2_vec, m0_stk0, m1_stk0, m2_stk0);
 }
@@ -89,13 +94,16 @@ UpdateMomentsVec(
   fVec m2_fvec0(0), m2_fvec1(0);
   for (const auto j : c10::irange(m0)) {
     const Vec x_bvec = Vec::loadu(X_ptr + j * Vec::size());
+    const fVec tmpVec = c_vecs[j];
     auto [x_fvec0, x_fvec1] = convert_to_float<T>(x_bvec);
     const fVec delta_fvec0 = x_fvec0 - m1_fvec0;
     const fVec delta_fvec1 = x_fvec1 - m1_fvec1;
-    m1_fvec0 += delta_fvec0 * c_vecs[j];
-    m1_fvec1 += delta_fvec1 * c_vecs[j];
-    m2_fvec0 += delta_fvec0 * (x_fvec0 - m1_fvec0);
-    m2_fvec1 += delta_fvec1 * (x_fvec1 - m1_fvec1);
+    m1_fvec0 = fmadd(delta_fvec0, tmpVec, m1_fvec0);
+    m1_fvec1 = fmadd(delta_fvec1, tmpVec, m1_fvec1);
+    const fVec delta_fvec2 = x_fvec0 - m1_fvec0;
+    const fVec delta_fvec3 = x_fvec1 - m1_fvec1;
+    m2_fvec0 = fmadd(delta_fvec0, delta_fvec2, m2_fvec0);
+    m2_fvec1 = fmadd(delta_fvec1, delta_fvec3, m2_fvec1);
   }
   AddMomentsVec(m0, m1_fvec0, m2_fvec0, m0_stk0, m1_stk0, m2_stk0);
   AddMomentsVec(m0, m1_fvec1, m2_fvec1, m0_stk0, m1_stk0, m2_stk0);

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/GroupMM.cu

@@ -346,8 +346,9 @@ void dispatch_bf16_grouped_kernel_on_tile_size(
   bool small = (M <= 128 || N <= 128);
   cudaDeviceProp* properties = at::cuda::getCurrentDeviceProperties();
   const bool sm10x = properties != nullptr && properties->major == 10;
+  const bool sm11x = properties != nullptr && properties->major == 11;
 
-  if (sm10x) {
+  if (sm10x || sm11x) {
     if (small){
       bf16bf16_grouped_gemm_impl_sm90_sm100<
         cutlass::arch::Sm100,

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

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

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

@@ -212,7 +212,7 @@ void nonzero_cuda_out_impl(const Tensor& self, Tensor& out) {
       std::nullopt /* memory format */
   );
   at::cuda::memcpy_and_sync(
-      (void*)pinned_num_nonzeros_h.const_data_ptr<int>(),
+      pinned_num_nonzeros_h.data_ptr<int>(),
       num_nonzeros.get(),
       sizeof(int) * num_chunks,
       cudaMemcpyDeviceToHost,

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

@@ -958,8 +958,9 @@ void dispatch_fp8_rowwise_kernel_on_sm(
   const bool sm89 = properties != nullptr && properties->major == 8 && properties->minor == 9;
   const bool sm9x = properties != nullptr && properties->major == 9;
   const bool sm10x = properties != nullptr && properties->major == 10;
+  const bool sm11x = properties != nullptr && properties->major == 11;
   const bool sm12x = properties != nullptr && properties->major == 12;
-  if (!(sm89 || sm9x || sm10x || sm12x)) {
+  if (!(sm89 || sm9x || sm10x || sm11x || sm12x)) {
     TORCH_CHECK(
         false, "Rowwise scaling is not currently supported on your device");
   }
@@ -968,7 +969,7 @@ void dispatch_fp8_rowwise_kernel_on_sm(
     dispatch_fp8_rowwise_kernel_on_cluster_size_and_transpose<
       /*ArchTag=*/cutlass::arch::Sm90,
       Types...>(XQ, WQ, x_scale, w_scale, bias, out);
-  } else if (sm10x) {
+  } else if (sm10x || sm11x) {
     dispatch_fp8_rowwise_kernel_on_cluster_size_and_transpose<
       /*ArchTag=*/cutlass::arch::Sm100,
       Types...>(XQ, WQ, x_scale, w_scale, bias, out);

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

@@ -1101,6 +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(
@@ -1113,6 +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/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/MetalShaderLibrary.h

@@ -147,6 +147,19 @@ class MetalShaderLibrary {
       const std::optional<c10::Scalar> alpha = std::nullopt,
       const std::optional<c10::ScalarType> scalar_arg_type = std::nullopt);
 
+  template <typename T>
+  void exec_unary_kernel_with_params(
+      TensorIteratorBase& iter,
+      const std::string& name,
+      T params,
+      const std::string& params_type_name);
+  template <typename T>
+  void exec_binary_kernel_with_params(
+      TensorIteratorBase& iter,
+      const std::string& name,
+      T params,
+      const std::string& params_type_name);
+
  protected:
   virtual MTLLibrary_t getLibrary();
   virtual MTLLibrary_t getLibrary(

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

@@ -7,10 +7,12 @@
 #include <ATen/Tensor.h>
 #include <ATen/TensorIterator.h>
 #include <ATen/Utils.h>
+#include <ATen/mps/MPSProfiler.h>
 #include <ATen/mps/MPSStream.h>
 #include <ATen/native/mps/MetalShaderLibrary.h>
 #include <ATen/native/mps/TensorFactory.h>
 #include <c10/core/ScalarType.h>
+#include <fmt/format.h>
 #include <torch/library.h>
 #include <unordered_map>
 
@@ -630,4 +632,147 @@ inline bool needsGather(const TensorBase& t) {
   return !is_macOS_15_0_or_newer && (!t.is_contiguous() || t.storage_offset());
 }
 
+template <typename T>
+void MetalShaderLibrary::exec_unary_kernel_with_params(TensorIteratorBase& iter,
+                                                       const std::string& name,
+                                                       T params,
+                                                       const std::string& params_type_name) {
+  using namespace at::mps;
+  // Decompose 64-bit tensor into 32-bit ones
+  if (!iter.can_use_32bit_indexing()) {
+    for (auto&& sub_iter : iter.with_32bit_indexing()) {
+      exec_unary_kernel_with_params(sub_iter, name, params, params_type_name);
+    }
+    return;
+  }
+
+  auto inputTensor = iter.input(0);
+  auto outputTensor = iter.output(0);
+  uint32_t length = iter.numel();
+  if (length == 0) {
+    return;
+  }
+  auto kernel_name = fmt::format("{}_{}_{}_{}{}",
+                                 name,
+                                 iter.is_contiguous() ? "dense" : "strided",
+                                 scalarToMetalTypeString(outputTensor),
+                                 scalarToMetalTypeString(inputTensor),
+                                 fmt::format("_{}", params_type_name));
+  @autoreleasepool {
+    auto cplState = getPipelineStateForFunc(kernel_name);
+
+    MPSStream* mpsStream = getCurrentMPSStream();
+    dispatch_sync(mpsStream->queue(), ^() {
+      auto computeEncoder = mpsStream->commandEncoder();
+
+      getMPSProfiler().beginProfileKernel(cplState, name, {inputTensor});
+
+      [computeEncoder setComputePipelineState:cplState];
+      bind_iter_tensors(computeEncoder, iter);
+      if (!iter.is_contiguous()) {
+        mtl_setArgs<2>(computeEncoder,
+                       outputTensor.sizes(),
+                       inputTensor.strides(),
+                       outputTensor.strides(),
+                       inputTensor.ndimension());
+      }
+      detail::mtl_setArg(computeEncoder, params, iter.is_contiguous() ? 2 : 6);
+      mtl_dispatch1DJob(computeEncoder, cplState, length);
+
+      getMPSProfiler().endProfileKernel(cplState);
+    });
+  }
+}
+
+template <typename T>
+void MetalShaderLibrary::exec_binary_kernel_with_params(TensorIteratorBase& iter,
+                                                        const std::string& name,
+                                                        T params,
+                                                        const std::string& params_type_name) {
+  using namespace mps;
+  // TODO: Figure a better place to downcast double scalars (probably in tensor iterator itself?)
+  // Right now running something like 1.0-torch.rand(5, device='mps') will create iterator with
+  // double as common dtype (because Python floating point are always 64-bit values)
+  TORCH_CHECK(iter.output().scalar_type() != at::kDouble, "float64 is not supported on MPS");
+
+  // Skip for empty iterators
+  if (iter.numel() == 0) {
+    return;
+  }
+
+  // Decompose 64-bit tensor into 32-bit ones
+  if (!iter.can_use_32bit_indexing()) {
+    for (auto&& sub_iter : iter.with_32bit_indexing()) {
+      exec_binary_kernel_with_params(sub_iter, name, params, params_type_name);
+    }
+    return;
+  }
+
+  auto convert_double_scalar = [](Tensor& t) {
+    if (t.dim() != 0) {
+      return;
+    }
+    if (t.scalar_type() == kDouble) {
+      t = t.to(kFloat);
+    } else if (t.scalar_type() == kComplexDouble) {
+      t = t.to(kComplexFloat);
+    }
+  };
+
+  Tensor input = iter.input(0);
+  Tensor other = iter.input(1);
+  Tensor out = iter.output();
+
+  convert_double_scalar(input);
+  convert_double_scalar(other);
+
+  MPSStream* mpsStream = getCurrentMPSStream();
+  const auto cast_needed = input.scalar_type() != other.scalar_type();
+  const auto suffix = iter.is_contiguous() ? "dense" : "strided";
+  // TODO: Implicitly pass both input and output types to non-cast kernels
+  const auto kernel_name = cast_needed
+      ? fmt::format("{}_{}_cast_{}_{}", name, suffix, scalarToMetalTypeString(out), params_type_name)
+      : fmt::format("{}_{}_{}_{}_{}",
+                    name,
+                    suffix,
+                    scalarToMetalTypeString(out),
+                    scalarToMetalTypeString(input),
+                    params_type_name);
+  dispatch_sync_with_rethrow(mpsStream->queue(), ^() {
+    @autoreleasepool {
+      auto computeEncoder = mpsStream->commandEncoder();
+      auto binaryPSO = getPipelineStateForFunc(kernel_name);
+      // this function call is a no-op if MPS Profiler is not enabled
+      getMPSProfiler().beginProfileKernel(binaryPSO, kernel_name, {input, other});
+      [computeEncoder setComputePipelineState:binaryPSO];
+      // Set input and output tensors
+      bind_iter_tensors(computeEncoder, iter);
+      // Iterator is contiguous if all of its elements are dense in storage,
+      // i.e. it's true for both row-first and column-first tensors
+      if (iter.is_contiguous()) {
+        detail::mtl_setArg(computeEncoder, params, 3);
+        if (cast_needed) {
+          std::array<int, 4> size_and_types = {static_cast<int>(c10::elementSize(input.scalar_type())),
+                                               static_cast<int>(c10::elementSize(other.scalar_type())),
+                                               static_cast<int>(input.scalar_type()),
+                                               static_cast<int>(other.scalar_type())};
+          mtl_setBytes(computeEncoder, size_and_types, 4);
+        }
+      } else {
+        // Please note that shapes and strides of the iterator might be
+        // different than that of its operands, for example binary op
+        // between 4x4 tensor and scalar will result in 1D 16 element iterator
+        std::array<int, 4> ndim_and_types = {iter.ndim(),
+                                             static_cast<int>(input.scalar_type()),
+                                             static_cast<int>(other.scalar_type()),
+                                             static_cast<int>(out.scalar_type())};
+        mtl_setArgs<3>(
+            computeEncoder, params, iter.shape(), iter.strides(0), iter.strides(1), iter.strides(2), ndim_and_types);
+      }
+      mtl_dispatch1DJob(computeEncoder, binaryPSO, iter.numel());
+      getMPSProfiler().endProfileKernel(binaryPSO);
+    }
+  });
+}
+
 } // namespace at::native::mps

aten/src/ATen/native/mps/kernels/Activation.h

@@ -0,0 +1,16 @@
+#pragma once
+
+template <typename T>
+struct ELUParams {
+  T alpha;
+  T scale;
+  T input_scale;
+};
+
+template <typename T>
+struct ELUBackwardParams {
+  T alpha;
+  T scale;
+  T input_scale;
+  bool is_result;
+};

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

@@ -1,3 +1,4 @@
+#include <ATen/native/mps/kernels/Activation.h>
 #include <c10/metal/indexing.h>
 #include <c10/metal/special_math.h>
 #include <metal_stdlib>
@@ -99,6 +100,59 @@ REGISTER_BINARY_OP(hardswish_backward, float, float);
 REGISTER_BINARY_OP(hardswish_backward, half, half);
 REGISTER_BINARY_OP(hardswish_backward, bfloat, bfloat);
 
