bench

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -3,15 +3,22 @@ set -eux -o pipefail
 
 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
-if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="9.0"
-elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="9.0;10.0;12.0"
-fi
-
 SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
 source $SCRIPTPATH/aarch64_ci_setup.sh
 
+tagged_version() {
+  GIT_DESCRIBE="git --git-dir /pytorch/.git describe --tags --match v[0-9]*.[0-9]*.[0-9]*"
+  if ${GIT_DESCRIBE} --exact >/dev/null; then
+    ${GIT_DESCRIBE}
+  else
+    return 1
+  fi
+}
+
+if tagged_version >/dev/null; then
+  export OVERRIDE_PACKAGE_VERSION="$(tagged_version | sed -e 's/^v//' -e 's/-.*$//')"
+fi
+
 ###############################################################################
 # Run aarch64 builder python
 ###############################################################################

.ci/aarch64_linux/aarch64_ci_setup.sh

@@ -5,14 +5,16 @@ set -eux -o pipefail
 # 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
+PYGIT2_VERSION=1.15.1
+if [[ "$DESIRED_PYTHON"  == "3.13" ]]; then
     NUMPY_VERSION=2.1.2
+    PYGIT2_VERSION=1.16.0
 fi
 
 SCRIPTPATH="$( cd "$(dirname "$0")" ; pwd -P )"
 source $SCRIPTPATH/../manywheel/set_desired_python.sh
 
-pip install -q numpy==${NUMPY_VERSION} pyyaml==6.0.2 scons==4.7.0 ninja==1.11.1 patchelf==0.17.2
+pip install -q numpy==${NUMPY_VERSION} pyyaml==6.0.2 scons==4.7.0 ninja==1.11.1 patchelf==0.17.2 pygit2==${PYGIT2_VERSION}
 
 for tool in python python3 pip pip3 ninja scons patchelf; do
     ln -sf ${DESIRED_PYTHON_BIN_DIR}/${tool} /usr/local/bin;

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -4,9 +4,12 @@
 import os
 import shutil
 from subprocess import check_call, check_output
+from typing import List
+
+from pygit2 import Repository
 
 
-def list_dir(path: str) -> list[str]:
+def list_dir(path: str) -> List[str]:
     """'
     Helper for getting paths for Python
     """
@@ -55,7 +58,7 @@ def build_ArmComputeLibrary() -> None:
         shutil.copytree(f"{acl_checkout_dir}/{d}", f"{acl_install_dir}/{d}")
 
 
-def update_wheel(wheel_path, desired_cuda) -> None:
+def update_wheel(wheel_path) -> None:
     """
     Update the cuda wheel libraries
     """
@@ -77,6 +80,7 @@ def update_wheel(wheel_path, desired_cuda) -> None:
         "/usr/local/cuda/lib64/libnvToolsExt.so.1",
         "/usr/local/cuda/lib64/libnvJitLink.so.12",
         "/usr/local/cuda/lib64/libnvrtc.so.12",
+        "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.6",
         "/usr/local/cuda/lib64/libcudnn_adv.so.9",
         "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
         "/usr/local/cuda/lib64/libcudnn_graph.so.9",
@@ -96,14 +100,6 @@ def update_wheel(wheel_path, desired_cuda) -> None:
             "/usr/local/lib/libnvpl_lapack_core.so.0",
             "/usr/local/lib/libnvpl_blas_core.so.0",
         ]
-        if "126" in desired_cuda:
-            libs_to_copy += [
-                "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.6",
-            ]
-        elif "128" in desired_cuda:
-            libs_to_copy += [
-                "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.8",
-            ]
     else:
         libs_to_copy += [
             "/opt/OpenBLAS/lib/libopenblas.so.0",
@@ -175,22 +171,22 @@ if __name__ == "__main__":
     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()
+    repo = Repository("/pytorch")
+    branch = repo.head.name
+    if branch == "HEAD":
+        branch = "master"
 
     print("Building PyTorch wheel")
     build_vars = "MAX_JOBS=5 CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
     os.system("cd /pytorch; python setup.py clean")
 
     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"]:
+    elif branch in ["nightly", "master"]:
         build_date = (
             check_output(["git", "log", "--pretty=format:%cs", "-1"], cwd="/pytorch")
             .decode()
@@ -200,6 +196,7 @@ if __name__ == "__main__":
             check_output(["cat", "version.txt"], cwd="/pytorch").decode().strip()[:-2]
         )
         if enable_cuda:
+            desired_cuda = os.getenv("DESIRED_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 "
@@ -228,6 +225,6 @@ if __name__ == "__main__":
         print("Updating Cuda Dependency")
         filename = os.listdir("/pytorch/dist/")
         wheel_path = f"/pytorch/dist/{filename[0]}"
-        update_wheel(wheel_path, desired_cuda)
+        update_wheel(wheel_path)
     pytorch_wheel_name = complete_wheel("/pytorch/")
     print(f"Build Complete. Created {pytorch_wheel_name}..")

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -12,7 +12,7 @@ import os
 import subprocess
 import sys
 import time
-from typing import Optional, Union
+from typing import Dict, List, Optional, Tuple, Union
 
 import boto3
 
@@ -24,12 +24,10 @@ os_amis = {
     "ubuntu22_04": "ami-0c6c29c5125214c77",  # login_name: ubuntu
     "redhat8": "ami-0698b90665a2ddcf1",  # login_name: ec2-user
 }
-
 ubuntu18_04_ami = os_amis["ubuntu18_04"]
-ubuntu20_04_ami = os_amis["ubuntu20_04"]
 
 
-def compute_keyfile_path(key_name: Optional[str] = None) -> tuple[str, str]:
+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:
@@ -59,7 +57,7 @@ def ec2_instances_by_id(instance_id):
 
 
 def start_instance(
-    key_name, ami=ubuntu20_04_ami, instance_type="t4g.2xlarge", ebs_size: int = 50
+    key_name, ami=ubuntu18_04_ami, instance_type="t4g.2xlarge", ebs_size: int = 50
 ):
     inst = ec2.create_instances(
         ImageId=ami,
@@ -98,7 +96,7 @@ class RemoteHost:
         self.keyfile_path = keyfile_path
         self.login_name = login_name
 
-    def _gen_ssh_prefix(self) -> list[str]:
+    def _gen_ssh_prefix(self) -> List[str]:
         return [
             "ssh",
             "-o",
@@ -110,13 +108,13 @@ class RemoteHost:
         ]
 
     @staticmethod
-    def _split_cmd(args: Union[str, list[str]]) -> list[str]:
+    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:
+    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:
+    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")
@@ -159,7 +157,7 @@ class RemoteHost:
     def using_docker(self) -> bool:
         return self.container_id is not None
 
-    def run_cmd(self, args: Union[str, list[str]]) -> 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
@@ -180,7 +178,7 @@ class RemoteHost:
         if rc != 0:
             raise subprocess.CalledProcessError(rc, docker_cmd)
 
-    def check_output(self, args: Union[str, list[str]]) -> str:
+    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
@@ -232,7 +230,7 @@ class RemoteHost:
             )
         self.download_file(remote_file, local_file)
 
-    def list_dir(self, path: str) -> list[str]:
+    def list_dir(self, path: str) -> List[str]:
         return self.check_output(["ls", "-1", path]).split("\n")
 
 
@@ -360,7 +358,7 @@ def checkout_repo(
     branch: str = "main",
     url: str,
     git_clone_flags: str,
-    mapping: dict[str, tuple[str, str]],
+    mapping: Dict[str, Tuple[str, str]],
 ) -> Optional[str]:
     for prefix in mapping:
         if not branch.startswith(prefix):
@@ -683,7 +681,7 @@ def build_domains(
     branch: str = "main",
     use_conda: bool = True,
     git_clone_flags: str = "",
-) -> tuple[str, str, str, str]:
+) -> Tuple[str, str, str, str]:
     vision_wheel_name = build_torchvision(
         host, branch=branch, use_conda=use_conda, git_clone_flags=git_clone_flags
     )
@@ -710,7 +708,7 @@ def start_build(
     pytorch_build_number: Optional[str] = None,
     shallow_clone: bool = True,
     enable_mkldnn: bool = False,
-) -> tuple[str, str, str, str, str]:
+) -> 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")
@@ -934,9 +932,9 @@ def parse_arguments():
     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(
+        "--os", type=str, choices=list(os_amis.keys()), default="ubuntu20_04"
+    )
     parser.add_argument(
         "--python-version",
         type=str,
@@ -966,13 +964,7 @@ def parse_arguments():
 
 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
-    )
+    ami = os_amis[args.os]
     keyfile_path, key_name = compute_keyfile_path(args.key_name)
 
     if args.list_instances:

.ci/docker/aotriton_version.txt

@@ -0,0 +1,5 @@
+0.8b
+manylinux_2_28
+rocm6.2
+6f8cbcac8a92775291bb1ba8f514d4beb350baf4
+e938def5d32869fe2e00aec0300f354c9f157867bebdf2e104d732b94cb238d8

.ci/docker/build.sh

@@ -86,10 +86,6 @@ CMAKE_VERSION=3.18.5
 
 _UCX_COMMIT=7bb2722ff2187a0cad557ae4a6afa090569f83fb
 _UCC_COMMIT=20eae37090a4ce1b32bcce6144ccad0b49943e0b
-if [[ "$image" == *rocm* ]]; then
-  _UCX_COMMIT=cc312eaa4655c0cc5c2bcd796db938f90563bcf6
-  _UCC_COMMIT=0c0fc21559835044ab107199e334f7157d6a0d3d
-fi
 
 # It's annoying to rename jobs every time you want to rewrite a
 # configuration, so we hardcode everything here rather than do it
@@ -109,6 +105,20 @@ case "$image" in
     CONDA_CMAKE=yes
     TRITON=yes
     ;;
+  pytorch-linux-focal-cuda12.1-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.1.1
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
+    TRITON=yes
+    ;;
   pytorch-linux-focal-cuda12.4-cudnn9-py3-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.4.1
     CUDNN_VERSION=9
@@ -124,6 +134,36 @@ case "$image" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
+  pytorch-linux-focal-cuda12.1-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.1.1
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-focal-cuda12.1-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.1.1
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=9
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
   pytorch-linux-focal-cuda12.4-cudnn9-py3.12-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.4.1
     CUDNN_VERSION=9
@@ -168,6 +208,48 @@ case "$image" in
     CONDA_CMAKE=yes
     TRITON=yes
     ;;
+  pytorch-linux-focal-cuda12.4-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.4.1
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-focal-cuda12.1-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.1.1
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-focal-cuda12.4-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.4.1
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
+    TRITON=yes
+    ;;
   pytorch-linux-focal-py3-clang10-onnx)
     ANACONDA_PYTHON_VERSION=3.9
     CLANG_VERSION=10
@@ -210,7 +292,18 @@ case "$image" in
     ;;
   pytorch-linux-focal-rocm-n-1-py3)
     ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
+    GCC_VERSION=9
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    ROCM_VERSION=6.1
+    NINJA_VERSION=1.9.0
+    CONDA_CMAKE=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-focal-rocm-n-py3)
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
     PROTOBUF=yes
     DB=yes
     VISION=yes
@@ -218,25 +311,6 @@ case "$image" in
     NINJA_VERSION=1.9.0
     CONDA_CMAKE=yes
     TRITON=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    INDUCTOR_BENCHMARKS=yes
-    ;;
-  pytorch-linux-focal-rocm-n-py3)
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
-    PROTOBUF=yes
-    DB=yes
-    VISION=yes
-    ROCM_VERSION=6.3
-    NINJA_VERSION=1.9.0
-    CONDA_CMAKE=yes
-    TRITON=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    INDUCTOR_BENCHMARKS=yes
     ;;
   pytorch-linux-jammy-xpu-2024.0-py3)
     ANACONDA_PYTHON_VERSION=3.9
@@ -451,7 +525,7 @@ docker build \
        --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
-       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH:-gfx90a;gfx942}" \
+       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH:-gfx90a}" \
        --build-arg "IMAGE_NAME=${IMAGE_NAME}" \
        --build-arg "UCX_COMMIT=${UCX_COMMIT}" \
        --build-arg "UCC_COMMIT=${UCC_COMMIT}" \

.ci/docker/centos-rocm/Dockerfile

@@ -113,6 +113,13 @@ COPY triton_version.txt triton_version.txt
 RUN if [ -n "${TRITON}" ]; then bash ./install_triton.sh; fi
 RUN rm install_triton.sh common_utils.sh triton.txt triton_version.txt
 
+# Install AOTriton (Early fail)
+COPY ./aotriton_version.txt aotriton_version.txt
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./common/install_aotriton.sh install_aotriton.sh
+RUN ["/bin/bash", "-c", "./install_aotriton.sh /opt/rocm && rm -rf install_aotriton.sh aotriton_version.txt common_utils.sh"]
+ENV AOTRITON_INSTALLED_PREFIX /opt/rocm/aotriton
+
 # Install ccache/sccache (do this last, so we get priority in PATH)
 COPY ./common/install_cache.sh install_cache.sh
 ENV PATH /opt/cache/bin:$PATH

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-5e4d6b6380d575e48e37e9d987fded4ec588e7bc
+6f638937d64e3396793956d75ee3e14802022745

.ci/docker/ci_commit_pins/nccl-cu11.txt

@@ -1 +0,0 @@
-v2.21.5-1

.ci/docker/ci_commit_pins/nccl-cu12.txt

@@ -1 +0,0 @@
-v2.25.1-1

.ci/docker/ci_commit_pins/timm.txt

@@ -1 +1 @@
-5d535d7a2d4b435b1b5c1177fd8f04a12b942b9a
+ac3470188b914c5d7a5058a7e28b9eb685a62427

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-4b3bb1f8da0ded6ccd572dd1358ef45af5a1befe
+0d4682f073ded4d1a8260dd4208a43d735ae3a2b

.ci/docker/common/install_acl.sh

@@ -1,7 +1,7 @@
 set -euo pipefail
 
 readonly version=v24.04
-readonly src_host=https://github.com/ARM-software
+readonly src_host=https://review.mlplatform.org/ml
 readonly src_repo=ComputeLibrary
 
 # Clone ACL

.ci/docker/common/install_aotriton.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+
+set -ex
+
+source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
+
+TARBALL='aotriton.tar.gz'
+# This read command alwasy returns with exit code 1
+read -d "\n" VER MANYLINUX ROCMBASE PINNED_COMMIT SHA256 < aotriton_version.txt || true
+ARCH=$(uname -m)
+AOTRITON_INSTALL_PREFIX="$1"
+AOTRITON_URL="https://github.com/ROCm/aotriton/releases/download/${VER}/aotriton-${VER}-${MANYLINUX}_${ARCH}-${ROCMBASE}-shared.tar.gz"
+
+cd "${AOTRITON_INSTALL_PREFIX}"
+# Must use -L to follow redirects
+curl -L --retry 3 -o "${TARBALL}" "${AOTRITON_URL}"
+ACTUAL_SHA256=$(sha256sum "${TARBALL}" | cut -d " " -f 1)
+if [ "${SHA256}" != "${ACTUAL_SHA256}" ]; then
+  echo -n "Error: The SHA256 of downloaded tarball is ${ACTUAL_SHA256},"
+  echo " which does not match the expected value ${SHA256}."
+  exit
+fi
+tar xf "${TARBALL}" && rm -rf "${TARBALL}"

.ci/docker/common/install_base.sh

@@ -32,12 +32,8 @@ install_ubuntu() {
 
   # HACK: UCC testing relies on libnccl library from NVIDIA repo, and version 2.16 crashes
   # See https://github.com/pytorch/pytorch/pull/105260#issuecomment-1673399729
-  # TODO: Eliminate this hack, we should not relay on apt-get installation
-  # See https://github.com/pytorch/pytorch/issues/144768
   if [[ "$UBUNTU_VERSION" == "20.04"* && "$CUDA_VERSION" == "11.8"* ]]; then
     maybe_libnccl_dev="libnccl2=2.15.5-1+cuda11.8 libnccl-dev=2.15.5-1+cuda11.8 --allow-downgrades --allow-change-held-packages"
-  elif [[ "$UBUNTU_VERSION" == "20.04"* && "$CUDA_VERSION" == "12.4"* ]]; then
-    maybe_libnccl_dev="libnccl2=2.25.1-1+cuda12.4 libnccl-dev=2.25.1-1+cuda12.4 --allow-downgrades --allow-change-held-packages"
   else
     maybe_libnccl_dev=""
   fi

.ci/docker/common/install_cache.sh

@@ -9,7 +9,7 @@ install_ubuntu() {
   # Instead use lib and headers from OpenSSL1.1 installed in `install_openssl.sh``
   apt-get install -y cargo
   echo "Checking out sccache repo"
-  git clone https://github.com/mozilla/sccache -b v0.9.1
+  git clone https://github.com/mozilla/sccache -b v0.9.0
   cd sccache
   echo "Building sccache"
   cargo build --release
@@ -36,7 +36,11 @@ sed -e 's|PATH="\(.*\)"|PATH="/opt/cache/bin:\1"|g' -i /etc/environment
 export PATH="/opt/cache/bin:$PATH"
 
 # Setup compiler cache
-install_ubuntu
+if [ -n "$ROCM_VERSION" ]; then
+  curl --retry 3 http://repo.radeon.com/misc/.sccache_amd/sccache -o /opt/cache/bin/sccache
+else
+  install_ubuntu
+fi
 chmod a+x /opt/cache/bin/sccache
 
 function write_sccache_stub() {

.ci/docker/common/install_cuda.sh

@@ -2,7 +2,7 @@
 
 set -ex
 
-NCCL_VERSION=v2.25.1-1
+NCCL_VERSION=v2.21.5-1
 CUDNN_VERSION=9.5.1.17
 
 function install_cusparselt_040 {
@@ -16,6 +16,17 @@ function install_cusparselt_040 {
     rm -rf tmp_cusparselt
 }
 
+function install_cusparselt_052 {
+    # cuSparseLt license: https://docs.nvidia.com/cuda/cusparselt/license.html
+    mkdir tmp_cusparselt && pushd tmp_cusparselt
+    wget -q https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-x86_64/libcusparse_lt-linux-x86_64-0.5.2.1-archive.tar.xz
+    tar xf libcusparse_lt-linux-x86_64-0.5.2.1-archive.tar.xz
+    cp -a libcusparse_lt-linux-x86_64-0.5.2.1-archive/include/* /usr/local/cuda/include/
+    cp -a libcusparse_lt-linux-x86_64-0.5.2.1-archive/lib/* /usr/local/cuda/lib64/
+    popd
+    rm -rf tmp_cusparselt
+}
+
 function install_cusparselt_062 {
     # cuSparseLt license: https://docs.nvidia.com/cuda/cusparselt/license.html
     mkdir tmp_cusparselt && pushd tmp_cusparselt
@@ -40,7 +51,6 @@ function install_cusparselt_063 {
 
 function install_118 {
     CUDNN_VERSION=9.1.0.70
-    NCCL_VERSION=v2.21.5-1
     echo "Installing CUDA 11.8 and cuDNN ${CUDNN_VERSION} and NCCL ${NCCL_VERSION} and cuSparseLt-0.4.0"
     rm -rf /usr/local/cuda-11.8 /usr/local/cuda
     # install CUDA 11.8.0 in the same container
@@ -73,6 +83,39 @@ function install_118 {
     ldconfig
 }
 
+function install_121 {
+    echo "Installing CUDA 12.1 and cuDNN ${CUDNN_VERSION} and NCCL ${NCCL_VERSION} and cuSparseLt-0.5.2"
+    rm -rf /usr/local/cuda-12.1 /usr/local/cuda
+    # install CUDA 12.1.0 in the same container
+    wget -q https://developer.download.nvidia.com/compute/cuda/12.1.1/local_installers/cuda_12.1.1_530.30.02_linux.run
+    chmod +x cuda_12.1.1_530.30.02_linux.run
+    ./cuda_12.1.1_530.30.02_linux.run --toolkit --silent
+    rm -f cuda_12.1.1_530.30.02_linux.run
+    rm -f /usr/local/cuda && ln -s /usr/local/cuda-12.1 /usr/local/cuda
+
+    # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
+    mkdir tmp_cudnn && cd tmp_cudnn
+    wget -q https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive.tar.xz -O cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive.tar.xz
+    tar xf cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive.tar.xz
+    cp -a cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive/include/* /usr/local/cuda/include/
+    cp -a cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive/lib/* /usr/local/cuda/lib64/
+    cd ..
+    rm -rf tmp_cudnn
+
+    # NCCL license: https://docs.nvidia.com/deeplearning/nccl/#licenses
+    # Follow build: https://github.com/NVIDIA/nccl/tree/master?tab=readme-ov-file#build
+    git clone -b $NCCL_VERSION --depth 1 https://github.com/NVIDIA/nccl.git
+    cd nccl && make -j src.build
+    cp -a build/include/* /usr/local/cuda/include/
+    cp -a build/lib/* /usr/local/cuda/lib64/
+    cd ..
+    rm -rf nccl
+
+    install_cusparselt_052
+
+    ldconfig
+}
+
 function install_124 {
   CUDNN_VERSION=9.1.0.70
   echo "Installing CUDA 12.4.1 and cuDNN ${CUDNN_VERSION} and NCCL ${NCCL_VERSION} and cuSparseLt-0.6.2"
@@ -171,6 +214,37 @@ function prune_118 {
     rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2022.3.0 $CUDA_BASE/nsight-systems-2022.4.2/
 }
 
+function prune_121 {
+  echo "Pruning CUDA 12.1"
+  #####################################################################################
+  # CUDA 12.1 prune static libs
+  #####################################################################################
+    export NVPRUNE="/usr/local/cuda-12.1/bin/nvprune"
+    export CUDA_LIB_DIR="/usr/local/cuda-12.1/lib64"
+
+    export GENCODE="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+    export GENCODE_CUDNN="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+
+    if [[ -n "$OVERRIDE_GENCODE" ]]; then
+        export GENCODE=$OVERRIDE_GENCODE
+    fi
+
+    # all CUDA libs except CuDNN and CuBLAS
+    ls $CUDA_LIB_DIR/ | grep "\.a" | grep -v "culibos" | grep -v "cudart" | grep -v "cudnn" | grep -v "cublas" | grep -v "metis"  \
+      | xargs -I {} bash -c \
+                "echo {} && $NVPRUNE $GENCODE $CUDA_LIB_DIR/{} -o $CUDA_LIB_DIR/{}"
+
+    # prune CuDNN and CuBLAS
+    $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublas_static.a -o $CUDA_LIB_DIR/libcublas_static.a
+    $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublasLt_static.a -o $CUDA_LIB_DIR/libcublasLt_static.a
+
+    #####################################################################################
+    # CUDA 12.1 prune visual tools
+    #####################################################################################
+    export CUDA_BASE="/usr/local/cuda-12.1/"
+    rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2023.1.0 $CUDA_BASE/nsight-systems-2023.1.2/
+}
+
 function prune_124 {
   echo "Pruning CUDA 12.4"
   #####################################################################################
@@ -239,52 +313,18 @@ function prune_126 {
   rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2024.3.2 $CUDA_BASE/nsight-systems-2024.5.1/
 }
 
-function install_128 {
-  CUDNN_VERSION=9.7.1.26
-  echo "Installing CUDA 12.8.0 and cuDNN ${CUDNN_VERSION} and NCCL ${NCCL_VERSION} and cuSparseLt-0.6.3"
-  rm -rf /usr/local/cuda-12.8 /usr/local/cuda
-  # install CUDA 12.8.0 in the same container
-  wget -q https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_570.86.10_linux.run
-  chmod +x cuda_12.8.0_570.86.10_linux.run
-  ./cuda_12.8.0_570.86.10_linux.run --toolkit --silent
-  rm -f cuda_12.8.0_570.86.10_linux.run
-  rm -f /usr/local/cuda && ln -s /usr/local/cuda-12.8 /usr/local/cuda
-
-  # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
-  mkdir tmp_cudnn && cd tmp_cudnn
-  wget -q https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive.tar.xz -O cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive.tar.xz
-  tar xf cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive.tar.xz
-  cp -a cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive/include/* /usr/local/cuda/include/
-  cp -a cudnn-linux-x86_64-${CUDNN_VERSION}_cuda12-archive/lib/* /usr/local/cuda/lib64/
-  cd ..
-  rm -rf tmp_cudnn
-
-  # NCCL license: https://docs.nvidia.com/deeplearning/nccl/#licenses
-  # Follow build: https://github.com/NVIDIA/nccl/tree/master?tab=readme-ov-file#build
-  git clone -b $NCCL_VERSION --depth 1 https://github.com/NVIDIA/nccl.git
-  cd nccl && make -j src.build
-  cp -a build/include/* /usr/local/cuda/include/
-  cp -a build/lib/* /usr/local/cuda/lib64/
-  cd ..
-  rm -rf nccl
-
-  install_cusparselt_063
-
-  ldconfig
-}
-
 # idiomatic parameter and option handling in sh
 while test $# -gt 0
 do
     case "$1" in
     11.8) install_118; prune_118
         ;;
+    12.1) install_121; prune_121
+        ;;
     12.4) install_124; prune_124
         ;;
     12.6) install_126; prune_126
         ;;
-    12.8) install_128;
-        ;;
     *) echo "bad argument $1"; exit 1
         ;;
     esac

.ci/docker/common/install_cuda_aarch64.sh

@@ -57,7 +57,7 @@ function install_124 {
   cd ..
   rm -rf nccl
 
-  install_cusparselt_063
+  install_cusparselt_062
 
   ldconfig
 }
@@ -160,40 +160,6 @@ function prune_126 {
   rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2024.3.2 $CUDA_BASE/nsight-systems-2024.5.1/
 }
 