+struct elu_functor {
+  template <typename T>
+  inline T operator()(const T self_, const ELUParams<T> params) {
+    using op_T = opmath_t<T>;
+    auto alpha = static_cast<op_T>(params.alpha);
+    auto scale = static_cast<op_T>(params.scale);
+    auto input_scale = static_cast<op_T>(params.input_scale);
+    auto self = static_cast<op_T>(self_);
+    auto neg_res = alpha * (::metal::precise::exp(self * input_scale) - 1);
+    return static_cast<T>(scale * (self < 0 ? neg_res : self));
+  }
+};
+
+struct elu_backward_functor {
+  template <typename T>
+  inline T operator()(
+      const T grad_output_,
+      const T self_,
+      ELUBackwardParams<T> params) {
+    using op_T = opmath_t<T>;
+    auto alpha = static_cast<op_T>(params.alpha);
+    auto scale = static_cast<op_T>(params.scale);
+    auto input_scale = static_cast<op_T>(params.input_scale);
+    auto grad_output = static_cast<op_T>(grad_output_);
+    auto self = static_cast<op_T>(self_);
+
+    if (params.is_result) {
+      auto neg_coef = input_scale * (self + alpha * scale);
+      return static_cast<T>(grad_output * (self <= 0 ? neg_coef : scale));
+    } else {
+      auto neg_coef = input_scale * alpha * scale *
+          ::metal::precise::exp(self * input_scale);
+      return static_cast<T>(grad_output * (self <= 0 ? neg_coef : scale));
+    }
+  }
+};
+
+#define REGISTER_ELU_OP(T)            \
+  typedef ELUParams<T> ELUParams_##T; \
+  REGISTER_UNARY_ALPHA_OP(elu, T, ELUParams_##T, T);
+
+REGISTER_ELU_OP(float);
+REGISTER_ELU_OP(half);
+REGISTER_ELU_OP(bfloat);
+
+#define REGISTER_ELU_BACKWARD_OP(T)                   \
+  typedef ELUBackwardParams<T> ELUBackwardParams_##T; \
+  REGISTER_BINARY_ALPHA_OP(elu_backward, T, ELUBackwardParams_##T, T);
+
+REGISTER_ELU_BACKWARD_OP(float);
+REGISTER_ELU_BACKWARD_OP(half);
+REGISTER_ELU_BACKWARD_OP(bfloat);
+
 struct leaky_relu_functor {
   template <typename T>
   inline T operator()(const T x, const T negative_slope) {

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

@@ -11,8 +11,6 @@
 #include <ATen/ops/_log_softmax_native.h>
 #include <ATen/ops/_prelu_kernel_backward_native.h>
 #include <ATen/ops/_prelu_kernel_native.h>
-#include <ATen/ops/elu_backward_native.h>
-#include <ATen/ops/elu_native.h>
 #include <ATen/ops/gelu_backward_native.h>
 #include <ATen/ops/gelu_native.h>
 #include <ATen/ops/glu_backward_native.h>
@@ -119,6 +117,10 @@ Tensor& relu_mps_(Tensor& self) {
 
 TORCH_IMPL_FUNC(log_softmax_mps_out)
 (const Tensor& self, const int64_t dim, const bool half_to_float, const Tensor& out) {
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "MPS doesn't know how to do exponent_i64");
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(self.scalar_type()),
+                              "log_softmax for complex is not supported for MPS");
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kBool, "log_softmax for bool is not supported for MPS");
   using namespace mps;
   using CachedGraph = MPSUnaryCachedGraph;
 
@@ -162,6 +164,10 @@ TORCH_IMPL_FUNC(log_softmax_mps_out)
 
 TORCH_IMPL_FUNC(log_softmax_backward_mps_out)
 (const Tensor& grad_output, const Tensor& output, int64_t dim, ScalarType input_dtype, const Tensor& out) {
+  TORCH_CHECK_NOT_IMPLEMENTED(grad_output.scalar_type() != kLong, "MPS doesn't know how to do exponent_i64");
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(grad_output.scalar_type()),
+                              "log_softmax for complex is not supported for MPS");
+  TORCH_CHECK_NOT_IMPLEMENTED(grad_output.scalar_type() != kBool, "log_softmax for bool is not supported for MPS");
   using namespace mps;
   using CachedGraph = MPSUnaryGradCachedGraph;
 
@@ -202,6 +208,7 @@ TORCH_IMPL_FUNC(log_softmax_backward_mps_out)
 }
 
 std::tuple<Tensor&, Tensor&> log_sigmoid_forward_out_mps(const Tensor& self, Tensor& output, Tensor& buffer) {
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "MPS doesn't know how to do exponent_i64");
   // NOTE: buffer is only used by CPU dispatch, we just ignore it here
   using namespace mps;
   using CachedGraph = MPSUnaryCachedGraph;
@@ -698,194 +705,6 @@ TORCH_IMPL_FUNC(gelu_backward_out_mps)
   }
 }
 
-static void elu_variants_out_mps(const Tensor& self,
-                                 const Scalar& alpha,
-                                 const Scalar& scale,
-                                 const Scalar& input_scale,
-                                 const Tensor& result,
-                                 std::string func_name) {
-  using namespace mps;
-  using CachedGraph = MPSUnaryCachedGraph;
-
-  auto resultMemFormat = result.suggest_memory_format();
-  bool executeGatherOp = !(self.is_contiguous(resultMemFormat) && result.is_contiguous(resultMemFormat));
-  Tensor out;
-  if (executeGatherOp) {
-    out = at::empty_like(result, MemoryFormat::Contiguous);
-  }
-
-  // Empty output
-  if (result.numel() == 0) {
-    return;
-  }
-
-  MPSStream* stream = getCurrentMPSStream();
-
-  @autoreleasepool {
-    std::string key = func_name + ":" + getTensorsStringKey({self}) + ":" + std::to_string(alpha.to<double>()) + ":" +
-        std::to_string(scale.to<double>()) + ":" + std::to_string(input_scale.to<double>());
-
-    auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
-      MPSGraphTensor* inputTensor = mpsGraphRankedPlaceHolder(mpsGraph, self);
-
-      // scale * (max(0, x) + min(0, alpha * (exp(input_scale * x) - 1) ))
-
-      MPSGraphTensor* alphaTensor = [mpsGraph constantWithScalar:alpha.to<double>()
-                                                           shape:@[ @1 ]
-                                                        dataType:getMPSDataType(self)];
-
-      MPSGraphTensor* inputScaleTensor = [mpsGraph constantWithScalar:input_scale.to<double>()
-                                                                shape:@[ @1 ]
-                                                             dataType:getMPSDataType(self)];
-
-      MPSGraphTensor* scaleTensor = [mpsGraph constantWithScalar:scale.to<double>()
-                                                           shape:@[ @1 ]
-                                                        dataType:getMPSDataType(self)];
-      MPSGraphTensor* unitTensor = [mpsGraph constantWithScalar:1.0f shape:@[ @1 ] dataType:getMPSDataType(self)];
-      MPSGraphTensor* zeroTensor = [mpsGraph constantWithScalar:0.0f shape:@[ @1 ] dataType:getMPSDataType(self)];
-
-      MPSGraphTensor* scaledInputTensor = [mpsGraph multiplicationWithPrimaryTensor:inputTensor
-                                                                    secondaryTensor:inputScaleTensor
-                                                                               name:nil];
-      MPSGraphTensor* exponentTensor = [mpsGraph exponentWithTensor:scaledInputTensor name:nil];
-      MPSGraphTensor* exponentMinusOneTensor = [mpsGraph subtractionWithPrimaryTensor:exponentTensor
-                                                                      secondaryTensor:unitTensor
-                                                                                 name:nil];
-      MPSGraphTensor* alphaTimesTensor = [mpsGraph multiplicationWithPrimaryTensor:exponentMinusOneTensor
-                                                                   secondaryTensor:alphaTensor
-                                                                              name:nil];
-      MPSGraphTensor* predicateTensor = [mpsGraph greaterThanWithPrimaryTensor:inputTensor
-                                                               secondaryTensor:zeroTensor
-                                                                          name:nil];
-      MPSGraphTensor* fusedOutput = [mpsGraph selectWithPredicateTensor:predicateTensor
-                                                    truePredicateTensor:inputTensor
-                                                   falsePredicateTensor:alphaTimesTensor
-                                                                   name:nil];
-      MPSGraphTensor* outputTensor = [mpsGraph multiplicationWithPrimaryTensor:fusedOutput
-                                                               secondaryTensor:scaleTensor
-                                                                          name:nil];
-
-      newCachedGraph->inputTensor_ = inputTensor;
-      newCachedGraph->outputTensor_ = outputTensor;
-    });
-
-    auto selfPlaceholder = Placeholder(cachedGraph->inputTensor_, self, nil, executeGatherOp);
-    auto outputPlaceholder = Placeholder(cachedGraph->outputTensor_, out.has_storage() ? out : result, nil, false);
-    auto feeds = dictionaryFromPlaceholders(selfPlaceholder);
-    runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
-    if (out.has_storage()) {
-      result.copy_(out);
-    }
-  }
-}
-
-// scale * (max(0, x) + min(0, alpha * (exp(input_scale * x) - 1) ))
-TORCH_IMPL_FUNC(elu_out_mps)
-(const Tensor& self, const Scalar& alpha, const Scalar& scale, const Scalar& input_scale, const Tensor& result) {
-  elu_variants_out_mps(self, alpha, scale, input_scale, result, "elu_out_mps");
-}
-
-TORCH_IMPL_FUNC(elu_backward_out_mps)
-(const Tensor& grad_output,
- const Scalar& alpha,
- const Scalar& scale,
- const Scalar& input_scale,
- bool is_result,
- const Tensor& self_or_result,
- const Tensor& grad_input) {
-  using namespace mps;
-  using CachedGraph = MPSUnaryGradCachedGraph;
-  auto gradMemFormat = grad_input.suggest_memory_format();
-  bool executeGatherOp = !(grad_output.is_contiguous(gradMemFormat) && self_or_result.is_contiguous(gradMemFormat) &&
-                           grad_input.is_contiguous(gradMemFormat));
-  Tensor out;
-  if (executeGatherOp && gradMemFormat == MemoryFormat::ChannelsLast) {
-    out = at::empty_like(grad_input, MemoryFormat::Contiguous);
-  }
-
-  // Empty output
-  if (grad_input.numel() == 0) {
-    return;
-  }
-
-  MPSStream* stream = getCurrentMPSStream();
-
-  @autoreleasepool {
-    std::string key = "elu_backward_out_mps:" + getTensorsStringKey({grad_output, self_or_result}) + ":" +
-        std::to_string(alpha.to<double>()) + ":" + std::to_string(scale.to<double>()) + ":" +
-        std::to_string(input_scale.to<double>()) + ":" + std::to_string(is_result);
-
-    auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
-      MPSGraphTensor* gradOutputTensor = mpsGraphRankedPlaceHolder(mpsGraph, grad_output);
-      MPSGraphTensor* selfOrResultTensor = mpsGraphRankedPlaceHolder(mpsGraph, self_or_result);
-      MPSGraphTensor* lessThanZeroGradTensor = nil;
-
-      if (is_result) {
-        MPSGraphTensor* alphaTensor = [mpsGraph constantWithScalar:alpha.to<double>()
-                                                             shape:@[ @1 ]
-                                                          dataType:getMPSDataType(grad_output)];
-        MPSGraphTensor* resultPlusAlphaTensor = [mpsGraph additionWithPrimaryTensor:selfOrResultTensor
-                                                                    secondaryTensor:alphaTensor
-                                                                               name:nil];
-        auto constMul = scale.to<double>() * input_scale.to<double>();
-        MPSGraphTensor* constMulTensor = [mpsGraph constantWithScalar:constMul
-                                                                shape:@[ @1 ]
-                                                             dataType:getMPSDataType(grad_output)];
-        lessThanZeroGradTensor = [mpsGraph multiplicationWithPrimaryTensor:resultPlusAlphaTensor
-                                                           secondaryTensor:constMulTensor
-                                                                      name:nil];
-      } else {
-        MPSGraphTensor* inputScaleTensor = [mpsGraph constantWithScalar:input_scale.to<double>()
-                                                                  shape:@[ @1 ]
-                                                               dataType:getMPSDataType(grad_output)];
-        MPSGraphTensor* scaledInputTensor = [mpsGraph multiplicationWithPrimaryTensor:selfOrResultTensor
-                                                                      secondaryTensor:inputScaleTensor
-                                                                                 name:nil];
-        MPSGraphTensor* expTensor = [mpsGraph exponentWithTensor:scaledInputTensor name:nil];
-        auto constMul = scale.to<double>() * input_scale.to<double>() * alpha.to<double>();
-        MPSGraphTensor* constMulTensor = [mpsGraph constantWithScalar:constMul
-                                                                shape:@[ @1 ]
-                                                             dataType:getMPSDataType(grad_output)];
-        lessThanZeroGradTensor = [mpsGraph multiplicationWithPrimaryTensor:expTensor
-                                                           secondaryTensor:constMulTensor
-                                                                      name:nil];
-      }
-
-      MPSGraphTensor* scaleTensor = [mpsGraph constantWithScalar:scale.to<double>()
-                                                           shape:@[ @1 ]
-                                                        dataType:getMPSDataType(grad_output)];
-      MPSGraphTensor* zeroTensor = [mpsGraph constantWithScalar:0.0f
-                                                          shape:@[ @1 ]
-                                                       dataType:getMPSDataType(grad_output)];
-      MPSGraphTensor* predicateTensor = [mpsGraph greaterThanWithPrimaryTensor:selfOrResultTensor
-                                                               secondaryTensor:zeroTensor
-                                                                          name:nil];
-      MPSGraphTensor* gradTensor = [mpsGraph selectWithPredicateTensor:predicateTensor
-                                                   truePredicateTensor:scaleTensor
-                                                  falsePredicateTensor:lessThanZeroGradTensor
-                                                                  name:nil];
-      MPSGraphTensor* gradInputTensor = [mpsGraph multiplicationWithPrimaryTensor:gradTensor
-                                                                  secondaryTensor:gradOutputTensor
-                                                                             name:nil];
-
-      newCachedGraph->gradOutputTensor_ = gradOutputTensor;
-      newCachedGraph->inputTensor_ = selfOrResultTensor;
-      newCachedGraph->gradInputTensor_ = gradInputTensor;
-    });
-
-    Placeholder gradOutputPlaceholder = Placeholder(cachedGraph->gradOutputTensor_, grad_output, nil, executeGatherOp);
-    Placeholder selfOrResultPlaceholder = Placeholder(cachedGraph->inputTensor_, self_or_result, nil, executeGatherOp);
-    Placeholder gradInputPlaceholder =
-        Placeholder(cachedGraph->gradInputTensor_, out.has_storage() ? out : grad_input, nil, false);
-
-    auto feeds = dictionaryFromPlaceholders(gradOutputPlaceholder, selfOrResultPlaceholder);
-    runMPSGraph(stream, cachedGraph->graph(), feeds, gradInputPlaceholder);
-    if (out.has_storage()) {
-      grad_input.copy_(out);
-    }
-  }
-}
-
 TORCH_IMPL_FUNC(glu_out_mps)(const Tensor& self, const int64_t dim, const Tensor& output) {
   using namespace mps;
   using CachedGraph = MPSUnaryCachedGraph;
@@ -896,6 +715,7 @@ TORCH_IMPL_FUNC(glu_out_mps)(const Tensor& self, const int64_t dim, const Tensor
   if (output.numel() == 0)
     return;
 
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "MPS doesn't know how to do exponent_i64");
   // this can't pass anyway because a 0-dimensional tensor has "size" 1, which
   // can't be evenly halved, but give a nicer error message here.
   TORCH_CHECK(self.dim() > 0, "glu does not support 0-dimensional tensors");
@@ -1009,6 +829,7 @@ TORCH_IMPL_FUNC(softplus_out_mps)
 (const Tensor& self, const Scalar& beta, const Scalar& threshold, const Tensor& result) {
   using namespace mps;
   TORCH_CHECK(self.is_mps());
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "Not implemented for long");
   // Applies the Softplus function :math:`\text{Softplus}(x) = \frac{1}{\beta} *
   // \log(1 + \exp(\beta * x))` element-wise.
   // For numerical stability the implementation reverts to the linear function
@@ -1159,6 +980,8 @@ TORCH_IMPL_FUNC(mish_out_mps)
 (const Tensor& self, const Tensor& result) {
   using namespace mps;
   TORCH_CHECK(self.is_mps());
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "MPS doesn't know how to do exponent_i64");
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(self.scalar_type()), "Mish for complex is not supported for MPS");
 
   if (result.numel() == 0)
     return;
@@ -1207,6 +1030,8 @@ TORCH_IMPL_FUNC(mish_out_mps)
 Tensor mish_backward_mps(const Tensor& grad_output, const Tensor& self) {
   using namespace mps;
   TORCH_CHECK(self.is_mps());
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "MPS doesn't know how to do exponent_i64");
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(self.scalar_type()), "Mish for complex is not supported for MPS");
 
   Tensor grad_input = at::empty_like(self, self.suggest_memory_format());
   if (grad_input.numel() == 0)
@@ -1396,6 +1221,7 @@ TORCH_IMPL_FUNC(silu_out_mps)(const Tensor& self, const Tensor& result) {
   using CachedGraph = MPSUnaryCachedGraph;
 
   TORCH_CHECK(self.is_mps());
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "MPS doesn't know how to do exponent_i64");
 
   // Empty output
   if (result.numel() == 0)

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

@@ -1,8 +1,10 @@
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
+#include <ATen/Dispatch.h>
 #include <ATen/TensorIterator.h>
 #include <ATen/mps/MPSProfiler.h>
 #include <ATen/native/Activation.h>
 #include <ATen/native/mps/OperationUtils.h>
+#include <ATen/native/mps/kernels/Activation.h>
 #include <fmt/format.h>
 
 namespace at::native {
@@ -41,6 +43,30 @@ static void hardswish_backward_kernel(at::TensorIterator& iter) {
   lib.exec_binary_kernel(iter, "hardswish_backward");
 }
 
+static void elu_kernel(TensorIteratorBase& iter, const Scalar& alpha, const Scalar& scale, const Scalar& input_scale) {
+  AT_DISPATCH_FLOATING_TYPES_AND2(c10::kHalf, c10::kBFloat16, iter.common_dtype(), "elu_mps", [&]() {
+    ELUParams<scalar_t> params{alpha.to<scalar_t>(), scale.to<scalar_t>(), input_scale.to<scalar_t>()};
+    lib.exec_unary_kernel_with_params(
+        iter, "elu", params, fmt::format("ELUParams_{}", mps::scalarToMetalTypeString(iter.common_dtype())));
+  });
+}
+
+static void elu_backward_kernel(TensorIteratorBase& iter,
+                                const Scalar& alpha,
+                                const Scalar& scale,
+                                const Scalar& input_scale,
+                                bool is_result) {
+  AT_DISPATCH_FLOATING_TYPES_AND2(c10::kHalf, c10::kBFloat16, iter.common_dtype(), "elu_backward_mps", [&]() {
+    ELUBackwardParams<scalar_t> params{
+        alpha.to<scalar_t>(), scale.to<scalar_t>(), input_scale.to<scalar_t>(), is_result};
+    lib.exec_binary_kernel_with_params(
+        iter,
+        "elu_backward",
+        params,
+        fmt::format("ELUBackwardParams_{}", mps::scalarToMetalTypeString(iter.common_dtype())));
+  });
+}
+
 static void leaky_relu_kernel(TensorIteratorBase& iter, const Scalar& negative_slope) {
   lib.exec_unary_kernel(iter, "leaky_relu", negative_slope);
 }
@@ -56,6 +82,8 @@ REGISTER_DISPATCH(hardsigmoid_stub, hardsigmoid_kernel);
 REGISTER_DISPATCH(hardsigmoid_backward_stub, hardsigmoid_backward_kernel);
 REGISTER_DISPATCH(hardswish_stub, hardswish_kernel);
 REGISTER_DISPATCH(hardswish_backward_stub, hardswish_backward_kernel);
+REGISTER_DISPATCH(elu_stub, elu_kernel);
+REGISTER_DISPATCH(elu_backward_stub, elu_backward_kernel);
 REGISTER_DISPATCH(leaky_relu_stub, leaky_relu_kernel);
 REGISTER_DISPATCH(leaky_relu_backward_stub, leaky_relu_backward_kernel);
 

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

@@ -80,6 +80,11 @@ static void grid_sampler_2d_mps_impl(Tensor& output,
     MPSGraphTensor* outputTensor_ = nil;
   };
 
+  // Crashes with
+  // MPSGraphUtilities.mm:97:0: error: 'mps.sample_grid' op operand #0 must be tensor of mps native type values, but got
+  // 'tensor<2x3x5x20xcomplex<f32>>'
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(input.scalar_type()),
+                              "grid_sampler_2d is not supported for complex on MPS");
   @autoreleasepool {
     std::string key = "grid_sampler_2d_mps" + getTensorsStringKey({input, grid}) + ":" +
         std::to_string(interpolation_mode) + ":" + std::to_string(padding_mode) + ":" + std::to_string(align_corners);

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

@@ -240,7 +240,7 @@ static void linalg_lu_factor_ex_out_mps_impl(const Tensor& A,
                                              bool check_errors) {
   using namespace mps;
 
-  TORCH_CHECK(!c10::isComplexType(A.scalar_type()) && !c10::isComplexType(LU.scalar_type()),
+  TORCH_CHECK(A.scalar_type() == kFloat && LU.scalar_type() == kFloat,
               "linalg.lu_factor(): MPS doesn't support complex types.");
   TORCH_CHECK(pivot, "linalg.lu_factor(): MPS doesn't allow pivot == False.");
 
@@ -364,8 +364,7 @@ static void linalg_solve_out_mps_impl(const Tensor& A,
                                       const Tensor& info) {
   using namespace mps;
 
-  TORCH_CHECK(!c10::isComplexType(A.scalar_type()) && !c10::isComplexType(LU.scalar_type()),
-              "linalg.lu_factor(): MPS doesn't support complex types.");
+  TORCH_CHECK(A.scalar_type() == kFloat && LU.scalar_type() == kFloat, "linalg.lu_factor(): MPS only supports floats.");
   Tensor A_t, B_t;
   // If 'left' is false, reinterpret the problem so that Ax = B becomes A^T ⋅ (x^T) = B^T
   // Then we solve the normal "left" case on the transposed matrices and transpose x finally to get the output
@@ -1058,7 +1057,8 @@ static Tensor& linalg_solve_triangular_mps_impl(const Tensor& A,
   using namespace mps;
 
   checkInputsSolver(A, B, left, "linalg.solve_triangular");
-  TORCH_CHECK(!A.is_complex() && !B.is_complex(), "linalg.solve.triangular(); Not supported for complex yet!");
+  TORCH_CHECK(A.scalar_type() == kFloat && B.scalar_type() == kFloat,
+              "linalg.solve.triangular(); Only float is supported!");
   Tensor A_t, B_t;
   std::tie(B_t, A_t) = _linalg_broadcast_batch_dims(B, A, /*don't check errors*/ nullptr);
   at::native::resize_output(out, B_t.sizes());

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

@@ -416,6 +416,8 @@ static void nllnd_loss_forward_impl(Tensor& output,
                                     int64_t reduction,
                                     int64_t ignore_index,
                                     bool is2D) {
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(output.scalar_type()),
+                              "nlld_loss for complex is not supported for MPS");
   std::vector<long long> reshapedTarget(target_arg.sizes().begin(), target_arg.sizes().end());
   reshapedTarget.push_back(1);
 
@@ -824,6 +826,9 @@ static void smooth_l1_loss_backward_impl(const Tensor& grad_output,
 Tensor& huber_loss_out_mps(const Tensor& input, const Tensor& target, int64_t reduction, double delta, Tensor& output) {
   std::string op_name = __func__;
   using namespace mps;
+  TORCH_CHECK_NOT_IMPLEMENTED(input.scalar_type() != kLong, "MPS doesn't know how to do square_i64");
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(input.scalar_type()),
+                              "huber_loss for complex is not supported for MPS");
   TORCH_CHECK(delta > 0, "huber_loss does not support non-positive values for delta.")
   TORCH_CHECK(target.is_same_size(input), op_name + ": target and input tensors must have identical shapes")
   TORCH_CHECK(output.is_mps());

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

@@ -597,6 +597,7 @@ static void avg_pool2d_template(const Tensor& input,
                                 bool count_include_pad,
                                 const std::optional<int64_t> divisor_override,
                                 const std::string& op_name) {
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(input.scalar_type()), "Not implemented for complex");
   const Tensor& grad_output = *(at::borrow_from_optional_tensor(grad_output_opt));
   const bool is_backward_pass = grad_output.defined();
   const bool use_divisor = divisor_override.has_value() && divisor_override.value() != 0;
@@ -915,6 +916,8 @@ TORCH_IMPL_FUNC(max_pool2d_with_indices_out_mps)
  bool ceil_mode,
  const Tensor& output,
  const Tensor& indices) {
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(input.scalar_type()),
+                              "Max pooling for complex is not supported for MPS");
   bool use_graph = use_graph_for_max_pool2d(kernel_size, stride);
   if (use_graph) {
     auto indices_memory_format = indices.suggest_memory_format();
@@ -967,6 +970,8 @@ TORCH_IMPL_FUNC(max_pool2d_with_indices_backward_out_mps)
  bool ceil_mode,
  const Tensor& indices,
  const Tensor& grad_input) {
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(input.scalar_type()),
+                              "Max pooling for complex is not supported for MPS");
   mps::PoolingOpBlock pooling_op_block = ^PoolingOpFn(cachedGraph, desc) {
     MPSGraph* mpsGraph = cachedGraph.graph();
     return [mpsGraph maxPooling2DGradientWithGradientTensor:cachedGraph.gradOutputTensor