-function install_128 {
-  CUDNN_VERSION=9.7.1.26
-  echo "Installing CUDA 12.8.0 and cuDNN ${CUDNN_VERSION} and NCCL ${NCCL_VERSION} and cuSparseLt-0.6.3"
-  rm -rf /usr/local/cuda-12.8 /usr/local/cuda
-  # install CUDA 12.8.0 in the same container
-  wget -q https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_570.86.10_linux_sbsa.run
-  chmod +x cuda_12.8.0_570.86.10_linux_sbsa.run
-  ./cuda_12.8.0_570.86.10_linux_sbsa.run --toolkit --silent
-  rm -f cuda_12.8.0_570.86.10_linux_sbsa.run
-  rm -f /usr/local/cuda && ln -s /usr/local/cuda-12.8 /usr/local/cuda
-
-  # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
-  mkdir tmp_cudnn && cd tmp_cudnn
-  wget -q https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-sbsa/cudnn-linux-sbsa-${CUDNN_VERSION}_cuda12-archive.tar.xz -O cudnn-linux-sbsa-${CUDNN_VERSION}_cuda12-archive.tar.xz
-  tar xf cudnn-linux-sbsa-${CUDNN_VERSION}_cuda12-archive.tar.xz
-  cp -a cudnn-linux-sbsa-${CUDNN_VERSION}_cuda12-archive/include/* /usr/local/cuda/include/
-  cp -a cudnn-linux-sbsa-${CUDNN_VERSION}_cuda12-archive/lib/* /usr/local/cuda/lib64/
-  cd ..
-  rm -rf tmp_cudnn
-
-  # NCCL license: https://docs.nvidia.com/deeplearning/nccl/#licenses
-  # Follow build: https://github.com/NVIDIA/nccl/tree/master?tab=readme-ov-file#build
-  git clone -b ${NCCL_VERSION} --depth 1 https://github.com/NVIDIA/nccl.git
-  cd nccl && make -j src.build
-  cp -a build/include/* /usr/local/cuda/include/
-  cp -a build/lib/* /usr/local/cuda/lib64/
-  cd ..
-  rm -rf nccl
-
-  install_cusparselt_063
-
-  ldconfig
-}
-
 # idiomatic parameter and option handling in sh
 while test $# -gt 0
 do
@@ -202,8 +168,6 @@ do
         ;;
     12.6) install_126; prune_126
         ;;
-    12.8) install_128;
-        ;;
     *) echo "bad argument $1"; exit 1
         ;;
     esac

.ci/docker/common/install_cudnn.sh

@@ -4,9 +4,7 @@ if [[ -n "${CUDNN_VERSION}" ]]; then
     # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
     mkdir tmp_cudnn
     pushd tmp_cudnn
-    if [[ ${CUDA_VERSION:0:4} == "12.8" ]]; then
-        CUDNN_NAME="cudnn-linux-x86_64-9.7.1.26_cuda12-archive"
-    elif [[ ${CUDA_VERSION:0:4} == "12.6" ]]; then
+    if [[ ${CUDA_VERSION:0:4} == "12.6" ]]; then
         CUDNN_NAME="cudnn-linux-x86_64-9.5.1.17_cuda12-archive"
     elif [[ ${CUDA_VERSION:0:2} == "12" ]]; then
         CUDNN_NAME="cudnn-linux-x86_64-9.1.0.70_cuda12-archive"

.ci/docker/common/install_cusparselt.sh

@@ -5,15 +5,7 @@ set -ex
 # cuSPARSELt license: https://docs.nvidia.com/cuda/cusparselt/license.html
 mkdir tmp_cusparselt && cd tmp_cusparselt
 
-if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-8]$ ]]; then
-    arch_path='sbsa'
-    export TARGETARCH=${TARGETARCH:-$(uname -m)}
-    if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
-        arch_path='x86_64'
-    fi
-    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.6.3.2-archive"
-    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
-elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
+if [[ ${CUDA_VERSION:0:4} =~ ^12\.[2-6]$ ]]; then
     arch_path='sbsa'
     export TARGETARCH=${TARGETARCH:-$(uname -m)}
     if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
@@ -21,11 +13,17 @@ elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
     fi
     CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.6.2.3-archive"
     curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
+elif [[ ${CUDA_VERSION:0:4} == "12.1" ]]; then
+    arch_path='sbsa'
+    export TARGETARCH=${TARGETARCH:-$(uname -m)}
+    if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
+        arch_path='x86_64'
+    fi
+    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.5.2.1-archive"
+    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
 elif [[ ${CUDA_VERSION:0:4} == "11.8" ]]; then
     CUSPARSELT_NAME="libcusparse_lt-linux-x86_64-0.4.0.7-archive"
     curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-x86_64/${CUSPARSELT_NAME}.tar.xz
-else
-    echo "Not sure which libcusparselt version to install for this ${CUDA_VERSION}"
 fi
 
 tar xf ${CUSPARSELT_NAME}.tar.xz

.ci/docker/common/install_executorch.sh

@@ -37,12 +37,7 @@ install_conda_dependencies() {
 
 install_pip_dependencies() {
   pushd executorch
-  as_jenkins bash install_executorch.sh
-
-  # A workaround, ExecuTorch has moved to numpy 2.0 which is not compatible with the current
-  # numba and scipy version used in PyTorch CI
-  conda_run pip uninstall -y numba scipy
-
+  as_jenkins bash install_requirements.sh --pybind xnnpack
   popd
 }
 

.ci/docker/common/install_onnx.sh

@@ -31,15 +31,15 @@ pip_install \
 pip_install coloredlogs packaging
 
 pip_install onnxruntime==1.18.1
-pip_install onnx==1.17.0
-pip_install onnxscript==0.1.0 --no-deps
+pip_install onnx==1.16.2
+pip_install onnxscript==0.1.0.dev20241124 --no-deps
 # required by onnxscript
 pip_install ml_dtypes
 
 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/
 IMPORT_SCRIPT_FILENAME="/tmp/onnx_import_script.py"
-as_jenkins echo 'import transformers; transformers.GPTJForCausalLM.from_pretrained("hf-internal-testing/tiny-random-gptj");' > "${IMPORT_SCRIPT_FILENAME}"
+as_jenkins echo 'import transformers; transformers.AutoModel.from_pretrained("sshleifer/tiny-gpt2"); transformers.AutoTokenizer.from_pretrained("sshleifer/tiny-gpt2"); transformers.AutoModelForSpeechSeq2Seq.from_pretrained("openai/whisper-large-v3");' > "${IMPORT_SCRIPT_FILENAME}"
 
 # Need a PyTorch version for transformers to work
 pip_install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu

.ci/docker/common/install_rocm.sh

@@ -62,22 +62,6 @@ install_ubuntu() {
         sqlite3 $kdb "PRAGMA journal_mode=off; PRAGMA VACUUM;"
     done
 
-    # ROCm 6.3 had a regression where initializing static code objects had significant overhead
-    if [[ $(ver $ROCM_VERSION) -eq $(ver 6.3) ]]; then
-        # clr build needs CppHeaderParser but can only find it using conda's python
-        /opt/conda/bin/python -m pip install CppHeaderParser
-        git clone https://github.com/ROCm/HIP -b rocm-6.3.x
-        HIP_COMMON_DIR=$(readlink -f HIP)
-        git clone https://github.com/jeffdaily/clr -b release/rocm-rel-6.3-statco-hotfix
-        mkdir -p clr/build
-        pushd clr/build
-        cmake .. -DCLR_BUILD_HIP=ON -DHIP_COMMON_DIR=$HIP_COMMON_DIR
-        make -j
-        cp hipamd/lib/libamdhip64.so.6.3.* /opt/rocm/lib/libamdhip64.so.6.3.*
-        popd
-        rm -rf HIP clr
-    fi
-
     # Cleanup
     apt-get autoclean && apt-get clean
     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*

.ci/docker/common/install_ucc.sh

@@ -8,12 +8,6 @@ else
   with_cuda=no
 fi
 
-if [[ -d "/opt/rocm" ]]; then
-  with_rocm=/opt/rocm
-else
-  with_rocm=no
-fi
-
 function install_ucx() {
   set -ex
   git clone --recursive https://github.com/openucx/ucx.git
@@ -25,7 +19,6 @@ function install_ucx() {
   ./configure --prefix=$UCX_HOME      \
       --enable-mt                     \
       --with-cuda=$with_cuda          \
-      --with-rocm=$with_rocm          \
       --enable-profiling              \
       --enable-stats
   time make -j
@@ -43,29 +36,12 @@ function install_ucc() {
   git submodule update --init --recursive
 
   ./autogen.sh
-
   # We only run distributed tests on Tesla M60 and A10G
   NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_86,code=compute_86"
-
-  if [[ -n "$ROCM_VERSION" ]]; then
-    if [[ -n "$PYTORCH_ROCM_ARCH" ]]; then
-      amdgpu_targets=`echo $PYTORCH_ROCM_ARCH | sed 's/;/ /g'`
-    else
-      amdgpu_targets=`rocm_agent_enumerator | grep -v gfx000 | sort -u | xargs`
-    fi
-    for arch in $amdgpu_targets; do
-      HIP_OFFLOAD="$HIP_OFFLOAD --offload-arch=$arch"
-    done
-  else
-    HIP_OFFLOAD="all-arch-no-native"
-  fi
-
   ./configure --prefix=$UCC_HOME          \
     --with-ucx=$UCX_HOME                  \
     --with-cuda=$with_cuda                \
-    --with-nvcc-gencode="${NVCC_GENCODE}" \
-    --with-rocm=$with_rocm                \
-    --with-rocm-arch="${HIP_OFFLOAD}"
+    --with-nvcc-gencode="${NVCC_GENCODE}"
   time make -j
   sudo make install
 

.ci/docker/libtorch/Dockerfile

@@ -56,6 +56,11 @@ RUN bash ./install_cuda.sh 11.8
 RUN bash ./install_magma.sh 11.8
 RUN ln -sf /usr/local/cuda-11.8 /usr/local/cuda
 
+FROM cuda as cuda12.1
+RUN bash ./install_cuda.sh 12.1
+RUN bash ./install_magma.sh 12.1
+RUN ln -sf /usr/local/cuda-12.1 /usr/local/cuda
+
 FROM cuda as cuda12.4
 RUN bash ./install_cuda.sh 12.4
 RUN bash ./install_magma.sh 12.4
@@ -66,11 +71,6 @@ RUN bash ./install_cuda.sh 12.6
 RUN bash ./install_magma.sh 12.6
 RUN ln -sf /usr/local/cuda-12.6 /usr/local/cuda
 
-FROM cuda as cuda12.8
-RUN bash ./install_cuda.sh 12.8
-RUN bash ./install_magma.sh 12.8
-RUN ln -sf /usr/local/cuda-12.8 /usr/local/cuda
-
 FROM cpu as rocm
 ARG PYTORCH_ROCM_ARCH
 ENV PYTORCH_ROCM_ARCH ${PYTORCH_ROCM_ARCH}
@@ -92,6 +92,13 @@ RUN apt-get update -y && \
 RUN bash ./install_rocm_drm.sh && rm install_rocm_drm.sh
 RUN bash ./install_rocm_magma.sh && rm install_rocm_magma.sh
 
+# Install AOTriton
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./aotriton_version.txt aotriton_version.txt
+COPY ./common/install_aotriton.sh install_aotriton.sh
+RUN bash ./install_aotriton.sh /opt/rocm && rm install_aotriton.sh aotriton_version.txt
+ENV AOTRITON_INSTALLED_PREFIX /opt/rocm/aotriton
+
 FROM ${BASE_TARGET} as final
 COPY --from=openssl            /opt/openssl           /opt/openssl
 # Install patchelf

.ci/docker/manywheel/Dockerfile

@@ -198,3 +198,10 @@ ADD ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh && rm install_rocm_magma.sh
 ADD ./common/install_miopen.sh install_miopen.sh
 RUN bash ./install_miopen.sh ${ROCM_VERSION} && rm install_miopen.sh
+
+# Install AOTriton
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./aotriton_version.txt aotriton_version.txt
+COPY ./common/install_aotriton.sh install_aotriton.sh
+RUN bash ./install_aotriton.sh /opt/rocm && rm install_aotriton.sh aotriton_version.txt
+ENV AOTRITON_INSTALLED_PREFIX /opt/rocm/aotriton

.ci/docker/requirements-ci.txt

@@ -304,7 +304,7 @@ pytest-cpp==2.3.0
 #Pinned versions: 2.3.0
 #test that import:
 
-z3-solver==4.12.6.0
+z3-solver==4.12.2.0
 #Description: The Z3 Theorem Prover Project
 #Pinned versions:
 #test that import:
@@ -329,7 +329,7 @@ lxml==5.3.0
 
 PyGithub==2.3.0
 
-sympy==1.13.3
+sympy==1.13.1 ; python_version >= "3.9"
 #Description: Required by coremltools, also pinned in .github/requirements/pip-requirements-macOS.txt
 #Pinned versions:
 #test that import:
@@ -339,7 +339,7 @@ onnx==1.17.0
 #Pinned versions:
 #test that import:
 
-onnxscript==0.1.0
+onnxscript==0.1.0.dev20240817
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -362,7 +362,6 @@ pwlf==2.2.1 ; python_version >= "3.8"
 # To build PyTorch itself
 astunparse
 PyYAML
-pyzstd
 setuptools
 
 ninja==1.11.1 ; platform_machine == "aarch64"
@@ -372,8 +371,3 @@ pulp==2.9.0 ; python_version >= "3.8"
 #Description: required for testing ilp formulaiton under torch/distributed/_tools
 #Pinned versions: 2.9.0
 #test that import: test_sac_ilp.py
-
-dataclasses_json==0.6.7
-#Description: required for data pipeline and scripts under tools/stats
-#Pinned versions: 0.6.7
-#test that import:

.ci/docker/ubuntu-rocm/Dockerfile

@@ -14,20 +14,21 @@ ENV PYTORCH_ROCM_ARCH ${PYTORCH_ROCM_ARCH}
 COPY ./common/install_base.sh install_base.sh
 RUN bash ./install_base.sh && rm install_base.sh
 
+# Install clang
+ARG LLVMDEV
+ARG CLANG_VERSION
+COPY ./common/install_clang.sh install_clang.sh
+RUN bash ./install_clang.sh && rm install_clang.sh
+
 # Install user
 COPY ./common/install_user.sh install_user.sh
 RUN bash ./install_user.sh && rm install_user.sh
 
-# Install katex
-ARG KATEX
-COPY ./common/install_docs_reqs.sh install_docs_reqs.sh
-RUN bash ./install_docs_reqs.sh && rm install_docs_reqs.sh
-
 # Install conda and other packages (e.g., numpy, pytest)
 ARG ANACONDA_PYTHON_VERSION
+ARG CONDA_CMAKE
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
-ARG CONDA_CMAKE
 COPY requirements-ci.txt /opt/conda/requirements-ci.txt
 COPY ./common/install_conda.sh install_conda.sh
 COPY ./common/common_utils.sh common_utils.sh
@@ -38,11 +39,6 @@ ARG GCC_VERSION
 COPY ./common/install_gcc.sh install_gcc.sh
 RUN bash ./install_gcc.sh && rm install_gcc.sh
 
-# Install clang
-ARG CLANG_VERSION
-COPY ./common/install_clang.sh install_clang.sh
-RUN bash ./install_clang.sh && rm install_clang.sh
-
 # (optional) Install protobuf for ONNX
 ARG PROTOBUF
 COPY ./common/install_protobuf.sh install_protobuf.sh
@@ -89,32 +85,6 @@ COPY ./common/install_amdsmi.sh install_amdsmi.sh
 RUN bash ./install_amdsmi.sh
 RUN rm install_amdsmi.sh
 
-# (optional) Install UCC
-ARG UCX_COMMIT
-ARG UCC_COMMIT
-ENV UCX_COMMIT $UCX_COMMIT
-ENV UCC_COMMIT $UCC_COMMIT
-ENV UCX_HOME /usr
-ENV UCC_HOME /usr
-ADD ./common/install_ucc.sh install_ucc.sh
-RUN if [ -n "${UCX_COMMIT}" ] && [ -n "${UCC_COMMIT}" ]; then bash ./install_ucc.sh; fi
-RUN rm install_ucc.sh
-
-COPY ./common/install_openssl.sh install_openssl.sh
-ENV OPENSSL_ROOT_DIR /opt/openssl
-RUN bash ./install_openssl.sh
-ENV OPENSSL_DIR /opt/openssl
-
-ARG INDUCTOR_BENCHMARKS
-ARG ANACONDA_PYTHON_VERSION
-ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
-COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
-COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface.txt huggingface.txt
-COPY ci_commit_pins/timm.txt timm.txt
-RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
-
 # (optional) Install non-default CMake version
 ARG CMAKE_VERSION
 COPY ./common/install_cmake.sh install_cmake.sh
@@ -137,17 +107,18 @@ COPY triton_version.txt triton_version.txt
 RUN if [ -n "${TRITON}" ]; then bash ./install_triton.sh; fi
 RUN rm install_triton.sh common_utils.sh triton.txt triton_version.txt
 
+# Install AOTriton
+COPY ./aotriton_version.txt aotriton_version.txt
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./common/install_aotriton.sh install_aotriton.sh
+RUN ["/bin/bash", "-c", "./install_aotriton.sh /opt/rocm && rm -rf install_aotriton.sh aotriton_version.txt common_utils.sh"]
+ENV AOTRITON_INSTALLED_PREFIX /opt/rocm/aotriton
 
 # Install ccache/sccache (do this last, so we get priority in PATH)
 COPY ./common/install_cache.sh install_cache.sh
 ENV PATH /opt/cache/bin:$PATH
 RUN bash ./install_cache.sh && rm install_cache.sh
 
-# Install Open MPI for ROCm
-COPY ./common/install_openmpi.sh install_openmpi.sh
-RUN if [ -n "${CUDA_VERSION}" ]; then bash install_openmpi.sh; fi
-RUN rm install_openmpi.sh
-
 # Include BUILD_ENVIRONMENT environment variable in image
 ARG BUILD_ENVIRONMENT
 ENV BUILD_ENVIRONMENT ${BUILD_ENVIRONMENT}

.ci/magma/Makefile

@@ -16,9 +16,9 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma/build_magma.sh
 
 .PHONY: all
-all: magma-cuda128
 all: magma-cuda126
 all: magma-cuda124
+all: magma-cuda121
 all: magma-cuda118
 
 .PHONY:
@@ -26,12 +26,6 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
-.PHONY: magma-cuda128
-magma-cuda128: DESIRED_CUDA := 12.8
-magma-cuda128: CUDA_ARCH_LIST += -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
-magma-cuda128:
-	$(DOCKER_RUN)
-
 .PHONY: magma-cuda126
 magma-cuda126: DESIRED_CUDA := 12.6
 magma-cuda126:
@@ -42,6 +36,11 @@ magma-cuda124: DESIRED_CUDA := 12.4
 magma-cuda124:
 	$(DOCKER_RUN)
 
+.PHONY: magma-cuda121
+magma-cuda121: DESIRED_CUDA := 12.1
+magma-cuda121:
+	$(DOCKER_RUN)
+
 .PHONY: magma-cuda118
 magma-cuda118: DESIRED_CUDA := 11.8
 magma-cuda118: CUDA_ARCH_LIST += -gencode arch=compute_37,code=sm_37

.ci/manywheel/build_cuda.sh

@@ -14,7 +14,6 @@ export USE_CUDA_STATIC_LINK=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
 export USE_CUPTI_SO=0
 export USE_CUSPARSELT=${USE_CUSPARSELT:-1} # Enable if not disabled by libtorch build
-export USE_CUFILE=${USE_CUFILE:-1}
 
 # Keep an array of cmake variables to add to
 if [[ -z "$CMAKE_ARGS" ]]; then
@@ -53,15 +52,23 @@ cuda_version_nodot=$(echo $CUDA_VERSION | tr -d '.')
 
 TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6"
 case ${CUDA_VERSION} in
-    12.8)
-        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0;10.0;12.0+PTX" #Ripping out 5.0 and 6.0 due to ld error
-        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
-        ;;
     12.6)
-        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0+PTX"
+        if [[ "$GPU_ARCH_TYPE" = "cuda-aarch64" ]]; then
+            TORCH_CUDA_ARCH_LIST="9.0"
+        else
+            TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0+PTX"
+        fi
         EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         ;;
     12.4)
+        if [[ "$GPU_ARCH_TYPE" = "cuda-aarch64" ]]; then
+            TORCH_CUDA_ARCH_LIST="9.0"
+        else
+            TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
+        fi
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    12.1)
         TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
         EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         ;;
@@ -119,16 +126,7 @@ if [[ $USE_CUSPARSELT == "1" && $CUDA_VERSION == "11.8" ]]; then
         )
 fi
 
-
-# Turn USE_CUFILE off for CUDA 11.8, 12.4 since nvidia-cufile-cu11 and 1.9.0.20 are
-# not available in PYPI
-if [[ $CUDA_VERSION == "11.8" || $CUDA_VERSION == "12.4" ]]; then
-    export USE_CUFILE=0
-fi
-
-
-# CUDA_VERSION 12.4, 12.6, 12.8
-if [[ $CUDA_VERSION == 12* ]]; then
+if [[ $CUDA_VERSION == "12.4" || $CUDA_VERSION == "12.6" ]]; then
     export USE_STATIC_CUDNN=0
     # Try parallelizing nvcc as well
     export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
@@ -169,16 +167,6 @@ if [[ $CUDA_VERSION == 12* ]]; then
             "libnvrtc.so.12"
             "libnvrtc-builtins.so"
         )
-        if [[ $USE_CUFILE == 1 ]]; then
-            DEPS_LIST+=(
-                "/usr/local/cuda/lib64/libcufile.so.0"
-                "/usr/local/cuda/lib64/libcufile_rdma.so.1"
-            )
-            DEPS_SONAME+=(
-                "libcufile.so.0"
-                "libcufile_rdma.so.1"
-            )
-        fi
     else
         echo "Using nvidia libs from pypi."
         CUDA_RPATHS=(
@@ -195,11 +183,6 @@ if [[ $CUDA_VERSION == 12* ]]; then
             '$ORIGIN/../../nvidia/nccl/lib'
             '$ORIGIN/../../nvidia/nvtx/lib'
         )
-        if [[ $USE_CUFILE == 1 ]]; then
-            CUDA_RPATHS+=(
-                '$ORIGIN/../../nvidia/cufile/lib'
-            )
-        fi
         CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
         export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
         export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'

.ci/manywheel/build_rocm.sh

@@ -186,6 +186,15 @@ do
     OS_SO_FILES[${#OS_SO_FILES[@]}]=$file_name # Append lib to array
 done
 
+# FIXME: Temporary until https://github.com/pytorch/pytorch/pull/137443 lands
+# Install AOTriton
+if [ -e ${PYTORCH_ROOT}/.ci/docker/aotriton_version.txt ]; then
+    cp -a ${PYTORCH_ROOT}/.ci/docker/aotriton_version.txt aotriton_version.txt
+    bash ${PYTORCH_ROOT}/.ci/docker/common/install_aotriton.sh ${ROCM_HOME} && rm aotriton_version.txt
+    export AOTRITON_INSTALLED_PREFIX=${ROCM_HOME}/aotriton
+    ROCM_SO_FILES+=("libaotriton_v2.so")
+fi
+
 # rocBLAS library files
 ROCBLAS_LIB_SRC=$ROCM_HOME/lib/rocblas/library
 ROCBLAS_LIB_DST=lib/rocblas/library
@@ -257,6 +266,20 @@ RCCL_SHARE_FILES=($(ls $RCCL_SHARE_SRC))
 DEPS_AUX_SRCLIST+=(${RCCL_SHARE_FILES[@]/#/$RCCL_SHARE_SRC/})
 DEPS_AUX_DSTLIST+=(${RCCL_SHARE_FILES[@]/#/$RCCL_SHARE_DST/})
 
+# PyTorch 2.6+ (AOTriton 0.8b+)
+# AKS = "AOTriton Kernel Storage", a file format to store GPU kernels compactly
+if (( $(echo "${PYTORCH_VERSION} 2.6" | awk '{print ($1 >= $2)}') )); then
+    LIBAOTRITON_DIR=$(find "$ROCM_HOME/lib/" -name "libaotriton_v2.so" -printf '%h\n')
+    if [[ -z ${LIBAOTRITON_DIR} ]]; then
+        LIBAOTRITON_DIR=$(find "$ROCM_HOME/" -name "libaotriton_v2.so" -printf '%h\n')
+    fi
+    AKS_FILES=($(find "${LIBAOTRITON_DIR}/aotriton.images" -type f -name '*.aks?' -printf '%P\n'))
+    AKS_SRC="${LIBAOTRITON_DIR}/aotriton.images"
+    AKS_DST="lib/aotriton.images"
+    DEPS_AUX_SRCLIST+=(${AKS_FILES[@]/#/${AKS_SRC}/})
+    DEPS_AUX_DSTLIST+=(${AKS_FILES[@]/#/${AKS_DST}/})
+fi
+
 echo "PYTORCH_ROCM_ARCH: ${PYTORCH_ROCM_ARCH}"
 
 SCRIPTPATH="$( cd "$(dirname "$0")" ; pwd -P )"

.ci/pytorch/build.sh

@@ -228,7 +228,7 @@ if [[ "$BUILD_ENVIRONMENT" == *-debug* ]]; then
   export CMAKE_BUILD_TYPE=RelWithAssert
 fi
 
-# Do not change workspace permissions for ROCm and s390x CI jobs
+# Do not change workspace permissions for ROCm CI jobs
 # as it can leave workspace with bad permissions for cancelled jobs
 if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /var/lib/jenkins/workspace ]]; then
   # Workaround for dind-rootless userid mapping (https://github.com/pytorch/ci-infra/issues/96)

.ci/pytorch/check_binary.sh

@@ -387,7 +387,7 @@ fi
 ###############################################################################
 # Check for C++ ABI compatibility between gcc7 and gcc9 compiled binaries
 ###############################################################################
-if [[ "$(uname)" == 'Linux' &&  "$PACKAGE_TYPE" == 'manywheel' ]]; then
+if [[ "$(uname)" == 'Linux' && ("$PACKAGE_TYPE" == 'conda' || "$PACKAGE_TYPE" == 'manywheel')]]; then
   pushd /tmp
   python -c "import torch; exit(0 if torch.compiled_with_cxx11_abi() else (0 if torch._C._PYBIND11_BUILD_ABI == '_cxxabi1011' else 1))"
   popd

.ci/pytorch/common_utils.sh

@@ -169,34 +169,24 @@ function install_torchrec_and_fbgemm() {
   torchrec_commit=$(get_pinned_commit torchrec)
   local fbgemm_commit
   fbgemm_commit=$(get_pinned_commit fbgemm)
-  if [[ "$BUILD_ENVIRONMENT" == *rocm* ]] ; then
-    fbgemm_commit=$(get_pinned_commit fbgemm_rocm)
-  fi
   pip_uninstall torchrec-nightly
   pip_uninstall fbgemm-gpu-nightly
   pip_install setuptools-git-versioning scikit-build pyre-extensions
 
-  if [[ "$BUILD_ENVIRONMENT" == *rocm* ]] ; then
-    # install torchrec first because it installs fbgemm nightly on top of rocm fbgemm
-    pip_install --no-use-pep517 --user "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}"
-    pip_uninstall fbgemm-gpu-nightly
+  # TODO (huydhn): I still have no clue on why sccache doesn't work with only fbgemm_gpu here, but it
+  # seems to be an sccache-related issue
+  if [[ "$IS_A100_RUNNER" == "1" ]]; then
+    unset CMAKE_CUDA_COMPILER_LAUNCHER
+    sudo mv /opt/cache/bin /opt/cache/bin-backup
+  fi
 