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

@@ -269,17 +269,22 @@ static void reduction_out_mps(const Tensor& input_t,
                                                                      name:nil];
         castOutputTensor = [mpsGraph reductionSumWithTensor:bandPartWithTensor axes:@[ @0, @1 ] name:nil];
       } else if (reduction_type == MPSReductionType::NANSUM) {
-        // Create a 0 tensor of the same shape as inputTensor
-        MPSGraphTensor* zeros = [mpsGraph constantWithScalar:0.0 dataType:castInputTensor.dataType];
-        // Find NaNs
-        MPSGraphTensor* nanMask = [mpsGraph isNaNWithTensor:castInputTensor name:nil];
-        // Replace NaNs with 0
-        MPSGraphTensor* nanReplaced = [mpsGraph selectWithPredicateTensor:nanMask
-                                                      truePredicateTensor:zeros
-                                                     falsePredicateTensor:castInputTensor
-                                                                     name:nil];
-        // Sum
-        castOutputTensor = [mpsGraph reductionSumWithTensor:nanReplaced axes:wrappedAxes name:nil];
+        // Integral types cannot contain NaN, so just do regular sum
+        if (([castInputTensor dataType] & MPSDataTypeFloatBit) == 0) {
+          castOutputTensor = [mpsGraph reductionSumWithTensor:castInputTensor axes:wrappedAxes name:nil];
+        } else {
+          // Create a 0 tensor of the same shape as inputTensor
+          auto zeros = [mpsGraph constantWithScalar:0.0 dataType:castInputTensor.dataType];
+          // Find NaNs
+          auto nanMask = [mpsGraph isNaNWithTensor:castInputTensor name:nil];
+          // Replace NaNs with 0
+          auto nanReplaced = [mpsGraph selectWithPredicateTensor:nanMask
+                                             truePredicateTensor:zeros
+                                            falsePredicateTensor:castInputTensor
+                                                            name:nil];
+          // Sum
+          castOutputTensor = [mpsGraph reductionSumWithTensor:nanReplaced axes:wrappedAxes name:nil];
+        }
       }
 
       MPSGraphTensor* outputTensor = castOutputTensor;
@@ -442,6 +447,7 @@ static Tensor std_var_common_impl_mps(const Tensor& input_t,
                                       const std::optional<Scalar>& correction,
                                       bool keepdim,
                                       StdVarType stdVarType) {
+  TORCH_CHECK_NOT_IMPLEMENTED(input_t.scalar_type() != kLong, "Not implemented for MPS");
   using CachedGraph = MPSUnaryCachedGraph;
 
   IntArrayRef input_shape = input_t.sizes();

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/SoftMax.mm

@@ -39,6 +39,7 @@ static void get_shapes(MPSShape* input_shape_readonly,
 TORCH_IMPL_FUNC(softmax_mps_out)
 (const Tensor& input_, const int64_t dim, const bool half_to_float, const Tensor& output) {
   TORCH_CHECK(!half_to_float, "softmax with half to float conversion is not supported on MPS");
+  TORCH_CHECK(c10::isFloatingType(input_.scalar_type()), "softmax only supported for floating types");
   static const bool is_macOS_15_0_or_newer = is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS);
 
   if (input_.numel() == 0) {

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

@@ -18,6 +18,10 @@ static Tensor& bincount_mps_impl(const Tensor& self, const Tensor& weights, Tens
   MPSStream* stream = getCurrentMPSStream();
   bool has_weights = weights.defined();
 
+  // Crashes with
+  // MPSGraphUtilities.mm:190:0: error: 'mps.scatter' op operand #2 must be tensor of int values, but got 'tensor<5xi1>'
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kBool, "bincount is not supported for Bool");
+
   @autoreleasepool {
     std::string key = "bincount_mps_impl" + getTensorsStringKey({self, weights});
     auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {

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 {

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
@@ -4617,7 +4617,7 @@
   dispatch:
     CompositeExplicitAutograd: permute
     MPS: permute_mps
-    SparseCPU, SparseCUDA: permute_sparse_coo
+    SparseCPU, SparseCUDA, SparseMPS: permute_sparse_coo
   tags: core
 
 - func: movedim.intlist(Tensor(a) self, int[] source, int[] destination) -> Tensor(a)
@@ -12064,8 +12064,7 @@
   device_check: NoCheck   # TensorIterator
   python_module: nn
   dispatch:
-    CPU, CUDA: elu_out
-    MPS: elu_out_mps
+    CPU, CUDA, MPS: elu_out
 
 - func: elu(Tensor self, Scalar alpha=1, Scalar scale=1, Scalar input_scale=1) -> Tensor
   structured_delegate: elu.out
@@ -12078,8 +12077,7 @@
   structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU, CUDA: elu_backward_out
-    MPS: elu_backward_out_mps
+    CPU, CUDA, MPS: elu_backward_out
 
 - func: elu_backward(Tensor grad_output, Scalar alpha, Scalar scale, Scalar input_scale, bool is_result, Tensor self_or_result) -> Tensor
   structured_delegate: elu_backward.grad_input

aten/src/ATen/test/CMakeLists.txt

@@ -61,9 +61,11 @@ list(APPEND ATen_CUDA_TEST_SRCS
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_complex_math_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_complex_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_cub_test.cu
+  ${CMAKE_CURRENT_SOURCE_DIR}/cuda_cublas_handle_pool_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_device_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_distributions_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_dlconvertor_test.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/cuda_event_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_exchange_device_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_generator_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_half_test.cu

aten/src/ATen/test/cuda_cublas_handle_pool_test.cpp

@@ -0,0 +1,77 @@
+#include <gtest/gtest.h>
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDACachingAllocator.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include <atomic>
+#include <thread>
+#include <vector>
+
+// Test concurrent access to getCurrentCUDABlasHandle and getCUDABlasLtWorkspace
+// to verify that the data race fix is working correctly
+
+TEST(CUDABlasHandlePoolTest, ConcurrentGetAndClearWorkspaces) {
+  if (!at::cuda::is_available()) {
+    return;
+  }
+
+  constexpr int num_accessor_threads = 15;
+  constexpr int num_clear_threads = 5;
+  constexpr int iterations_per_thread = 50;
+
+  std::atomic<bool> stop{false};
+  std::atomic<int> error_count{0};
+  std::vector<std::thread> threads;
+  threads.reserve(num_accessor_threads + num_clear_threads);
+
+  // Launch accessor threads
+  for (int i = 0; i < num_accessor_threads; ++i) {
+    threads.emplace_back([&stop, &error_count]() {
+      try {
+        at::cuda::CUDAGuard device_guard(0);
+
+        while (!stop.load(std::memory_order_relaxed)) {
+          const auto handle = at::cuda::getCurrentCUDABlasHandle();
+          const auto workspace = at::cuda::getCUDABlasLtWorkspace();
+
+          if (handle == nullptr || workspace == nullptr) {
+            error_count++;
+          }
+        }
+      } catch (const std::exception& e) {
+        error_count++;
+      }
+    });
+  }
+
+  // Launch threads that clear workspaces
+  for (int i = 0; i < num_clear_threads; ++i) {
+    threads.emplace_back([&error_count]() {
+      try {
+        for (int j = 0; j < iterations_per_thread; ++j) {
+          at::cuda::clearCublasWorkspaces();
+          std::this_thread::yield();
+        }
+      } catch (const std::exception& e) {
+        error_count++;
+      }
+    });
+  }
+
+  // Let them run for a bit
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));
+  stop.store(true, std::memory_order_relaxed);
+
+  for (auto& thread : threads) {
+    thread.join();
+  }
+
+  EXPECT_EQ(error_count.load(), 0);
+}
+
+int main(int argc, char* argv[]) {
+  ::testing::InitGoogleTest(&argc, argv);
+  c10::cuda::CUDACachingAllocator::init(1);
+  return RUN_ALL_TESTS();
+}

aten/src/ATen/test/cuda_event_test.cpp

@@ -0,0 +1,36 @@
+#include <gtest/gtest.h>
+
+#include <ATen/cuda/CUDAEvent.h>
+#include <ATen/cuda/CUDAGraph.h>
+#include <ATen/cuda/Sleep.h>
+
+TEST(CUDAEventTest, testCUDAExternalEvent) {
+  if (!at::cuda::is_available()) {
+    return;
+  }
+
+  // Create two external CUDA events
+  unsigned int flags = cudaEventDefault | cudaEventExternal;
+  auto event1 = at::cuda::CUDAEvent(flags);
+  auto event2 = at::cuda::CUDAEvent(flags);
+  // Ensure external CUDAEvent remain valid and functional after being moved.
+  auto start_event = std::move(event1);
+  auto end_event = std::move(event2);
+
+  auto stream = at::cuda::getStreamFromPool();
+  at::cuda::setCurrentCUDAStream(stream);
+
+  auto graph = at::cuda::CUDAGraph();
+  graph.capture_begin();
+  start_event.record();
+  at::cuda::sleep(100000);
+  end_event.record();
+  graph.capture_end();
+
+  // External events should correctly record timestamps even when used inside
+  // CUDA graphs, and elapsed_time() between them should be positive.
+  stream.synchronize();
+  graph.replay();
+  at::cuda::device_synchronize();
+  EXPECT_TRUE(start_event.elapsed_time(end_event) > 0);
+}

aten/src/ATen/xpu/XPUScaledBlas.cpp

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

aten/src/ATen/xpu/XPUScaledBlas.h

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

aten/tools/valgrind.sup

@@ -10,6 +10,13 @@
    ...
 }
 
+{
+   ignore_empty_generic_uninitialised_conditional_jump
+   Memcheck:Cond
+   fun:_ZN2at6detail13empty_genericEN3c108ArrayRefIlEEPNS1_9AllocatorENS1_14DispatchKeySetENS1_10ScalarTypeESt8optionalINS1_12MemoryFormatEE
+   ...
+}
+
 {
    Cond_cuda
    Memcheck:Cond

benchmarks/dynamo/check_perf_csv.py

@@ -9,28 +9,61 @@ def check_perf_csv(filename, threshold, threshold_scale):
     """
     Basic performance checking.
     """
+    try:
+        df = pd.read_csv(filename)
+    except FileNotFoundError:
+        print(f"Error: File {filename} not found")
+        sys.exit(1)
 
-    df = pd.read_csv(filename)
+    effective_threshold = threshold * threshold_scale
+    print(f"Checking {filename} (speedup threshold >= {effective_threshold:.2f}x)\n")
 
     failed = []
     for _, row in df.iterrows():
         model_name = row["name"]
-        speedup = row["speedup"]
-        if speedup < threshold * threshold_scale:
-            failed.append(model_name)
+        speedup = float(row["speedup"])
+        abs_latency = float(row["abs_latency"])
+        compilation_latency = float(row["compilation_latency"])
+        compression_ratio = float(row["compression_ratio"])
+        eager_peak_mem = float(row["eager_peak_mem"])
+        dynamo_peak_mem = float(row["dynamo_peak_mem"])
 
-        print(f"{model_name:34} {speedup}")
+        perf_summary = f"{model_name:34} speedup={speedup:.3f}x"
+        if pd.notna(abs_latency):
+            perf_summary += f", latency={abs_latency:.1f} ms/iter"
+        if pd.notna(compilation_latency):
+            perf_summary += f", compile={compilation_latency:.3f}s"
+        if pd.notna(compression_ratio):
+            perf_summary += f", mem_ratio={1 / compression_ratio:.2f}x"
+            if pd.notna(eager_peak_mem) and pd.notna(dynamo_peak_mem):
+                perf_summary += (
+                    f" (eager={eager_peak_mem:.1f} GB, dynamo={dynamo_peak_mem:.1f} GB)"
+                )
+
+        if speedup < effective_threshold:
+            failed.append((model_name, speedup))
+
+        print(perf_summary)
 
     if failed:
         print(
             textwrap.dedent(
                 f"""
-                Error {len(failed)} models performance regressed
-                    {" ".join(failed)}
+                Error {len(failed)} model(s) performance regressed
+                    {" ".join([name for name, _ in failed])}
                 """
             )
         )
+        for name, sp in sorted(failed, key=lambda x: x[1]):
+            pct_from_target = (sp / effective_threshold - 1.0) * 100.0
+            print(
+                f"  - {name}: {sp:.3f}x (< {effective_threshold:.2f}x; {pct_from_target:.1f}% from target)"
+            )
         sys.exit(1)
+    else:
+        print(
+            f"\nAll {len(df)} model(s) passed threshold check (>= {effective_threshold:.2f}x)"
+        )
 
 
 if __name__ == "__main__":
@@ -44,7 +77,7 @@ if __name__ == "__main__":
         "-s",
         type=float,
         default=1.0,
-        help="multiple threshold by this value to relax the check",
+        help="multiply threshold by this value to relax the check",
     )
     args = parser.parse_args()
     check_perf_csv(args.file, args.threshold, args.threshold_scale)

benchmarks/dynamo/common.py

@@ -2379,7 +2379,9 @@ class BenchmarkRunner:
                     print(
                         f"Load model outputs from {self.args.compare_model_outputs_with} to compare"
                     )
-                    saved_result = torch.load(self.args.compare_model_outputs_with)
+                    saved_result = torch.load(
+                        self.args.compare_model_outputs_with, weights_only=False
+                    )
                     is_bitwise_same = bitwise_same(saved_result, new_result)
                     if not is_bitwise_same:
                         print(

c10/core/SafePyObject.h

@@ -44,7 +44,7 @@ struct C10_API SafePyObject {
       (*other.pyinterpreter_)->incref(other.data_);
     }
     if (data_ != nullptr) {
-      (*pyinterpreter_)->decref(data_, /*has_pyobj_slot*/ false);
+      (*pyinterpreter_)->decref(data_);
     }
     data_ = other.data_;
     pyinterpreter_ = other.pyinterpreter_;
@@ -53,7 +53,7 @@ struct C10_API SafePyObject {
 
   ~SafePyObject() {
     if (data_ != nullptr) {
-      (*pyinterpreter_)->decref(data_, /*has_pyobj_slot*/ false);
+      (*pyinterpreter_)->decref(data_);
     }
   }
 

c10/core/ScalarType.h

@@ -92,13 +92,6 @@ inline bool isComplexType(ScalarType t) {
       t == ScalarType::ComplexDouble);
 }
 
-inline bool isQIntType(ScalarType t) {
-  // Don't forget to extend this when adding new QInt types
-  return t == ScalarType::QInt8 || t == ScalarType::QUInt8 ||
-      t == ScalarType::QInt32 || t == ScalarType::QUInt4x2 ||
-      t == ScalarType::QUInt2x4;
-}
-
 inline bool isBitsType(ScalarType t) {
   return t == ScalarType::Bits1x8 || t == ScalarType::Bits2x4 ||
       t == ScalarType::Bits4x2 || t == ScalarType::Bits8 ||

c10/core/StorageImpl.cpp

@@ -48,6 +48,30 @@ void warnDeprecatedDataPtr() {
   TORCH_CHECK(false, "Cannot access data pointer of Storage that is invalid.");
 }
 
+void StorageImpl::incref_pyobject() const noexcept {
+  // Because intrusive_ptr incref uses relaxed memory order, we need to
+  // do an acquire fence to ensure that the kHasPyObject bit was
+  // observed before the load of the PyObject* below.
+  // NB: This is a no-op on x86/x86-64
+  std::atomic_thread_fence(std::memory_order_acquire);
+
+  PyObject* obj = pyobj_slot_.load_pyobj();
+  (*pyobj_slot_.pyobj_interpreter())->incref(obj);
+}
+
+void StorageImpl::decref_pyobject() const noexcept {
+  PyObject* obj = pyobj_slot_.load_pyobj();
+  (*pyobj_slot_.pyobj_interpreter())->decref(obj);
+}
+
+bool StorageImpl::try_incref_pyobject() const noexcept {
+  c10::impl::PyInterpreter* interp = pyobj_slot_.pyobj_interpreter();
+  if (C10_UNLIKELY(!interp)) {
+    return false;
+  }
+  return (*interp)->try_incref(pyobj_slot_);
+}
+
 void SetStorageImplCreate(DeviceType t, StorageImplCreateHelper fptr) {
   // Allowlist verification.
   // Only if the devicetype is in the allowlist,

c10/core/StorageImpl.h

@@ -105,6 +105,12 @@ struct C10_API StorageImpl : public c10::intrusive_ptr_target {
     data_ptr_.clear();
   }
 
+  void incref_pyobject() const noexcept override final;
+
+  void decref_pyobject() const noexcept override final;
+
+  bool try_incref_pyobject() const noexcept override final;
+
   size_t nbytes() const {
     // OK to do this instead of maybe_as_int as nbytes is guaranteed positive
     TORCH_CHECK(!size_bytes_is_heap_allocated_);
@@ -370,4 +376,18 @@ C10_API c10::intrusive_ptr<c10::StorageImpl> make_storage_impl(
     bool resizable,
     std::optional<at::Device> device_opt);
 
+namespace detail {
+
+#ifndef C10_MOBILE
+template <class T>
+struct TargetTraits<
+    T,
+    std::enable_if_t<
+        std::is_base_of_v<c10::StorageImpl, std::remove_cv_t<T>>>> {
+  static constexpr bool can_have_pyobject = true;
+};
+#endif
+
+} // namespace detail
+
 } // namespace c10

c10/core/TensorImpl.cpp

@@ -277,7 +277,6 @@ void TensorImpl::release_resources() {
   if (storage_) {
     storage_ = {};
   }
-  pyobj_slot_.maybe_destroy_pyobj();
 }
 
 #ifndef C10_DISABLE_TENSORIMPL_EXTENSIBILITY
@@ -989,6 +988,30 @@ void TensorImpl::empty_tensor_restride_symint(MemoryFormat memory_format) {
   }
 }
 
+void TensorImpl::incref_pyobject() const noexcept {
+  // Because intrusive_ptr incref uses relaxed memory order, we need to
+  // do an acquire fence to ensure that the kHasPyObject bit was
+  // observed before the load of the PyObject* below.
+  // NB: This is a no-op on x86/x86-64
+  std::atomic_thread_fence(std::memory_order_acquire);
+
+  PyObject* obj = pyobj_slot_.load_pyobj();
+  (*pyobj_slot_.pyobj_interpreter())->incref(obj);
+}
+
+void TensorImpl::decref_pyobject() const noexcept {
+  PyObject* obj = pyobj_slot_.load_pyobj();
+  (*pyobj_slot_.pyobj_interpreter())->decref(obj);
+}
+
+bool TensorImpl::try_incref_pyobject() const noexcept {
+  c10::impl::PyInterpreter* interp = pyobj_slot_.pyobj_interpreter();
+  if (C10_UNLIKELY(!interp)) {
+    return false;
+  }
+  return (*interp)->try_incref(pyobj_slot_);
+}
+
 namespace impl {
 
 namespace {

c10/core/TensorImpl.h

@@ -2178,6 +2178,12 @@ struct C10_API TensorImpl : public c10::intrusive_ptr_target {
     return &pyobj_slot_;
   }
 
+  void incref_pyobject() const noexcept override final;
+
+  void decref_pyobject() const noexcept override final;
+
+  bool try_incref_pyobject() const noexcept override final;
+
  private:
   // See NOTE [std::optional operator usage in CUDA]
   // We probably don't want to expose this publicly until
@@ -3079,6 +3085,19 @@ struct C10_API TensorImpl : public c10::intrusive_ptr_target {
   friend class C10_TensorImpl_Size_Check_Dummy_Class;
 };
 
+namespace detail {
+
+#ifndef C10_MOBILE
+template <class T>
+struct TargetTraits<
+    T,
+    std::enable_if_t<std::is_base_of_v<c10::TensorImpl, std::remove_cv_t<T>>>> {
+  static constexpr bool can_have_pyobject = true;
+};
+#endif
+
+} // namespace detail
+
 // Note [TensorImpl size constraints]
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 // Changed the size of TensorImpl?  If the size went down, good for

c10/core/impl/PyInterpreter.cpp

@@ -11,8 +11,11 @@ struct NoopPyInterpreterVTable final : public PyInterpreterVTable {
 
   void incref(PyObject* pyobj) const override {} // do nothing
 
-  void decref(PyObject* pyobj, bool has_pyobj_slot) const override {
-  } // do nothing
+  void decref(PyObject* pyobj) const override {} // do nothing
+
+  bool try_incref(const c10::impl::PyObjectSlot& pyobj_slot) const override {
+    return false;
+  }
 
 #define PANIC(m)              \
   TORCH_INTERNAL_ASSERT(      \
@@ -20,6 +23,10 @@ struct NoopPyInterpreterVTable final : public PyInterpreterVTable {
       "attempted to call " #m \
       " on a Tensor with nontrivial PyObject after corresponding interpreter died")
 
+  size_t refcnt(PyObject* pyobj) const override {
+    PANIC(refcnt);
+  }
+
   c10::intrusive_ptr<TensorImpl> detach(const TensorImpl* self) const override {
     PANIC(detach);
   }

c10/core/impl/PyInterpreter.h

@@ -18,6 +18,9 @@ namespace c10 {
 struct IValue;
 class OperatorHandle;
 struct TensorImpl;
+namespace impl {
+struct PyObjectSlot;
+} // namespace impl
 } // namespace c10
 
 namespace torch::jit {
@@ -126,9 +129,12 @@ struct C10_API PyInterpreterVTable {
 
   // Run Py_INCREF on a PyObject.
   virtual void incref(PyObject* pyobj) const = 0;
-  // Run Py_DECREF on a PyObject.  We DO NOT assume the GIL is held on call
-  // See NOTE [PyInterpreter::decref takes a `has_pyobj_slot` arg]
-  virtual void decref(PyObject* pyobj, bool has_pyobj_slot) const = 0;
+  // Run Py_DECREF on a PyObject.  We DO NOT assume the GIL is held on call.
+  virtual void decref(PyObject* pyobj) const = 0;
+  // Run PyUnstable_TryIncRef on a PyObject if it's not NULL.
+  virtual bool try_incref(const c10::impl::PyObjectSlot& pyobj_slot) const = 0;
+  // Run Py_REFCNT on a PyObject.
+  virtual size_t refcnt(PyObject* pyobj) const = 0;
 
   // Perform a detach by deferring to the __torch_dispatch__ implementation of
   // detach, which will also arrange for the PyObject to get copied in this

c10/core/impl/PyObjectSlot.cpp

@@ -1,56 +0,0 @@
-#include <c10/core/impl/PyObjectSlot.h>
-
-namespace c10::impl {
-
-PyObjectSlot::PyObjectSlot() : pyobj_interpreter_(nullptr), pyobj_(nullptr) {}
-
-PyObjectSlot::~PyObjectSlot() {
-  maybe_destroy_pyobj();
-}
-
-void PyObjectSlot::maybe_destroy_pyobj() {
-  if (owns_pyobj()) {
-    TORCH_INTERNAL_ASSERT(pyobj_interpreter_ != nullptr);
-    TORCH_INTERNAL_ASSERT(pyobj_ != nullptr);
-    (*pyobj_interpreter_.load(std::memory_order_acquire))
-        ->decref(_unchecked_untagged_pyobj(), /*has_pyobj_slot*/ true);
-    // NB: this destructor can only be entered when there are no
-    // references to this C++ object (obviously), NOR any references
-    // to the PyObject (if there are references to the PyObject,
-    // then the PyObject holds an owning reference to the tensor).
-    // So it is OK to clear pyobj_ here as it is impossible for it to
-    // be used again (modulo weak reference races)
-    pyobj_ = nullptr; // for safety
-  }
-}
-
-PyInterpreter* PyObjectSlot::pyobj_interpreter() {
-  return pyobj_interpreter_.load(std::memory_order_acquire);
-}
-
-PyObject* PyObjectSlot::_unchecked_untagged_pyobj() const {
-  // NOLINTNEXTLINE(performance-no-int-to-ptr)
-  return reinterpret_cast<PyObject*>(
-      reinterpret_cast<uintptr_t>(pyobj_) & ~0x1ULL);
-}
-
-PyInterpreter& PyObjectSlot::load_pyobj_interpreter() const {
-  auto interpreter = pyobj_interpreter_.load(std::memory_order_acquire);
-  if (interpreter) {
-    return *interpreter;
-  }
-  TORCH_CHECK(false, "cannot access PyObject for Tensor - no interpreter set");
-}
-
-bool PyObjectSlot::owns_pyobj() {
-  // NOLINTNEXTLINE(performance-no-int-to-ptr)
-  return reinterpret_cast<uintptr_t>(pyobj_) & 1;
-}
-
-void PyObjectSlot::set_owns_pyobj(bool b) {
-  // NOLINTNEXTLINE(performance-no-int-to-ptr)
-  pyobj_ = reinterpret_cast<PyObject*>(
-      reinterpret_cast<uintptr_t>(_unchecked_untagged_pyobj()) | b);
-}
-
-} // namespace c10::impl

c10/core/impl/PyObjectSlot.h

@@ -8,117 +8,58 @@
 
 #include <atomic>
 
+namespace torch::utils {
+class PyObjectPreservation;
+}
+
 namespace c10::impl {
 
 struct C10_API PyObjectSlot {
  public:
-  PyObjectSlot();
-
-  ~PyObjectSlot();
-
-  void maybe_destroy_pyobj();
-
-  // Associate the TensorImpl with the specified PyObject, and, if necessary,
-  // also tag the interpreter.
-  //
-  // NB: This lives in a header so that we can inline away the switch on status
-  //
-  // NB: THIS FUNCTION CAN RAISE AN EXCEPTION.  Make sure to clean up after
-  // PyObject if necessary!
-  void init_pyobj(PyObject* pyobj) {
-    pyobj_interpreter_.store(
-        getGlobalPyInterpreter(), std::memory_order_relaxed);
-    pyobj_ = pyobj;
-  }
+  PyObjectSlot() : pyobj_interpreter_(nullptr), pyobj_(nullptr) {}
 