-    pip_install tabulate  # needed for newer fbgemm
-    pip_install patchelf  # needed for rocm fbgemm
-    git clone --recursive https://github.com/pytorch/fbgemm
-    pushd fbgemm/fbgemm_gpu
-    git checkout "${fbgemm_commit}"
-    python setup.py install \
-      --package_variant=rocm \
-      -DHIP_ROOT_DIR="${ROCM_PATH}" \
-      -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
-      -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
-    popd
-    rm -rf fbgemm
-  else
-    # See https://github.com/pytorch/pytorch/issues/106971
-    CUDA_PATH=/usr/local/cuda-12.1 pip_install --no-use-pep517 --user "git+https://github.com/pytorch/FBGEMM.git@${fbgemm_commit}#egg=fbgemm-gpu&subdirectory=fbgemm_gpu"
-    pip_install --no-use-pep517 --user "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}"
+  # See https://github.com/pytorch/pytorch/issues/106971
+  CUDA_PATH=/usr/local/cuda-12.1 pip_install --no-use-pep517 --user "git+https://github.com/pytorch/FBGEMM.git@${fbgemm_commit}#egg=fbgemm-gpu&subdirectory=fbgemm_gpu"
+  pip_install --no-use-pep517 --user "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}"
+
+  if [[ "$IS_A100_RUNNER" == "1" ]]; then
+    export CMAKE_CUDA_COMPILER_LAUNCHER=/opt/cache/bin/sccache
+    sudo mv /opt/cache/bin-backup /opt/cache/bin
   fi
 }
 
@@ -226,11 +216,6 @@ function checkout_install_torchbench() {
     # to install and test other models
     python install.py --continue_on_fail
   fi
-
-  # TODO (huydhn): transformers-4.44.2 added by https://github.com/pytorch/benchmark/pull/2488
-  # is regressing speedup metric. This needs to be investigated further
-  pip install transformers==4.38.1
-
   echo "Print all dependencies after TorchBench is installed"
   python -mpip freeze
   popd

.ci/pytorch/macos-test.sh

@@ -18,9 +18,6 @@ if [[ ! $(python -c "import torch; print(int(torch.backends.openmp.is_available(
 fi
 popd
 
-# enable debug asserts in serialization
-export TORCH_SERIALIZATION_DEBUG=1
-
 setup_test_python() {
   # The CircleCI worker hostname doesn't resolve to an address.
   # This environment variable makes ProcessGroupGloo default to

.ci/pytorch/multigpu-test.sh

@@ -8,62 +8,55 @@
 source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
 
 echo "Testing pytorch"
-# When adding more tests, please use HUD to see which shard is shorter
-if [[ "${SHARD_NUMBER:-1}" == "1" ]]; then
-    # FSDP tests
-    for f in test/distributed/fsdp/*.py ; do time python test/run_test.py --verbose -i "${f#*/}" ; done
-fi
+time python test/run_test.py --include test_cuda_multigpu test_cuda_primary_ctx --verbose
 
-if [[ "${SHARD_NUMBER:-2}" == "2" ]]; then
-    time python test/run_test.py --include test_cuda_multigpu test_cuda_primary_ctx --verbose
+# Disabling tests to see if they solve timeout issues; see https://github.com/pytorch/pytorch/issues/70015
+# python tools/download_mnist.py --quiet -d test/cpp/api/mnist
+# OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="test/cpp/api/mnist" build/bin/test_api
+time python test/run_test.py --verbose -i distributed/test_c10d_common
+time python test/run_test.py --verbose -i distributed/test_c10d_gloo
+time python test/run_test.py --verbose -i distributed/test_c10d_nccl
+time python test/run_test.py --verbose -i distributed/test_c10d_spawn_gloo
+time python test/run_test.py --verbose -i distributed/test_c10d_spawn_nccl
+time python test/run_test.py --verbose -i distributed/test_compute_comm_reordering
+time python test/run_test.py --verbose -i distributed/test_store
+time python test/run_test.py --verbose -i distributed/test_symmetric_memory
+time python test/run_test.py --verbose -i distributed/test_pg_wrapper
+time python test/run_test.py --verbose -i distributed/rpc/cuda/test_tensorpipe_agent
+# FSDP tests
+for f in test/distributed/fsdp/*.py ; do time python test/run_test.py --verbose -i "${f#*/}" ; done
+# ShardedTensor tests
+time python test/run_test.py --verbose -i distributed/checkpoint/test_checkpoint
+time python test/run_test.py --verbose -i distributed/checkpoint/test_file_system_checkpoint
+time python test/run_test.py --verbose -i distributed/_shard/sharding_spec/test_sharding_spec
+time python test/run_test.py --verbose -i distributed/_shard/sharding_plan/test_sharding_plan
+time python test/run_test.py --verbose -i distributed/_shard/sharded_tensor/test_sharded_tensor
+time python test/run_test.py --verbose -i distributed/_shard/sharded_tensor/test_sharded_tensor_reshard
 
-    # Disabling tests to see if they solve timeout issues; see https://github.com/pytorch/pytorch/issues/70015
-    # python tools/download_mnist.py --quiet -d test/cpp/api/mnist
-    # OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="test/cpp/api/mnist" build/bin/test_api
-    time python test/run_test.py --verbose -i distributed/test_c10d_common
-    time python test/run_test.py --verbose -i distributed/test_c10d_gloo
-    time python test/run_test.py --verbose -i distributed/test_c10d_nccl
-    time python test/run_test.py --verbose -i distributed/test_c10d_spawn_gloo
-    time python test/run_test.py --verbose -i distributed/test_c10d_spawn_nccl
-    time python test/run_test.py --verbose -i distributed/test_compute_comm_reordering
-    time python test/run_test.py --verbose -i distributed/test_store
-    time python test/run_test.py --verbose -i distributed/test_symmetric_memory
-    time python test/run_test.py --verbose -i distributed/test_pg_wrapper
-    time python test/run_test.py --verbose -i distributed/rpc/cuda/test_tensorpipe_agent
+# functional collective tests
+time python test/run_test.py --verbose -i distributed/test_functional_api
 
-    # ShardedTensor tests
-    time python test/run_test.py --verbose -i distributed/checkpoint/test_checkpoint
-    time python test/run_test.py --verbose -i distributed/checkpoint/test_file_system_checkpoint
-    time python test/run_test.py --verbose -i distributed/_shard/sharding_spec/test_sharding_spec
-    time python test/run_test.py --verbose -i distributed/_shard/sharding_plan/test_sharding_plan
-    time python test/run_test.py --verbose -i distributed/_shard/sharded_tensor/test_sharded_tensor
-    time python test/run_test.py --verbose -i distributed/_shard/sharded_tensor/test_sharded_tensor_reshard
+# DTensor tests
+time python test/run_test.py --verbose -i distributed/_tensor/test_random_ops
+time python test/run_test.py --verbose -i distributed/_tensor/test_dtensor_compile
 
-    # functional collective tests
-    time python test/run_test.py --verbose -i distributed/test_functional_api
+# DeviceMesh test
+time python test/run_test.py --verbose -i distributed/test_device_mesh
 
-    # DTensor tests
-    time python test/run_test.py --verbose -i distributed/tensor/test_random_ops
-    time python test/run_test.py --verbose -i distributed/tensor/test_dtensor_compile
+# DTensor/TP tests
+time python test/run_test.py --verbose -i distributed/tensor/parallel/test_tp_examples
+time python test/run_test.py --verbose -i distributed/tensor/parallel/test_tp_random_state
 
-    # DeviceMesh test
-    time python test/run_test.py --verbose -i distributed/test_device_mesh
+# FSDP2 tests
+time python test/run_test.py --verbose -i distributed/_composable/fsdp/test_fully_shard_training -- -k test_2d_mlp_with_nd_mesh
 
-    # DTensor/TP tests
-    time python test/run_test.py --verbose -i distributed/tensor/parallel/test_tp_examples
-    time python test/run_test.py --verbose -i distributed/tensor/parallel/test_tp_random_state
+# ND composability tests
+time python test/run_test.py --verbose -i distributed/_composable/test_composability/test_2d_composability
+time python test/run_test.py --verbose -i distributed/_composable/test_composability/test_pp_composability
 
-    # FSDP2 tests
-    time python test/run_test.py --verbose -i distributed/_composable/fsdp/test_fully_shard_training -- -k test_2d_mlp_with_nd_mesh
-
-    # ND composability tests
-    time python test/run_test.py --verbose -i distributed/_composable/test_composability/test_2d_composability
-    time python test/run_test.py --verbose -i distributed/_composable/test_composability/test_pp_composability
-
-    # Other tests
-    time python test/run_test.py --verbose -i test_cuda_primary_ctx
-    time python test/run_test.py --verbose -i test_optim -- -k test_forloop_goes_right_direction_multigpu
-    time python test/run_test.py --verbose -i test_optim -- -k test_mixed_device_dtype
-    time python test/run_test.py --verbose -i test_foreach -- -k test_tensors_grouping
-fi
+# Other tests
+time python test/run_test.py --verbose -i test_cuda_primary_ctx
+time python test/run_test.py --verbose -i test_optim -- -k test_forloop_goes_right_direction_multigpu
+time python test/run_test.py --verbose -i test_optim -- -k test_mixed_device_dtype
+time python test/run_test.py --verbose -i test_foreach -- -k test_tensors_grouping
 assert_git_not_dirty

.ci/pytorch/run_tests.sh

@@ -40,7 +40,7 @@ retry () {
 if [[ "$#" != 3 ]]; then
   if [[ -z "${DESIRED_PYTHON:-}" || -z "${DESIRED_CUDA:-}" || -z "${PACKAGE_TYPE:-}" ]]; then
     echo "USAGE: run_tests.sh  PACKAGE_TYPE  DESIRED_PYTHON  DESIRED_CUDA"
-    echo "The env variable PACKAGE_TYPE must be set to 'manywheel' or 'libtorch'"
+    echo "The env variable PACKAGE_TYPE must be set to 'conda' or 'manywheel' or 'libtorch'"
     echo "The env variable DESIRED_PYTHON must be set like '2.7mu' or '3.6m' etc"
     echo "The env variable DESIRED_CUDA must be set like 'cpu' or 'cu80' etc"
     exit 1

.ci/pytorch/smoke_test/check_binary_symbols.py

@@ -6,7 +6,7 @@ import itertools
 import os
 import re
 from pathlib import Path
-from typing import Any
+from typing import Any, List, Tuple
 
 
 # We also check that there are [not] cxx11 symbols in libtorch
@@ -46,17 +46,17 @@ LIBTORCH_PRE_CXX11_PATTERNS = _apply_libtorch_symbols(PRE_CXX11_SYMBOLS)
 
 
 @functools.lru_cache(100)
-def get_symbols(lib: str) -> list[tuple[str, str, str]]:
+def get_symbols(lib: str) -> List[Tuple[str, str, str]]:
     from subprocess import check_output
 
     lines = check_output(f'nm "{lib}"|c++filt', shell=True)
     return [x.split(" ", 2) for x in lines.decode("latin1").split("\n")[:-1]]
 
 
-def grep_symbols(lib: str, patterns: list[Any]) -> list[str]:
+def grep_symbols(lib: str, patterns: List[Any]) -> List[str]:
     def _grep_symbols(
-        symbols: list[tuple[str, str, str]], patterns: list[Any]
-    ) -> list[str]:
+        symbols: List[Tuple[str, str, str]], patterns: List[Any]
+    ) -> List[str]:
         rc = []
         for _s_addr, _s_type, s_name in symbols:
             for pattern in patterns:

.ci/pytorch/smoke_test/smoke_test.py

@@ -6,7 +6,6 @@ import re
 import subprocess
 import sys
 from pathlib import Path
-from tempfile import NamedTemporaryFile
 
 import torch
 import torch._dynamo
@@ -162,32 +161,6 @@ def test_cuda_runtime_errors_captured() -> None:
         raise RuntimeError("Expected CUDA RuntimeError but have not received!")
 
 
-def test_cuda_gds_errors_captured() -> None:
-    major_version = int(torch.version.cuda.split(".")[0])
-    minor_version = int(torch.version.cuda.split(".")[1])
-
-    if major_version < 12 or (major_version == 12 and minor_version < 6):
-        print("CUDA version is not supported for GDS smoke test")
-        return
-
-    cuda_exception_missed = True
-    try:
-        print("Testing test_cuda_gds_errors_captured")
-        with NamedTemporaryFile() as f:
-            torch.cuda.gds.GdsFile(f.name, os.O_CREAT | os.O_RDWR)
-    except RuntimeError as e:
-        expected_error = "cuFileHandleRegister failed"
-        if re.search(expected_error, f"{e}"):
-            print(f"Caught CUDA exception with success: {e}")
-            cuda_exception_missed = False
-        else:
-            raise e
-    if cuda_exception_missed:
-        raise RuntimeError(
-            "Expected cuFileHandleRegister failed RuntimeError but have not received!"
-        )
-
-
 def smoke_test_cuda(
     package: str, runtime_error_check: str, torch_compile_check: str
 ) -> None:
@@ -408,7 +381,6 @@ def main() -> None:
     test_numpy()
     if is_cuda_system:
         test_linalg("cuda")
-        test_cuda_gds_errors_captured()
 
     if options.package == "all":
         smoke_test_modules()

.ci/pytorch/test.sh

@@ -12,9 +12,9 @@ export TERM=vt100
 # shellcheck source=./common.sh
 source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
 
-# Do not change workspace permissions for ROCm and s390x CI jobs
+# Do not change workspace permissions for ROCm CI jobs
 # as it can leave workspace with bad permissions for cancelled jobs
-if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /var/lib/jenkins/workspace ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *rocm* && -d /var/lib/jenkins/workspace ]]; then
   # Workaround for dind-rootless userid mapping (https://github.com/pytorch/ci-infra/issues/96)
   WORKSPACE_ORIGINAL_OWNER_ID=$(stat -c '%u' "/var/lib/jenkins/workspace")
   cleanup_workspace() {
@@ -46,9 +46,6 @@ BUILD_BIN_DIR="$BUILD_DIR"/bin
 SHARD_NUMBER="${SHARD_NUMBER:=1}"
 NUM_TEST_SHARDS="${NUM_TEST_SHARDS:=1}"
 
-# enable debug asserts in serialization
-export TORCH_SERIALIZATION_DEBUG=1
-
 export VALGRIND=ON
 # export TORCH_INDUCTOR_INSTALL_GXX=ON
 if [[ "$BUILD_ENVIRONMENT" == *clang9* || "$BUILD_ENVIRONMENT" == *xpu* ]]; then
@@ -89,13 +86,6 @@ if [[ "$BUILD_ENVIRONMENT" == *clang9* || "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   export VALGRIND=OFF
 fi
 
-
-if [[ "$BUILD_ENVIRONMENT" == *s390x* ]]; then
-  # There are additional warnings on s390x, maybe due to newer gcc.
-  # Skip this check for now
-  export VALGRIND=OFF
-fi
-
 if [[ "${PYTORCH_TEST_RERUN_DISABLED_TESTS}" == "1" ]] || [[ "${CONTINUE_THROUGH_ERROR}" == "1" ]]; then
   # When rerunning disable tests, do not generate core dumps as it could consume
   # the runner disk space when crashed tests are run multiple times. Running out
@@ -139,7 +129,7 @@ if [[ "$TEST_CONFIG" == 'default' ]]; then
 fi
 
 if [[ "$TEST_CONFIG" == 'distributed' ]] && [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
-  export HIP_VISIBLE_DEVICES=0,1,2,3
+  export HIP_VISIBLE_DEVICES=0,1
 fi
 
 if [[ "$TEST_CONFIG" == 'slow' ]]; then
@@ -177,9 +167,6 @@ if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
   # Print GPU info
   rocminfo
   rocminfo | grep -E 'Name:.*\sgfx|Marketing'
-
-  # for benchmarks/dynamo/check_accuracy.py, we need to put results in a rocm specific directory to avoid clashes with cuda
-  MAYBE_ROCM="rocm/"
 fi
 
 if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
@@ -342,7 +329,7 @@ test_inductor_distributed() {
   python test/run_test.py -i inductor/test_aot_inductor.py -k test_non_default_cuda_device --verbose
   python test/run_test.py -i inductor/test_aot_inductor.py -k test_replicate_on_devices --verbose
   python test/run_test.py -i distributed/test_c10d_functional_native.py --verbose
-  python test/run_test.py -i distributed/tensor/test_dtensor_compile.py --verbose
+  python test/run_test.py -i distributed/_tensor/test_dtensor_compile.py --verbose
   python test/run_test.py -i distributed/tensor/parallel/test_micro_pipeline_tp.py --verbose
   python test/run_test.py -i distributed/_composable/test_replicate_with_compiler.py --verbose
   python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_comm.py --verbose
@@ -395,32 +382,15 @@ test_inductor_aoti() {
   CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference
 }
 
-test_inductor_cpp_wrapper_shard() {
-  if [[ -z "$NUM_TEST_SHARDS" ]]; then
-    echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
-    exit 1
-  fi
-
+test_inductor_cpp_wrapper() {
   export TORCHINDUCTOR_CPP_WRAPPER=1
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
 
-  if [[ "$1" -eq "2" ]]; then
-    # For now, manually put the opinfo tests in shard 2, and all other tests in
-    # shard 1.  Test specific things triggering past bugs, for now.
-    python test/run_test.py \
-      --include inductor/test_torchinductor_opinfo \
-      -k 'linalg or to_sparse' \
-      --verbose
-    exit
-  fi
+  # Run certain inductor unit tests with cpp wrapper. In the end state, we should be able to run all the inductor
+  # unit tests with cpp wrapper.
+  python test/run_test.py --include inductor/test_torchinductor.py --verbose
 
-  # Run certain inductor unit tests with cpp wrapper. In the end state, we
-  # should be able to run all the inductor unit tests with cpp_wrapper.
-  python test/run_test.py \
-    --include inductor/test_torchinductor inductor/test_max_autotune inductor/test_cpu_repro \
-    --verbose
-  python test/run_test.py --inductor --include test_torch -k 'take' --verbose
 
   # Run inductor benchmark tests with cpp wrapper.
   # Skip benchmark tests if it's in rerun-disabled-mode.
@@ -433,7 +403,7 @@ test_inductor_cpp_wrapper_shard() {
     --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_training.csv"
     python benchmarks/dynamo/check_accuracy.py \
       --actual "$TEST_REPORTS_DIR/inductor_cpp_wrapper_training.csv" \
-      --expected "benchmarks/dynamo/ci_expected_accuracy/${MAYBE_ROCM}inductor_timm_training.csv"
+      --expected "benchmarks/dynamo/ci_expected_accuracy/inductor_timm_training.csv"
 
     python benchmarks/dynamo/torchbench.py --device cuda --accuracy \
       --bfloat16 --inference --inductor --only hf_T5 --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv"
@@ -443,7 +413,7 @@ test_inductor_cpp_wrapper_shard() {
       --bfloat16 --inference --inductor --only moco --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv"
     python benchmarks/dynamo/check_accuracy.py \
       --actual "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv" \
-      --expected "benchmarks/dynamo/ci_expected_accuracy/${MAYBE_ROCM}inductor_torchbench_inference.csv"
+      --expected "benchmarks/dynamo/ci_expected_accuracy/inductor_torchbench_inference.csv"
   fi
 }
 
@@ -518,8 +488,6 @@ test_perf_for_dashboard() {
     test_inductor_set_cpu_affinity
   elif [[ "${TEST_CONFIG}" == *cuda_a10g* ]]; then
     device=cuda_a10g
-  elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
-    device=rocm
   fi
 
   for mode in "${modes[@]}"; do
@@ -552,7 +520,7 @@ test_perf_for_dashboard() {
             --dynamic-batch-only "$@" \
             --output "$TEST_REPORTS_DIR/${backend}_dynamic_${suite}_${dtype}_${mode}_${device}_${target}.csv"
       fi
-      if [[ "$DASHBOARD_TAG" == *cppwrapper-true* ]]; then
+      if [[ "$DASHBOARD_TAG" == *cppwrapper-true* ]] && [[ "$mode" == "inference" ]]; then
         TORCHINDUCTOR_CPP_WRAPPER=1 $TASKSET python "benchmarks/dynamo/$suite.py" \
             "${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" --disable-cudagraphs "$@" \
             --output "$TEST_REPORTS_DIR/${backend}_cpp_wrapper_${suite}_${dtype}_${mode}_${device}_${target}.csv"
@@ -636,16 +604,16 @@ test_single_dynamo_benchmark() {
       TEST_CONFIG=${TEST_CONFIG//_avx512/}
     fi
     python "benchmarks/dynamo/$suite.py" \
-      --ci --accuracy --timing --explain --print-compilation-time \
+      --ci --accuracy --timing --explain \
       "${DYNAMO_BENCHMARK_FLAGS[@]}" \
       "$@" "${partition_flags[@]}" \
       --output "$TEST_REPORTS_DIR/${name}_${suite}.csv"
     python benchmarks/dynamo/check_accuracy.py \
       --actual "$TEST_REPORTS_DIR/${name}_$suite.csv" \
-      --expected "benchmarks/dynamo/ci_expected_accuracy/${MAYBE_ROCM}${TEST_CONFIG}_${name}.csv"
+      --expected "benchmarks/dynamo/ci_expected_accuracy/${TEST_CONFIG}_${name}.csv"
     python benchmarks/dynamo/check_graph_breaks.py \
       --actual "$TEST_REPORTS_DIR/${name}_$suite.csv" \
-      --expected "benchmarks/dynamo/ci_expected_accuracy/${MAYBE_ROCM}${TEST_CONFIG}_${name}.csv"
+      --expected "benchmarks/dynamo/ci_expected_accuracy/${TEST_CONFIG}_${name}.csv"
   fi
 }
 
@@ -668,7 +636,7 @@ test_inductor_halide() {
 }
 
 test_inductor_triton_cpu() {
-  python test/run_test.py --include inductor/test_triton_cpu_backend.py inductor/test_torchinductor_strided_blocks.py --verbose
+  python test/run_test.py --include inductor/test_triton_cpu_backend.py --verbose
   assert_git_not_dirty
 }
 
@@ -732,7 +700,7 @@ test_inductor_torchbench_smoketest_perf() {
       --only $test --output "$TEST_REPORTS_DIR/inductor_warm_start_smoketest_$test.csv"
     python benchmarks/dynamo/check_accuracy.py \
       --actual "$TEST_REPORTS_DIR/inductor_warm_start_smoketest_$test.csv" \
-      --expected "benchmarks/dynamo/ci_expected_accuracy/${MAYBE_ROCM}inductor_huggingface_training.csv"
+      --expected "benchmarks/dynamo/ci_expected_accuracy/inductor_huggingface_training.csv"
   done
 }
 
@@ -928,20 +896,10 @@ test_libtorch_api() {
   else
     # Exclude IMethodTest that relies on torch::deploy, which will instead be ran in test_deploy
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_api -k "not IMethodTest"
-
-    # On s390x, pytorch is built without llvm.
-    # Even if it would be built with llvm, llvm currently doesn't support used features on s390x and
-    # test fails with errors like:
-    # JIT session error: Unsupported target machine architecture in ELF object pytorch-jitted-objectbuffer
-    # unknown file: Failure
-    # C++ exception with description "valOrErr INTERNAL ASSERT FAILED at "/var/lib/jenkins/workspace/torch/csrc/jit/tensorexpr/llvm_jit.h":34, please report a bug to PyTorch. Unexpected failure in LLVM JIT: Failed to materialize symbols: { (main, { func }) }
-    if [[ "${BUILD_ENVIRONMENT}" != *s390x* ]]; then
-      python test/run_test.py --cpp --verbose -i cpp/test_tensorexpr
-    fi
+    python test/run_test.py --cpp --verbose -i cpp/test_tensorexpr
   fi
 
-  # quantization is not fully supported on s390x yet
-  if [[ "${BUILD_ENVIRONMENT}" != *android* && "${BUILD_ENVIRONMENT}" != *cuda* && "${BUILD_ENVIRONMENT}" != *asan* && "${BUILD_ENVIRONMENT}" != *s390x* ]]; then
+  if [[ "${BUILD_ENVIRONMENT}" != *android* && "${BUILD_ENVIRONMENT}" != *cuda* && "${BUILD_ENVIRONMENT}" != *asan* ]]; then
     # NB: This test is not under TORCH_BIN_DIR but under BUILD_BIN_DIR
     export CPP_TESTS_DIR="${BUILD_BIN_DIR}"
     python test/run_test.py --cpp --verbose -i cpp/static_runtime_test
@@ -1439,7 +1397,7 @@ test_executorch() {
 test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
-        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops \
+        test_foreach test_reductions test_unary_ufuncs \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1542,7 +1500,7 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchaudio cuda
   install_torchvision
   checkout_install_torchbench hf_T5 llama moco
-  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
+  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   test_inductor_shard "${SHARD_NUMBER}"

.ci/pytorch/test_example_code/cnn_smoke_win_arm64.py

@@ -1,41 +0,0 @@
-r"""
-It's used to check basic rnn features with cpu-only.
-For example, it would throw exception if some components are missing
-"""
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import torch.optim as optim
-
-
-class SimpleCNN(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.conv = nn.Conv2d(1, 1, 3)
-        self.pool = nn.MaxPool2d(2, 2)
-
-    def forward(self, inputs):
-        output = self.pool(F.relu(self.conv(inputs)))
-        output = output.view(1)
-        return output
-
-
-try:
-    # Mock one infer
-    net = SimpleCNN()
-    net_inputs = torch.rand((1, 1, 5, 5))
-    outputs = net(net_inputs)
-    print(outputs)
-
-    criterion = nn.MSELoss()
-    optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.1)
-
-    # Mock one step training
-    label = torch.full((1,), 1.0, dtype=torch.float)
-    loss = criterion(outputs, label)
-    loss.backward()
-    optimizer.step()
-
-except Exception as e:
-    print(f"An error occurred: {e}")

.ci/pytorch/test_example_code/rnn_smoke_win_arm64.py

@@ -1,13 +0,0 @@
-r"""
-It's used to check basic rnn features with cpu-only.
-For example, it would throw exception if missing some components are missing
-"""
-
-import torch
-import torch.nn as nn
-
-
-rnn = nn.RNN(10, 20, 2)
-inputs = torch.randn(5, 3, 10)
-h0 = torch.randn(2, 3, 20)
-output, hn = rnn(inputs, h0)

.ci/pytorch/win-test.sh

@@ -18,9 +18,6 @@ export PYTORCH_FINAL_PACKAGE_DIR="${PYTORCH_FINAL_PACKAGE_DIR:-/c/w/build-result
 PYTORCH_FINAL_PACKAGE_DIR_WIN=$(cygpath -w "${PYTORCH_FINAL_PACKAGE_DIR}")
 export PYTORCH_FINAL_PACKAGE_DIR_WIN
 
-# enable debug asserts in serialization
-export TORCH_SERIALIZATION_DEBUG=1
-
 mkdir -p "$TMP_DIR"/build/torch
 
 export SCRIPT_HELPERS_DIR=$SCRIPT_PARENT_DIR/win-test-helpers

.ci/pytorch/windows/arm64/bootstrap_apl.bat

@@ -1,31 +0,0 @@
-@echo off
-
-echo Dependency ARM Performance Libraries (APL) installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
-if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
-
-:: Set download URL for the ARM Performance Libraries (APL)
-set DOWNLOAD_URL="https://developer.arm.com/-/cdn-downloads/permalink/Arm-Performance-Libraries/Version_24.10/arm-performance-libraries_24.10_Windows.msi"
-set INSTALLER_FILE=%DOWNLOADS_DIR%\arm-performance-libraries.msi
-
-:: Download installer
-echo Downloading ARM Performance Libraries (APL)...
-curl -L -o "%INSTALLER_FILE%" %DOWNLOAD_URL%
-
-:: Install ARM Performance Libraries (APL)
-echo Installing ARM Performance Libraries (APL)...
-msiexec /i "%INSTALLER_FILE%" /qn /norestart ACCEPT_EULA=1 INSTALLFOLDER="%DEPENDENCIES_DIR%"
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install ARM Performance Libraries (APL) components. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-:: Add to environment
-echo ARMPL_DIR=%DEPENDENCIES_DIR%\armpl_24.10\>> %GITHUB_ENV%
-echo %DEPENDENCIES_DIR%\armpl_24.10\bin\>> %GITHUB_PATH%
-
-echo Dependency ARM Performance Libraries (APL) installation finished.