   // Query the PyObject interpreter.  This may return null if there is no
-  // interpreter.  This is racy!
-  PyInterpreter* pyobj_interpreter();
-
-  PyObject* _unchecked_untagged_pyobj() const;
-
-  // Test the interpreter tag.  If tagged for the current interpreter, return
-  // a non-nullopt (but possibly null) PyObject.  If (possibly) untagged,
-  // returns a nullopt.  If it is definitely invalid, raises an error.
-  //
-  // If `ignore_hermetic_tls` is false and this function is called from a
-  // hermetic context (ie, `HermeticPyObjectTLS::get_state()` is true), then
-  // nullopt is returned. If `ignore_hermetic_tls` is true, then the hermetic
-  // context is ignored, allowing you to check the interpreter tag of a
-  // nonhermetic PyObject from within a hermetic context. This is necessary
-  // because there are some cases where the deallocator function of a
-  // nonhermetic PyObject is called from within a hermetic context, so it must
-  // be properly treated as a nonhermetic PyObject.
-  //
-  // NB: this lives in header so that we can avoid actually creating the
-  // std::optional
-
-  // @todo alban: I'm not too sure what's going on here, we can probably delete
-  // it but it's worthwhile making sure
-  std::optional<PyObject*> check_pyobj(bool ignore_hermetic_tls = false) const {
-    impl::PyInterpreter* interpreter =
-        pyobj_interpreter_.load(std::memory_order_acquire);
-    if (interpreter == nullptr) {
-      return std::nullopt;
-    }
-
-    if (!ignore_hermetic_tls && c10::impl::HermeticPyObjectTLS::get_state()) {
-      return std::nullopt;
-    } else {
-      return _unchecked_untagged_pyobj();
-    }
+  // interpreter.
+  PyInterpreter* pyobj_interpreter() const {
+    return pyobj_interpreter_.load(std::memory_order_acquire);
   }
 
-  PyInterpreter& load_pyobj_interpreter() const;
+  PyInterpreter& load_pyobj_interpreter() const {
+    auto interpreter = pyobj_interpreter_.load(std::memory_order_acquire);
+    TORCH_INTERNAL_ASSERT(
+        interpreter, "cannot access PyObject for Tensor - no interpreter set");
+    return *interpreter;
+  }
 
-  bool owns_pyobj();
+  PyObject* load_pyobj() const {
+    return pyobj_.load(std::memory_order_acquire);
+  }
 
-  void set_owns_pyobj(bool b);
+  void store_pyobj(PyObject* obj) {
+    pyobj_.store(obj, std::memory_order_release);
+  }
+
+  bool has_unique_reference() const {
+    PyObject* pyobj = load_pyobj();
+    return pyobj != nullptr && load_pyobj_interpreter()->refcnt(pyobj) == 1;
+  }
+
+  void clear() {
+    pyobj_.store(nullptr, std::memory_order_relaxed);
+    pyobj_interpreter_.store(nullptr, std::memory_order_relaxed);
+  }
 
  private:
-  // This field contains the interpreter tag for this object.  See
-  // Note [Python interpreter tag] for general context
-  //
-  // Note [Memory ordering on Python interpreter tag]
-  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-  // What memory_order do we need when accessing this atomic?  We don't
-  // need a single total modification order (as provided by
-  // memory_order_seq_cst) as pyobj_interpreter_ is monotonic: it can only
-  // transition from -1 to some positive integer and never changes afterwards.
-  // Because there is only one modification, it trivially already has a total
-  // modification order (e.g., we don't need fences or locked instructions on
-  // x86)
-  //
-  // In fact, one could make a reasonable argument that relaxed reads are OK,
-  // due to the presence of external locking (GIL) to ensure that interactions
-  // with other data structures are still correctly synchronized, so that
-  // we fall in the "Single-Location Data Structures" case as described in
-  // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p2055r0.pdf
-  // However, on x86, it doesn't matter if I use acquire or relaxed on the load
-  // as I get the same assembly in both cases.  So I just use the more
-  // conservative acquire (which will impede compiler optimizations but I don't
-  // care)
+  // This is now always the global interpreter if the PyObject is set.
+  // Maybe we can remove this field some day...
   std::atomic<PyInterpreter*> pyobj_interpreter_;
 
-  // This field contains a reference to a PyObject representing this Tensor.
-  // If pyobj is nullptr, when we transfer Tensor to Python, we allocate a new
-  // PyObject for it and set this field.  This field does not have to be
-  // protected by an atomic as it is only allowed to be accessed when you hold
-  // the GIL, or during destruction of the tensor.
-  //
-  // When a PyObject dies, you are obligated to clear this field
-  // (otherwise, you will try to use-after-free the pyobj); this currently
-  // occurs in THPVariable_clear in torch/csrc/autograd/python_variable.cpp
-  //
-  // NB: Ordinarily, this should not be a strong reference, as if the
-  // PyObject owns the Tensor, this would create a reference cycle.
-  // However, sometimes this ownership flips.  To track who owns
-  // who, this has a single pointer tag indicating whether or not the
-  // C++ object owns the PyObject (the common case, zero, means PyObject
-  // owns the C++ object); see _unchecked_untagged_pyobj for raw access
-  // or check_pyobj for checked access.  See references to PyObject
-  // resurrection in torch/csrc/autograd/python_variable.cpp
-  PyObject* pyobj_;
+  // The PyObject representing this Tensor or nullptr. Ownership is managed
+  // by intrusive_ptr. By the time the PyObjectSlot is destroyed, this
+  // reference is already dead.
+  std::atomic<PyObject*> pyobj_;
+
+  friend class torch::utils::PyObjectPreservation;
 };
 
 } // namespace c10::impl

c10/util/Exception.cpp

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

c10/util/Logging.cpp

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

c10/util/intrusive_ptr.h

@@ -12,6 +12,10 @@ template <typename, typename...>
 class class_;
 }
 
+namespace torch::utils {
+class PyObjectPreservation;
+}
+
 namespace c10 {
 class intrusive_ptr_target;
 namespace raw {
@@ -33,6 +37,8 @@ constexpr uint64_t kImpracticallyHugeWeakReferenceCount =
 constexpr uint64_t kReferenceCountOne = 1;
 constexpr uint64_t kWeakReferenceCountOne = (kReferenceCountOne << 32);
 constexpr uint64_t kUniqueRef = (kReferenceCountOne | kWeakReferenceCountOne);
+// Indicates whether the object has a PyObject wrapper.
+constexpr uint64_t kHasPyObject = (uint64_t(1) << 63);
 
 template <class TTarget>
 struct intrusive_target_default_null_type final {
@@ -55,7 +61,15 @@ inline uint32_t refcount(uint64_t combined_refcount) {
 }
 
 inline uint32_t weakcount(uint64_t combined_refcount) {
-  return static_cast<uint32_t>(combined_refcount >> 32);
+  return static_cast<uint32_t>((combined_refcount & ~kHasPyObject) >> 32);
+}
+
+inline bool has_pyobject(uint64_t combined_refcount) {
+  return (combined_refcount & kHasPyObject) != 0;
+}
+
+inline bool is_uniquely_owned(uint64_t combined_refcount) {
+  return (combined_refcount & ~detail::kHasPyObject) == detail::kUniqueRef;
 }
 
 // The only requirement for refcount increment is that it happens-before
@@ -66,12 +80,6 @@ inline uint64_t atomic_combined_refcount_increment(
   return combined_refcount.fetch_add(inc, std::memory_order_relaxed) + inc;
 }
 
-inline uint32_t atomic_refcount_increment(
-    std::atomic<uint64_t>& combined_refcount) {
-  return detail::refcount(atomic_combined_refcount_increment(
-      combined_refcount, kReferenceCountOne));
-}
-
 inline uint32_t atomic_weakcount_increment(
     std::atomic<uint64_t>& combined_refcount) {
   return detail::weakcount(atomic_combined_refcount_increment(
@@ -99,6 +107,11 @@ inline uint32_t atomic_weakcount_decrement(
       combined_refcount, kWeakReferenceCountOne));
 }
 
+template <class T, class = void>
+struct TargetTraits {
+  static constexpr bool can_have_pyobject = false;
+};
+
 } // namespace detail
 
 /**
@@ -155,6 +168,23 @@ class C10_API intrusive_ptr_target {
   // we can atomically operate on both at the same time for performance
   // and defined behaviors.
   //
+  // Note [PyObject preservation for Tensor and Storages]
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // intrusive_ptr has special support for preserving PyObject wrappers
+  // for TensorImpl and StorageImpl. The most significant bit (kHasPyObject) of
+  // the combined_refcount_ is used to indicate whether the object has a
+  // PyObject wrapper.
+  //
+  //   - The PyObject, if it exists, holds a strong reference to the
+  //     intrusive_ptr_target.
+  //
+  //   - When the refcount goes from 1 to 2, we incref the PyObject.
+  //
+  //   - When the refcount goes from 2 to 1, we decref the PyObject.
+  //
+  // In other words, the intrusive_ptr keeps the PyObject alive as long as there
+  // are other C++ references to the intrusive_ptr_target.
+
   mutable std::atomic<uint64_t> combined_refcount_;
   static_assert(sizeof(std::atomic<uint64_t>) == 8);
   static_assert(alignof(std::atomic<uint64_t>) == 8);
@@ -172,6 +202,8 @@ class C10_API intrusive_ptr_target {
   template <typename T>
   friend struct ExclusivelyOwnedTensorTraits;
 
+  friend class torch::utils::PyObjectPreservation;
+
  protected:
   // protected destructor. We never want to destruct intrusive_ptr_target*
   // directly.
@@ -255,6 +287,16 @@ class C10_API intrusive_ptr_target {
    */
   virtual void release_resources() {}
 
+  /**
+   * These two methods are called when the refcount transitions between one
+   * and two and the object has a PyObject wrapper.
+   */
+  virtual void incref_pyobject() const noexcept {}
+  virtual void decref_pyobject() const noexcept {}
+  virtual bool try_incref_pyobject() const noexcept {
+    return false;
+  }
+
   uint32_t refcount(std::memory_order order = std::memory_order_relaxed) const {
     return detail::refcount(combined_refcount_.load(order));
   }
@@ -265,6 +307,19 @@ class C10_API intrusive_ptr_target {
   }
 };
 
+namespace detail {
+
+#ifndef C10_MOBILE
+template <>
+struct TargetTraits<c10::intrusive_ptr_target> {
+  // A generic intrusive_ptr<intrusive_ptr_target> may actually be a TensorImpl
+  // or StorageImpl, so we have to allow for PyObject support.
+  static constexpr bool can_have_pyobject = true;
+};
+#endif
+
+} // namespace detail
+
 template <class TTarget, class NullType>
 class weak_intrusive_ptr;
 
@@ -312,51 +367,87 @@ class intrusive_ptr final {
   template <typename, typename...>
   friend class pybind11::class_;
 
-  void retain_() {
+  void retain_() noexcept {
     if (target_ != NullType::singleton()) {
-      uint32_t new_refcount =
-          detail::atomic_refcount_increment(target_->combined_refcount_);
+      uint64_t combined = detail::atomic_combined_refcount_increment(
+          target_->combined_refcount_, detail::kReferenceCountOne);
+      uint32_t new_refcount = detail::refcount(combined);
       TORCH_INTERNAL_ASSERT_DEBUG_ONLY(
           new_refcount != 1,
           "intrusive_ptr: Cannot increase refcount after it reached zero.");
+
+      if constexpr (detail::TargetTraits<TTarget>::can_have_pyobject) {
+        // If the refcount transitioned from 1 to 2, we need to incref the
+        // PyObject. In other words, we need to ensure that the PyObject stays
+        // alive now that we have a C++ reference to this object in addition to
+        // the PyObject itself.
+        if (detail::has_pyobject(combined) && detail::refcount(combined) == 2) {
+          target_->incref_pyobject();
+        }
+      } else {
+        TORCH_INTERNAL_ASSERT_DEBUG_ONLY(
+            !detail::has_pyobject(combined),
+            "TargetTraits indicates that type cannot have PyObject, but refcount has PyObject bit set.");
+      }
     }
   }
 
   void reset_() noexcept {
     if (target_ != NullType::singleton()) {
-      if (target_->combined_refcount_.load(std::memory_order_acquire) ==
-          detail::kUniqueRef) {
-        // Both counts are 1, so there are no weak references and
-        // we are releasing the last strong reference. No other
-        // threads can observe the effects of this target_ deletion
-        // call (e.g. calling use_count()) without a data race.
-        target_->combined_refcount_.store(0, std::memory_order_relaxed);
-        delete target_;
+      reset_not_null_(target_);
+    }
+  }
+
+  // C10_NOINLINE to keep binary size a bit smaller. We pass TTarget* here
+  // to avoid an extra pointer dereference in the call from reset_().
+  C10_NOINLINE static void reset_not_null_(TTarget* target) noexcept {
+    if (detail::is_uniquely_owned(
+            target->combined_refcount_.load(std::memory_order_acquire))) {
+      // Both counts are 1, so there are no weak references and
+      // we are releasing the last strong reference. No other
+      // threads can observe the effects of this target deletion
+      // call (e.g. calling use_count()) without a data race.
+      target->combined_refcount_.store(0, std::memory_order_relaxed);
+      delete target;
+      return;
+    }
+
+    auto combined_refcount = detail::atomic_combined_refcount_decrement(
+        target->combined_refcount_, detail::kReferenceCountOne);
+    uint32_t new_refcount = detail::refcount(combined_refcount);
+    bool has_pyobject = detail::has_pyobject(combined_refcount);
+    if (new_refcount == 0) {
+      if (detail::weakcount(combined_refcount) == 1) {
+        delete target;
         return;
       }
-
-      auto combined_refcount = detail::atomic_combined_refcount_decrement(
-          target_->combined_refcount_, detail::kReferenceCountOne);
-      if (detail::refcount(combined_refcount) == 0) {
-        bool should_delete =
-            (combined_refcount == detail::kWeakReferenceCountOne);
-        // See comment above about weakcount. As long as refcount>0,
-        // weakcount is one larger than the actual number of weak references.
-        // So we need to decrement it here.
-        if (!should_delete) {
-          // justification for const_cast: release_resources is basically a
-          // destructor and a destructor always mutates the object, even for
-          // const objects.
-          // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast)
-          const_cast<std::remove_const_t<TTarget>*>(target_)
-              ->release_resources();
-          should_delete = detail::atomic_weakcount_decrement(
-                              target_->combined_refcount_) == 0;
-        }
-        if (should_delete) {
-          delete target_;
-        }
+      // See comment above about weakcount. As long as refcount>0,
+      // weakcount is one larger than the actual number of weak references.
+      // So we need to decrement it here.
+      release_resources_and_decrement_weakrefs_(target);
+    } else if constexpr (detail::TargetTraits<TTarget>::can_have_pyobject) {
+      // If the refcount transitioned from 2 to 1, we need to decref the
+      // PyObject. In other words, we don't want to keep the PyObject alive if
+      // there are no C++ references to this object other than the PyObject
+      // itself.
+      if (has_pyobject && new_refcount == 1) {
+        target->decref_pyobject();
       }
+    } else {
+      TORCH_INTERNAL_ASSERT_DEBUG_ONLY(
+          !has_pyobject,
+          "TargetTraits indicates that type cannot have PyObject, but refcount has PyObject bit set.");
+    }
+  }
+
+  C10_NOINLINE static void release_resources_and_decrement_weakrefs_(
+      TTarget* target) noexcept {
+    // justification for const_cast: release_resources is basically a
+    // destructor and a destructor always mutates the object, even for
+    // const objects.
+    const_cast<std::remove_const_t<TTarget>*>(target)->release_resources();
+    if (detail::atomic_weakcount_decrement(target->combined_refcount_) == 0) {
+      delete target;
     }
   }
 
@@ -522,6 +613,15 @@ class intrusive_ptr final {
     return use_count() == 1;
   }
 
+  /**
+   * Stronger than unique() in that it must not have any weakrefs as well.
+   */
+  bool is_uniquely_owned() const noexcept {
+    TORCH_INTERNAL_ASSERT_DEBUG_ONLY(target_ != NullType::singleton());
+    return detail::is_uniquely_owned(
+        target_->combined_refcount_.load(std::memory_order_acquire));
+  }
+
   /**
    * Returns an owning (!) pointer to the underlying object and makes the
    * intrusive_ptr instance invalid. That means the refcount is not decreased.
@@ -932,6 +1032,7 @@ class weak_intrusive_ptr final {
     if (target_ == NullType::singleton()) {
       return intrusive_ptr<TTarget, NullType>();
     } else {
+      bool increfed = false;
       auto combined_refcount =
           target_->combined_refcount_.load(std::memory_order_relaxed);
       do {
@@ -940,12 +1041,31 @@ class weak_intrusive_ptr final {
           // Return nullptr.
           return intrusive_ptr<TTarget, NullType>();
         }
+        if constexpr (detail::TargetTraits<TTarget>::can_have_pyobject) {
+          if (detail::has_pyobject(combined_refcount) &&
+              detail::refcount(combined_refcount) == 1 && !increfed) {
+            // Object has a python wrapper with no other C++ references.
+            // We need to to incref the Python object before we acquire a
+            // strong reference to the C++ object to avoid a situation
+            // where the Python object is deallocated concurrently.
+            if (!target_->try_incref_pyobject()) {
+              return intrusive_ptr<TTarget, NullType>();
+            }
+            increfed = true;
+          }
+        }
       } while (!target_->combined_refcount_.compare_exchange_weak(
           combined_refcount,
           combined_refcount + detail::kReferenceCountOne,
           std::memory_order_acquire,
           std::memory_order_relaxed));
 
+      if constexpr (detail::TargetTraits<TTarget>::can_have_pyobject) {
+        if (increfed && detail::refcount(combined_refcount) != 1) {
+          target_->decref_pyobject();
+        }
+      }
+
       return intrusive_ptr<TTarget, NullType>(
           target_, raw::DontIncreaseRefcount{});
     }
@@ -1060,7 +1180,16 @@ namespace intrusive_ptr {
 // NullType::singleton to this function
 inline void incref(intrusive_ptr_target* self) {
   if (self) {
-    detail::atomic_refcount_increment(self->combined_refcount_);
+    uint64_t combined = detail::atomic_combined_refcount_increment(
+        self->combined_refcount_, detail::kReferenceCountOne);
+
+#ifndef C10_MOBILE
+    if (detail::has_pyobject(combined) && detail::refcount(combined) == 2) {
+      self->incref_pyobject();
+    }
+#else
+    TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!detail::has_pyobject(combined));
+#endif
   }
 }
 

c10/xpu/XPUCachingAllocator.cpp

@@ -15,6 +15,8 @@ using namespace c10::CachingDeviceAllocator;
 // newly allocated memory with 512-byte alignment.
 constexpr size_t kDeviceAlignment = 512;
 
+class XPUAllocator;
+
 namespace {
 using stream_set = ska::flat_hash_set<xpu::XPUStream>;
 
@@ -23,14 +25,19 @@ typedef bool (*Comparison)(const Block*, const Block*);
 bool BlockComparatorSize(const Block* a, const Block* b);
 bool BlockComparatorAddress(const Block* a, const Block* b);
 
+struct PrivatePool;
+
 struct BlockPool {
-  BlockPool(bool small)
+  BlockPool(bool small, PrivatePool* private_pool = nullptr)
       : blocks(BlockComparatorSize),
         unmapped(BlockComparatorAddress),
-        is_small(small) {}
+        is_small(small),
+        owner_PrivatePool(private_pool) {}
+
   std::set<Block*, Comparison> blocks;
   std::set<Block*, Comparison> unmapped;
   const bool is_small;
+  PrivatePool* owner_PrivatePool;
 };
 
 struct ExpandableSegment;
@@ -349,6 +356,43 @@ struct AllocParams {
   StatTypes stat_types = {};
 };
 
+// Internal implementation that manages actual memory blocks.
+// high level MemPool interface wraps PrivatePool via MempoolId.
+struct PrivatePool {
+  PrivatePool(MempoolId_t id, XPUAllocator* allocator = nullptr)
+      : id(std::move(id)),
+        allocator_(allocator),
+        large_blocks(/*small=*/false, this),
+        small_blocks(/*small=*/true, this) {}
+  PrivatePool(const PrivatePool&) = delete;
+  PrivatePool(PrivatePool&&) = delete;
+  PrivatePool& operator=(const PrivatePool&) = delete;
+  PrivatePool& operator=(PrivatePool&&) = delete;
+  ~PrivatePool() = default;
+
+  // default Mempool when no Mempool is specified
+  MempoolId_t id{0, 0};
+  // Number of live graphs using this pool
+  int use_count{1};
+  // Number of unfreed allocations made for this pool. When use_count and
+  // allocation_count drop to zero, we can delete this PrivatePool from
+  // graph_pools.
+  int allocation_count{0};
+  XPUAllocator* allocator_;
+  BlockPool large_blocks;
+  BlockPool small_blocks;
+
+ public:
+  XPUAllocator* allocator() {
+    return allocator_;
+  }
+};
+struct MempoolIdHash {
+  std::size_t operator()(const MempoolId_t& mempool_id) const noexcept {
+    return mempool_id.first != 0 ? mempool_id.first : mempool_id.second;
+  }
+};
+
 } // anonymous namespace
 
 class DeviceCachingAllocator {
@@ -365,6 +409,13 @@ class DeviceCachingAllocator {
   bool set_fraction = false;
   std::vector<ExpandableSegment*> expandable_segments;
   std::vector<c10::DeviceIndex> devices_with_peer_access; // reserved
+  std::vector<std::pair<MempoolId_t, std::function<bool(sycl::queue*)>>>
+      captures_underway;
+  ska::flat_hash_map<MempoolId_t, std::unique_ptr<PrivatePool>, MempoolIdHash>
+      graph_pools;
+  // Pools no longer referenced by any graph.
+  ska::flat_hash_map<MempoolId_t, PrivatePool*, MempoolIdHash>
+      graph_pools_freeable;
 
   size_t try_merge_blocks(Block* dst, Block* src, BlockPool& pool) {
     if (!src || src->allocated || src->event_count > 0 ||
@@ -463,7 +514,22 @@ class DeviceCachingAllocator {
     }
   }
 
-  BlockPool& get_pool(size_t size) {
+  BlockPool& get_pool(size_t size, sycl::queue* queue) {
+    if (C10_UNLIKELY(!captures_underway.empty())) {
+      for (auto& entry : captures_underway) {
+        // lookup for mempool id matching current capture graph
+        if (entry.second(queue)) {
+          auto it1 = graph_pools.find(entry.first);
+          // lookup mempool
+          TORCH_INTERNAL_ASSERT(it1 != graph_pools.end());
+          if (size <= kSmallSize) {
+            return it1->second->small_blocks;
+          } else {
+            return it1->second->large_blocks;
+          }
+        }
+      }
+    }
     if (size < kSmallSize) {
       return small_blocks;
     } else {
@@ -669,6 +735,10 @@ class DeviceCachingAllocator {
     if (!ptr) {
       return false;
     }
+
+    if (p.pool->owner_PrivatePool) {
+      p.pool->owner_PrivatePool->allocation_count++;
+    }
     p.block = new Block(device, p.queue(), size, p.pool, ptr);
     for_each_selected_stat_type(p.stat_types, [&](size_t stat_type) {
       stats.reserved_bytes[stat_type].increase(size);
@@ -677,11 +747,14 @@ class DeviceCachingAllocator {
     return true;
   }
 
-  void synchronize_and_free_events() {
+  void synchronize_and_free_events(PrivatePool* pool = nullptr) {
     for (auto& xe : xpu_events) {
       for (auto& e : xe.second) {
         auto event = e.first;
         auto* block = e.second;
+        if (pool && block->pool->owner_PrivatePool != pool) {
+          continue;
+        }
         event.wait();
         block->event_count--;
         if (block->event_count == 0) {
@@ -785,6 +858,13 @@ class DeviceCachingAllocator {
     for_each_selected_stat_type(stat_types, [&](size_t stat_type) {
       stats.reserved_bytes[stat_type].decrease(unmapped.size);
     });
+
+    if (block->pool->owner_PrivatePool) {
+      // The Freed block belonged to a XPU graph's PrivatePool.
+      TORCH_INTERNAL_ASSERT(
+          block->pool->owner_PrivatePool->allocation_count > 0);
+      block->pool->owner_PrivatePool->allocation_count--;
+    }
   }
 
   void release_blocks(BlockPool& pool) {
@@ -812,13 +892,49 @@ class DeviceCachingAllocator {
     }
   }
 