.ci/pytorch/windows/arm64/bootstrap_buildtools.bat

@@ -1,49 +0,0 @@
-@echo off
-
-echo Dependency MSVC Build Tools with C++ with ARM64/ARM64EC components installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir "%DOWNLOADS_DIR%"
-if not exist "%DEPENDENCIES_DIR%" mkdir "%DEPENDENCIES_DIR%"
-
-:: Set download URL for the Visual Studio Installer
-set DOWNLOAD_URL=https://aka.ms/vs/17/release/vs_BuildTools.exe
-set INSTALLER_FILE=%DOWNLOADS_DIR%\vs_BuildTools.exe
-
-:: Download installer
-echo Downloading Visual Studio Build Tools with C++ installer...
-curl -L -o "%INSTALLER_FILE%" %DOWNLOAD_URL%
-
-:: Install the Visual Studio Build Tools with C++ components
-echo Installing Visual Studio Build Tools with C++ components...
-echo Installing MSVC %MSVC_VERSION%
-if "%MSVC_VERSION%" == "latest" (
-    "%INSTALLER_FILE%" --norestart --nocache --quiet --wait --installPath "%DEPENDENCIES_DIR%\VSBuildTools" ^
-        --add Microsoft.VisualStudio.Component.Windows11SDK.22621 ^
-        --add Microsoft.VisualStudio.Component.VC.ASAN ^
-        --add Microsoft.VisualStudio.Component.VC.CMake.Project ^
-        --add Microsoft.VisualStudio.Component.VC.Tools.ARM64 ^
-        --add Microsoft.VisualStudio.Component.VC.Tools.x86.x64
-) else if "%MSVC_VERSION%" == "14.40" (
-    "%INSTALLER_FILE%" --norestart --nocache --quiet --wait --installPath "%DEPENDENCIES_DIR%\VSBuildTools" ^
-        --add Microsoft.VisualStudio.Component.Windows11SDK.22621 ^
-        --add Microsoft.VisualStudio.Component.VC.ASAN ^
-        --add Microsoft.VisualStudio.Component.VC.CMake.Project ^
-        --add Microsoft.VisualStudio.Component.VC.14.40.17.10.ARM64 ^
-        --add Microsoft.VisualStudio.Component.VC.14.40.17.10.x86.x64
-) else if "%MSVC_VERSION%" == "14.36" (
-    "%INSTALLER_FILE%" --norestart --nocache --quiet --wait --installPath "%DEPENDENCIES_DIR%\VSBuildTools" ^
-        --add Microsoft.VisualStudio.Component.Windows11SDK.22621 ^
-        --add Microsoft.VisualStudio.Component.VC.ASAN ^
-        --add Microsoft.VisualStudio.Component.VC.CMake.Project ^
-        --add Microsoft.VisualStudio.Component.VC.14.36.17.6.ARM64 ^
-        --add Microsoft.VisualStudio.Component.VC.14.36.17.6.x86.x64
-)
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install Visual Studio Build Tools with C++ components. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-echo Dependency Visual Studio Build Tools with C++ installation finished.

.ci/pytorch/windows/arm64/bootstrap_git.bat

@@ -1,37 +0,0 @@
-:: we need to install newer version of Git manually as "-submodules" function is not supported in the default version of runner.
-
-@echo off
-
-echo Dependency Git installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
-if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
-
-:: Set download URL for the Git
-set DOWNLOAD_URL="https://github.com/git-for-windows/git/releases/download/v2.46.0.windows.1/Git-2.46.0-64-bit.exe"
-set INSTALLER_FILE=%DOWNLOADS_DIR%\Git-2.46.0-64-bit.exe
-
-:: Download installer
-echo Downloading Git...
-curl -L -o "%INSTALLER_FILE%" %DOWNLOAD_URL%
-
-:: Install Git
-echo Installing Git...
-"%INSTALLER_FILE%" /VERYSILENT /DIR="%DEPENDENCIES_DIR%\git"
-
-dir %DEPENDENCIES_DIR%\git
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install Git. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-:: Enable long paths
-call "%DEPENDENCIES_DIR%\git\cmd\git.exe" config --system core.longpaths true
-
-:: Add to PATH
-echo %DEPENDENCIES_DIR%\git\cmd\;%DEPENDENCIES_DIR%\git\bin\>> %GITHUB_PATH%
-
-echo Dependency Git installation finished.

.ci/pytorch/windows/arm64/bootstrap_libuv.bat

@@ -1,33 +0,0 @@
-@echo off
-
-echo Dependency libuv installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
-if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
-
-:: activate visual studio
-call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
-where cl.exe
-
-cd %DEPENDENCIES_DIR%
-git clone https://github.com/libuv/libuv.git -b v1.39.0
-
-echo Configuring libuv...
-mkdir libuv\build
-cd libuv\build
-cmake .. -DBUILD_TESTING=OFF
-
-echo Building libuv...
-cmake --build . --config Release
-
-echo Installing libuv...
-cmake --install . --prefix ../install
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install libuv. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-echo Dependency libuv installation finished.

.ci/pytorch/windows/arm64/bootstrap_openblas.bat

@@ -1,46 +0,0 @@
-@echo off
-
-echo Dependency OpenBLAS installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
-if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
-
-:: activate visual studio
-call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
-where cl.exe
-
-:: Clone OpenBLAS
-cd %DEPENDENCIES_DIR%
-git clone https://github.com/OpenMathLib/OpenBLAS.git -b v0.3.29
-
-echo Configuring OpenBLAS...
-mkdir OpenBLAS\build
-cd OpenBLAS\build
-cmake .. -G Ninja ^
-  -DBUILD_TESTING=0 ^
-  -DBUILD_BENCHMARKS=0 ^
-  -DC_LAPACK=1 ^
-  -DNOFORTRAN=1 ^
-  -DDYNAMIC_ARCH=0 ^
-  -DARCH=arm64 ^
-  -DBINARY=64 ^
-  -DTARGET=GENERIC ^
-  -DUSE_OPENMP=1 ^
-  -DCMAKE_SYSTEM_PROCESSOR=ARM64 ^
-  -DCMAKE_SYSTEM_NAME=Windows ^
-  -DCMAKE_BUILD_TYPE=Release
-
-echo Building OpenBLAS...
-cmake --build . --config Release
-
-echo Installing OpenBLAS...
-cmake --install . --prefix ../install
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install OpenBLAS. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-echo Dependency OpenBLAS installation finished.

.ci/pytorch/windows/arm64/bootstrap_python.bat

@@ -1,41 +0,0 @@
-@echo off
-
-echo Dependency Python installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
-if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
-
-if "%PYTHON_VERSION%"=="Python312" (
-    echo Python version is set to Python312
-    set DOWNLOAD_URL="https://www.python.org/ftp/python/3.12.7/python-3.12.7-arm64.exe"
-) else if "%PYTHON_VERSION%"=="Python311" (
-    echo Python version is set to Python311
-    set DOWNLOAD_URL="https://www.python.org/ftp/python/3.11.9/python-3.11.9-arm64.exe"
-) else (
-    echo PYTHON_VERSION not defined, Python version is set to Python312
-    set DOWNLOAD_URL="https://www.python.org/ftp/python/3.12.7/python-3.12.7-arm64.exe"
-)
-
-set INSTALLER_FILE=%DOWNLOADS_DIR%\python-installer.exe
-
-:: Download installer
-echo Downloading Python...
-curl -L -o "%INSTALLER_FILE%" %DOWNLOAD_URL%
-
-:: Install Python
-echo Installing Python...
-"%INSTALLER_FILE%" /quiet Include_debug=1 TargetDir="%DEPENDENCIES_DIR%\Python"
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install Python. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-:: Add to PATH
-echo %DEPENDENCIES_DIR%\Python\>> %GITHUB_PATH%
-echo %DEPENDENCIES_DIR%\Python\scripts\>> %GITHUB_PATH%
-echo %DEPENDENCIES_DIR%\Python\libs\>> %GITHUB_PATH%
-
-echo Dependency Python installation finished.

.ci/pytorch/windows/arm64/bootstrap_rust.bat

@@ -1,33 +0,0 @@
-@echo off
-
-echo Dependency Rust installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
-if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
-
-set DOWNLOAD_URL="https://static.rust-lang.org/rustup/dist/x86_64-pc-windows-msvc/rustup-init.exe"
-set INSTALLER_FILE=%DOWNLOADS_DIR%\rustup-init.exe
-set RUSTUP_HOME=%DEPENDENCIES_DIR%\rust
-set CARGO_HOME=%DEPENDENCIES_DIR%\cargo
-
-:: Download installer
-echo Downloading Rust...
-curl -L -o "%INSTALLER_FILE%" %DOWNLOAD_URL%
-
-:: Install APL
-echo Installing Rust...
-"%INSTALLER_FILE%" -q -y --default-host aarch64-pc-windows-msvc --default-toolchain stable --profile default
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install Rust. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-:: Add to PATH
-echo %DEPENDENCIES_DIR%\cargo\bin\>> %GITHUB_PATH%
-echo RUSTUP_HOME=%DEPENDENCIES_DIR%\rust>> %GITHUB_ENV%
-echo CARGO_HOME=%DEPENDENCIES_DIR%\cargo>> %GITHUB_ENV%
-
-echo Dependency Rust installation finished.

.ci/pytorch/windows/arm64/bootstrap_sccache.bat

@@ -1,33 +0,0 @@
-@echo off
-
-echo Dependency sccache installation started.
-
-:: Pre-check for downloads and dependencies folders
-if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
-if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
-
-:: Set download URL for the sccache
-set DOWNLOAD_URL="https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-pc-windows-msvc.zip"
-set INSTALLER_FILE=%DOWNLOADS_DIR%\sccache.zip
-
-:: Download installer
-echo Downloading sccache.zip...
-curl -L -o "%INSTALLER_FILE%" %DOWNLOAD_URL%
-
-:: Install sccache
-echo Extracting sccache.zip...
-tar -xf "%INSTALLER_FILE%" -C %DEPENDENCIES_DIR%
-cd %DEPENDENCIES_DIR%
-ren sccache-v0.8.1-x86_64-pc-windows-msvc sccache
-cd ..
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed to install sccache. (exitcode = %errorlevel%)"
-    exit /b 1
-)
-
-:: Add to PATH
-echo %DEPENDENCIES_DIR%\sccache\>> %GITHUB_PATH%
-
-echo Dependency sccache installation finished.