-  bool release_cached_blocks() {
-    synchronize_and_free_events();
-    // See Note [Safe to Free Blocks on BlockPool]
-    c10::xpu::syncStreamsOnDevice(device_index);
+  bool release_cached_blocks(MempoolId_t mempool_id) {
+    bool streams_synced = false;
+    if (mempool_id.first == 0 && mempool_id.second == 0 &&
+        captures_underway.empty()) {
+      synchronize_and_free_events();
+      // See Note [Safe to Free Blocks on BlockPool]
+      c10::xpu::syncStreamsOnDevice(device_index);
+      streams_synced = true;
 
-    release_blocks(large_blocks);
-    release_blocks(small_blocks);
+      release_blocks(large_blocks);
+      release_blocks(small_blocks);
+    }
+
+    for (auto it = graph_pools_freeable.begin();
+         it != graph_pools_freeable.end();) {
+      if (mempool_id.first != 0 || mempool_id.second != 0) {
+        if (it->first == mempool_id) {
+          // If there is an active mempool, we sync only the events
+          // associated with the pool
+          synchronize_and_free_events(it->second);
+        } else {
+          // otherwise we move on
+          ++it;
+          continue;
+        }
+      }
+
+      if (!streams_synced) {
+        // See Note [Safe to Free Blocks on BlockPool]
+        c10::xpu::syncStreamsOnDevice(device_index);
+        streams_synced = true;
+      }
+      TORCH_INTERNAL_ASSERT(it->second->use_count == 0);
+      release_blocks(it->second->small_blocks);
+      release_blocks(it->second->large_blocks);
+      if (it->second->allocation_count == 0) {
+        auto erase_count = graph_pools.erase(it->first);
+        TORCH_INTERNAL_ASSERT(erase_count == 1);
+        it = graph_pools_freeable.erase(it);
+      } else {
+        ++it;
+      }
+    }
     return true;
   }
 
@@ -903,6 +1019,30 @@ class DeviceCachingAllocator {
     }
   }
 
+  void create_or_incref_pool(
+      MempoolId_t mempool_id,
+      XPUAllocator* allocator = nullptr) {
+    auto it = graph_pools.find(mempool_id);
+    if (it == graph_pools.end()) {
+      // mempool_id does not reference an existing pool.
+      // Make a new pool for XPU graph capture or memory pool usage.
+      graph_pools.emplace(
+          mempool_id, std::make_unique<PrivatePool>(mempool_id, allocator));
+    } else {
+      // mempool_id references an existing pool, which the current XPU graph
+      // capture will share.
+      TORCH_INTERNAL_ASSERT(it->second->use_count > 0);
+      TORCH_INTERNAL_ASSERT(allocator == nullptr);
+      it->second->use_count++;
+    }
+  }
+
+  PrivatePool* get_private_pool(MempoolId_t mempool_id) {
+    auto it = graph_pools.find(mempool_id);
+    TORCH_INTERNAL_ASSERT(it != graph_pools.end());
+    return it->second.get();
+  }
+
  public:
   DeviceCachingAllocator(DeviceIndex device_index)
       : large_blocks(/* small */ false),
@@ -911,9 +1051,11 @@ class DeviceCachingAllocator {
 
   Block* malloc(DeviceIndex device, size_t orig_size, sycl::queue& queue) {
     std::scoped_lock<std::recursive_mutex> lock(mutex);
-    process_events();
+    if (C10_LIKELY(captures_underway.empty())) {
+      process_events();
+    }
     size_t size = round_size(orig_size);
-    auto& pool = get_pool(size);
+    auto& pool = get_pool(size, &queue);
     const size_t alloc_size = get_allocation_size(size);
     AllocParams params(device, size, &queue, &pool, alloc_size);
     params.stat_types = get_stat_types_for_pool(pool);
@@ -923,7 +1065,7 @@ class DeviceCachingAllocator {
     // Can't reuse an existing block, try to get a new one.
     if (!block_found) {
       block_found = alloc_block(params, false) ||
-          (release_cached_blocks() && alloc_block(params, true));
+          (release_cached_blocks({0, 0}) && alloc_block(params, true));
     }
     if (!block_found) {
       const auto& raw_device = c10::xpu::get_raw_device(device);
@@ -1016,9 +1158,9 @@ class DeviceCachingAllocator {
     block->stream_uses.insert(stream);
   }
 
-  void emptyCache() {
+  void emptyCache(MempoolId_t mempool_id) {
     std::scoped_lock<std::recursive_mutex> lock(mutex);
-    release_cached_blocks();
+    release_cached_blocks(mempool_id);
   }
 
   DeviceStats getStats() {
@@ -1085,6 +1227,63 @@ class DeviceCachingAllocator {
     allowed_memory_maximum = static_cast<size_t>(fraction * device_total);
     set_fraction = true;
   }
+
+  void createOrIncrefPool(
+      MempoolId_t mempool_id,
+      XPUAllocator* allocator = nullptr) {
+    std::scoped_lock<std::recursive_mutex> lock(mutex);
+    create_or_incref_pool(mempool_id, allocator);
+  }
+
+  int getPoolUseCount(MempoolId_t mempool_id) {
+    std::scoped_lock<std::recursive_mutex> lock(mutex);
+    auto it = graph_pools.find(mempool_id);
+    if (it == graph_pools.end()) {
+      return 0;
+    }
+    return it->second->use_count;
+  }
+
+  // Called by XPUGraph::capture_begin
+  void beginAllocateToPool(
+      MempoolId_t mempool_id,
+      std::function<bool(sycl::queue*)> filter) {
+    std::lock_guard<std::recursive_mutex> lock(mutex);
+    create_or_incref_pool(mempool_id);
+    auto not_found = std::all_of(
+        captures_underway.begin(),
+        captures_underway.end(),
+        [&](const auto& entry) { return entry.first != mempool_id; });
+    TORCH_CHECK(
+        not_found, "beginAllocateToPool: already recording to mempool_id");
+    captures_underway.emplace_back(mempool_id, std::move(filter));
+  }
+
+  // Called by XPUGraph::capture_end
+  void endAllocateToPool(MempoolId_t mempool_id) {
+    std::lock_guard<std::recursive_mutex> lock(mutex);
+
+    auto it = std::find_if(
+        captures_underway.begin(),
+        captures_underway.end(),
+        [&](const auto& entry) { return entry.first == mempool_id; });
+    TORCH_INTERNAL_ASSERT(
+        it != captures_underway.end(),
+        "endAllocatePool: not currently recording to mempool_id");
+    captures_underway.erase(it);
+  }
+
+  // Called by XPUGraph::reset and MemPool::~MemPool()
+  void releasePool(MempoolId_t mempool_id) {
+    std::lock_guard<std::recursive_mutex> lock(mutex);
+    auto pp = get_private_pool(mempool_id);
+    auto uc = --(pp->use_count);
+    TORCH_INTERNAL_ASSERT(uc >= 0);
+    if (uc == 0) {
+      bool inserted = graph_pools_freeable.insert({mempool_id, pp}).second;
+      TORCH_INTERNAL_ASSERT(inserted);
+    }
+  }
 };
 
 static void local_raw_delete(void* ptr);
@@ -1172,9 +1371,9 @@ class XPUAllocator : public DeviceAllocator {
     }
   }
 
-  void emptyCache(MempoolId_t mempool_id [[maybe_unused]] = {0, 0}) override {
+  void emptyCache(MempoolId_t mempool_id) override {
     for (auto& da : device_allocators) {
-      da->emptyCache();
+      da->emptyCache(mempool_id);
     }
   }
 
@@ -1274,6 +1473,39 @@ class XPUAllocator : public DeviceAllocator {
         ". Please set within (0, 1].");
     device_allocators[device]->setMemoryFraction(fraction);
   }
+
+  void createOrIncrefPool(
+      c10::DeviceIndex device,
+      MempoolId_t mempool_id,
+      XPUAllocator* allocator) {
+    assertValidDevice(device);
+    device_allocators[device]->createOrIncrefPool(
+        std::move(mempool_id), allocator);
+  }
+
+  void beginAllocateToPool(
+      c10::DeviceIndex device,
+      MempoolId_t mempool_id,
+      std::function<bool(sycl::queue*)> filter) {
+    assertValidDevice(device);
+    device_allocators[device]->beginAllocateToPool(
+        std::move(mempool_id), std::move(filter));
+  }
+
+  void endAllocateToPool(c10::DeviceIndex device, MempoolId_t mempool_id) {
+    assertValidDevice(device);
+    device_allocators[device]->endAllocateToPool(mempool_id);
+  }
+
+  void releasePool(c10::DeviceIndex device, MempoolId_t mempool_id) {
+    assertValidDevice(device);
+    device_allocators[device]->releasePool(std::move(mempool_id));
+  }
+
+  int getPoolUseCount(c10::DeviceIndex device, MempoolId_t mempool_id) {
+    assertValidDevice(device);
+    return device_allocators[device]->getPoolUseCount(std::move(mempool_id));
+  }
 };
 
 static XPUAllocator allocator;
@@ -1290,8 +1522,8 @@ void init(DeviceIndex device_count) {
   return allocator.init(device_count);
 }
 
-void emptyCache() {
-  return allocator.emptyCache();
+void emptyCache(MempoolId_t mempool_id) {
+  return allocator.emptyCache(mempool_id);
 }
 
 void resetPeakStats(DeviceIndex device) {
@@ -1330,6 +1562,92 @@ void setMemoryFraction(double fraction, DeviceIndex device) {
   return allocator.setMemoryFraction(fraction, device);
 }
 
+void createOrIncrefPool(
+    c10::DeviceIndex device,
+    MempoolId_t mempool_id,
+    XPUAllocator* allocator_ptr) {
+  return allocator.createOrIncrefPool(device, mempool_id, allocator_ptr);
+}
+
+void beginAllocateToPool(
+    c10::DeviceIndex device,
+    MempoolId_t mempool_id,
+    std::function<bool(sycl::queue*)> filter) {
+  return allocator.beginAllocateToPool(device, mempool_id, std::move(filter));
+}
+
+void endAllocateToPool(c10::DeviceIndex device, MempoolId_t mempool_id) {
+  return allocator.endAllocateToPool(device, mempool_id);
+}
+
+void releasePool(c10::DeviceIndex device, MempoolId_t mempool_id) {
+  return allocator.releasePool(device, mempool_id);
+}
+
+int getPoolUseCount(c10::DeviceIndex device, MempoolId_t mempool_id) {
+  return allocator.getPoolUseCount(device, mempool_id);
+}
+
 REGISTER_ALLOCATOR(kXPU, &allocator)
 
 } // namespace c10::xpu::XPUCachingAllocator
+
+namespace c10::xpu {
+
+// uid_ is incremented when a user creates a MemPool,
+//
+// uuid_ is incremented when XPUGraph creates a MemPool
+// as a result of a user not providing a pool.
+
+std::atomic<CaptureId_t> MemPool::uid_{1};
+std::atomic<CaptureId_t> MemPool::uuid_{1};
+
+MemPool::MemPool(
+    XPUCachingAllocator::XPUAllocator* allocator,
+    bool is_user_created,
+    bool use_on_oom)
+    : allocator_(allocator), is_user_created_(is_user_created) {
+  if (is_user_created_) {
+    id_ = {0, uid_++};
+  } else {
+    id_ = {uuid_++, 0};
+  }
+  device_ = c10::xpu::current_device();
+  XPUCachingAllocator::createOrIncrefPool(device_, id_, allocator);
+  if (use_on_oom) {
+    // XPU doesn't support use_on_oom yet
+    TORCH_WARN(
+        "XPUCachingAllocator::MemPool: use_on_oom is not supported on XPU");
+  }
+}
+
+MemPool::~MemPool() {
+  TORCH_INTERNAL_ASSERT(use_count() == 1);
+  XPUCachingAllocator::releasePool(device_, id_);
+  c10::xpu::XPUCachingAllocator::emptyCache(id_); // release cached blocks
+}
+
+MempoolId_t MemPool::id() {
+  return id_;
+}
+
+XPUCachingAllocator::XPUAllocator* MemPool::allocator() {
+  return allocator_;
+}
+
+int MemPool::use_count() {
+  return XPUCachingAllocator::getPoolUseCount(device_, id_);
+}
+
+c10::DeviceIndex MemPool::device() {
+  return device_;
+}
+
+MempoolId_t MemPool::graph_pool_handle(bool is_user_created) {
+  if (is_user_created) {
+    return {0, uid_++};
+  }
+  return {uuid_++, 0};
+}
+
+} // namespace c10::xpu