.ci/pytorch/windows/arm64/bootstrap_tests.bat

@@ -1,22 +0,0 @@
-:: change to source directory
-cd %PYTORCH_ROOT%
-
-:: activate visual studio
-call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
-where cl.exe
-
-:: create virtual environment
-python -m venv .venv
-echo * > .venv\.gitignore
-call .\.venv\Scripts\activate
-where python
-
-:: install dependencies
-python -m pip install --upgrade pip
-pip install -r requirements.txt
-pip install pytest numpy
-
-:: find file name for pytorch wheel
-for /f "delims=" %%f in ('dir /b "%PYTORCH_FINAL_PACKAGE_DIR%" ^| findstr "torch-"') do set "TORCH_WHEEL_FILENAME=%PYTORCH_FINAL_PACKAGE_DIR%\%%f"
-
-pip install %TORCH_WHEEL_FILENAME%

.ci/pytorch/windows/arm64/build_libtorch.bat

@@ -1,101 +0,0 @@
-@echo on
-
-:: environment variables
-set CMAKE_BUILD_TYPE=%BUILD_TYPE%
-set CMAKE_C_COMPILER_LAUNCHER=sccache
-set CMAKE_CXX_COMPILER_LAUNCHER=sccache
-set libuv_ROOT=%DEPENDENCIES_DIR%\libuv\install
-set MSSdk=1
-if defined PYTORCH_BUILD_VERSION (
-  set PYTORCH_BUILD_VERSION=%PYTORCH_BUILD_VERSION%
-  set PYTORCH_BUILD_NUMBER=1
-)
-
-:: Set BLAS type
-if %ENABLE_APL% == 1 (
-    set BLAS=APL
-    set USE_LAPACK=1
-) else if %ENABLE_OPENBLAS% == 1 (
-    set BLAS=OpenBLAS
-    set OpenBLAS_HOME=%DEPENDENCIES_DIR%\OpenBLAS\install
-)
-
-:: activate visual studio
-call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
-where cl.exe
-
-:: change to source directory
-cd %PYTORCH_ROOT%
-
-:: copy libuv.dll
-copy %libuv_ROOT%\lib\Release\uv.dll torch\lib\uv.dll
-
-:: create virtual environment
-python -m venv .venv
-echo * > .venv\.gitignore
-call .\.venv\Scripts\activate
-where python
-
-:: python install dependencies
-python -m pip install --upgrade pip
-pip install -r requirements.txt
-:: DISTUTILS_USE_SDK should be set after psutil dependency
-set DISTUTILS_USE_SDK=1
-
-:: start sccache server and reset sccache stats
-sccache --start-server
-sccache --zero-stats
-sccache --show-stats
- 
-:: Prepare the environment
-mkdir libtorch
-mkdir libtorch\bin
-mkdir libtorch\cmake
-mkdir libtorch\include
-mkdir libtorch\lib
-mkdir libtorch\share
-mkdir libtorch\test
-
-:: Call LibTorch build script
-python ./tools/build_libtorch.py
-
-:: Check if there is an error
-IF ERRORLEVEL 1 exit /b 1
-IF NOT ERRORLEVEL 0 exit /b 1
- 
-:: Move the files to the correct location
-move /Y torch\bin\*.* libtorch\bin\
-move /Y torch\cmake\*.* libtorch\cmake\
-robocopy /move /e torch\include\ libtorch\include\
-move /Y torch\lib\*.* libtorch\lib\
-robocopy /move /e torch\share\ libtorch\share\
-move /Y torch\test\*.* libtorch\test\
-move /Y libtorch\bin\*.dll libtorch\lib\
-
-:: Set version
-echo %PYTORCH_BUILD_VERSION% > libtorch\build-version
-git rev-parse HEAD > libtorch\build-hash
-
-:: Set LIBTORCH_PREFIX
-IF "%DEBUG%" == "" (
-    set LIBTORCH_PREFIX=libtorch-win-arm64-shared-with-deps
-) ELSE (
-    set LIBTORCH_PREFIX=libtorch-win-arm64-shared-with-deps-debug
-)
-
-:: Create output
-C:\Windows\System32\tar.exe -cvaf %LIBTORCH_PREFIX%-%PYTORCH_BUILD_VERSION%.zip -C libtorch *
-
-:: Copy output to target directory
-if not exist ..\output mkdir ..\output
-copy /Y "%LIBTORCH_PREFIX%-%PYTORCH_BUILD_VERSION%.zip" "%PYTORCH_FINAL_PACKAGE_DIR%\"
-copy /Y "%LIBTORCH_PREFIX%-%PYTORCH_BUILD_VERSION%.zip" "%PYTORCH_FINAL_PACKAGE_DIR%\%LIBTORCH_PREFIX%-latest.zip"
-
-:: Cleanup raw data to save space
-rmdir /s /q libtorch
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed on build_libtorch. (exitcode = %errorlevel%)"
-    exit /b 1
-)

.ci/pytorch/windows/arm64/build_pytorch.bat

@@ -1,60 +0,0 @@
-@echo on
-
-:: environment variables
-set CMAKE_BUILD_TYPE=%BUILD_TYPE%
-set CMAKE_C_COMPILER_LAUNCHER=sccache
-set CMAKE_CXX_COMPILER_LAUNCHER=sccache
-set libuv_ROOT=%DEPENDENCIES_DIR%\libuv\install
-set MSSdk=1
-if defined PYTORCH_BUILD_VERSION (
-  set PYTORCH_BUILD_VERSION=%PYTORCH_BUILD_VERSION%
-  set PYTORCH_BUILD_NUMBER=1
-)
-
-:: Set BLAS type
-if %ENABLE_APL% == 1 (
-    set BLAS=APL
-    set USE_LAPACK=1
-) else if %ENABLE_OPENBLAS% == 1 (
-    set BLAS=OpenBLAS
-    set OpenBLAS_HOME=%DEPENDENCIES_DIR%\OpenBLAS\install
-)
-
-:: activate visual studio
-call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
-where cl.exe
-
-:: change to source directory
-cd %PYTORCH_ROOT%
-
-:: copy libuv.dll
-copy %libuv_ROOT%\lib\Release\uv.dll torch\lib\uv.dll
-
-:: create virtual environment
-python -m venv .venv
-echo * > .venv\.gitignore
-call .\.venv\Scripts\activate
-where python
-
-:: python install dependencies
-python -m pip install --upgrade pip
-pip install -r requirements.txt
-:: DISTUTILS_USE_SDK should be set after psutil dependency
-set DISTUTILS_USE_SDK=1
-
-:: start sccache server and reset sccache stats
-sccache --start-server
-sccache --zero-stats
-sccache --show-stats
-
-:: Call PyTorch build script
-python setup.py bdist_wheel -d "%PYTORCH_FINAL_PACKAGE_DIR%"
-
-:: show sccache stats
-sccache --show-stats
-
-:: Check if installation was successful
-if %errorlevel% neq 0 (
-    echo "Failed on build_pytorch. (exitcode = %errorlevel%)"
-    exit /b 1
-)

.ci/pytorch/windows/arm64/smoke_test.bat

@@ -1,65 +0,0 @@
-@echo off
-setlocal
-
-set "ORIG_PATH=%PATH%"
-
-if "%PACKAGE_TYPE%" == "wheel" goto wheel
-if "%PACKAGE_TYPE%" == "libtorch" goto libtorch
-
-echo "unknown package type"
-exit /b 1
-
-:wheel
-echo "install wheel package"
-
-echo Running pip install...
-pip install -q --pre numpy protobuf
-echo Error level after pip install: %ERRORLEVEL%
-if errorlevel 1 exit /b 1
-
-for /F "delims=" %%i in ('where /R "%PYTORCH_FINAL_PACKAGE_DIR:/=\%" *.whl') do pip install "%%i"
-if errorlevel 1 exit /b 1
-
-goto smoke_test
-
-:smoke_test
-python -c "import torch"
-if ERRORLEVEL 1 exit /b 1
-
-echo Running python rnn_smoke.py...
-python %PYTORCH_ROOT%\.ci\pytorch\test_example_code\rnn_smoke_win_arm64.py
-if errorlevel 1 exit /b 1
-
-echo Checking that basic CNN works...
-python %PYTORCH_ROOT%\.ci\pytorch\test_example_code\cnn_smoke_win_arm64.py
-if errorlevel 1 exit /b 1
-
-goto end
-
-:libtorch
-echo "install and test libtorch"
-
-for /F "delims=" %%i in ('where /R "%PYTORCH_FINAL_PACKAGE_DIR:/=\%" *-latest.zip') do tar -xf "%%i" -C tmp
-if ERRORLEVEL 1 exit /b 1
-
-pushd tmp\libtorch
-
-set VC_VERSION_LOWER=14
-set VC_VERSION_UPPER=36
-
-call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
-
-set install_root=%CD%
-set INCLUDE=%INCLUDE%;%install_root%\include;%install_root%\include\torch\csrc\api\include
-set LIB=%LIB%;%install_root%\lib
-set PATH=%PATH%;%install_root%\lib
-
-cl %PYTORCH_ROOT%\.ci\pytorch\test_example_code\simple-torch-test.cpp c10.lib torch_cpu.lib /EHsc /std:c++17
-if ERRORLEVEL 1 exit /b 1
-
-.\simple-torch-test.exe
-if ERRORLEVEL 1 exit /b 1
-
-:end
-set "PATH=%ORIG_PATH%"
-popd

.ci/pytorch/windows/condaenv.bat

@@ -9,13 +9,12 @@ FOR %%v IN (%DESIRED_PYTHON%) DO (
     set PYTHON_VERSION_STR=%%v
     set PYTHON_VERSION_STR=!PYTHON_VERSION_STR:.=!
     conda remove -n py!PYTHON_VERSION_STR! --all -y || rmdir %CONDA_HOME%\envs\py!PYTHON_VERSION_STR! /s
-    if "%%v" == "3.9" call conda create -n py!PYTHON_VERSION_STR! -y numpy=2.0.1 boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
-    if "%%v" == "3.10" call conda create -n py!PYTHON_VERSION_STR! -y -c=conda-forge numpy=2.0.1  boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
-    if "%%v" == "3.11" call conda create -n py!PYTHON_VERSION_STR! -y -c=conda-forge numpy=2.0.1  boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
-    if "%%v" == "3.12" call conda create -n py!PYTHON_VERSION_STR! -y -c=conda-forge numpy=2.0.1  boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
-    if "%%v" == "3.13" call conda create -n py!PYTHON_VERSION_STR! -y -c=conda-forge numpy=2.1.2  boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
-    if "%%v" == "3.13t" call conda create -n py!PYTHON_VERSION_STR! -y -c=conda-forge numpy=2.1.2 boto3 cmake ninja typing_extensions setuptools=72.1.0 python-freethreading python=3.13
-    call conda run -n py!PYTHON_VERSION_STR! pip install pyyaml
+    if "%%v" == "3.8" call conda create -n py!PYTHON_VERSION_STR! -y -q numpy=1.11 pyyaml boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
+    if "%%v" == "3.9" call conda create -n py!PYTHON_VERSION_STR! -y -q numpy=2.0.1 pyyaml boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
+    if "%%v" == "3.10" call conda create -n py!PYTHON_VERSION_STR! -y -q -c=conda-forge numpy=2.0.1 pyyaml boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
+    if "%%v" == "3.11" call conda create -n py!PYTHON_VERSION_STR! -y -q -c=conda-forge numpy=2.0.1 pyyaml boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
+    if "%%v" == "3.12" call conda create -n py!PYTHON_VERSION_STR! -y -q -c=conda-forge numpy=2.0.1 pyyaml boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
+    if "%%v" == "3.13" call conda create -n py!PYTHON_VERSION_STR! -y -q -c=conda-forge numpy=2.1.2 pyyaml boto3 cmake ninja typing_extensions setuptools=72.1.0 python=%%v
     call conda run -n py!PYTHON_VERSION_STR! pip install mkl-include
     call conda run -n py!PYTHON_VERSION_STR! pip install mkl-static
 )

.ci/pytorch/windows/cuda128.bat

@@ -1,59 +0,0 @@
-@echo off
-
-set MODULE_NAME=pytorch
-
-IF NOT EXIST "setup.py" IF NOT EXIST "%MODULE_NAME%" (
-    call internal\clone.bat
-    cd %~dp0
-) ELSE (
-    call internal\clean.bat
-)
-IF ERRORLEVEL 1 goto :eof
-
-call internal\check_deps.bat
-IF ERRORLEVEL 1 goto :eof
-
-REM Check for optional components
-
-set USE_CUDA=
-set CMAKE_GENERATOR=Visual Studio 15 2017 Win64
-
-IF "%NVTOOLSEXT_PATH%"=="" (
-    IF EXIST "C:\Program Files\NVIDIA Corporation\NvToolsExt\lib\x64\nvToolsExt64_1.lib"  (
-        set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
-    ) ELSE (
-        echo NVTX ^(Visual Studio Extension ^for CUDA^) ^not installed, failing
-        exit /b 1
-    )
-)
-
-IF "%CUDA_PATH_V128%"=="" (
-    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\bin\nvcc.exe" (
-        set "CUDA_PATH_V128=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8"
-    ) ELSE (
-        echo CUDA 12.8 not found, failing
-        exit /b 1
-    )
-)
-
-IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=5.0;6.0;6.1;7.0;7.5;8.0;8.6;9.0;10.0;12.0
-    set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
-) ELSE (
-    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90 -gencode=arch=compute_100,code=compute_100 -gencode=arch=compute_120,code=compute_120
-)
-
-set "CUDA_PATH=%CUDA_PATH_V128%"
-set "PATH=%CUDA_PATH_V128%\bin;%PATH%"
-
-:optcheck
-
-call internal\check_opts.bat
-IF ERRORLEVEL 1 goto :eof
-
-if exist "%NIGHTLIES_PYTORCH_ROOT%" cd %NIGHTLIES_PYTORCH_ROOT%\..
-call  %~dp0\internal\copy.bat
-IF ERRORLEVEL 1 goto :eof
-
-call  %~dp0\internal\setup.bat
-IF ERRORLEVEL 1 goto :eof

.ci/pytorch/windows/internal/cuda_install.bat

@@ -9,8 +9,7 @@ if "%CUDA_VERSION%" == "xpu" (
     exit /b 0
 )
 
-set SRC_DIR=%~dp0\..
-
+set SRC_DIR=%NIGHTLIES_PYTORCH_ROOT%
 if not exist "%SRC_DIR%\temp_build" mkdir "%SRC_DIR%\temp_build"
 
 set /a CUDA_VER=%CUDA_VERSION%
@@ -24,9 +23,9 @@ set CUDNN_LIB_FOLDER="lib\x64"
 if exist "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v%CUDA_VERSION_STR%\bin\nvcc.exe" goto set_cuda_env_vars
 
 if %CUDA_VER% EQU 118 goto cuda118
+if %CUDA_VER% EQU 121 goto cuda121
 if %CUDA_VER% EQU 124 goto cuda124
 if %CUDA_VER% EQU 126 goto cuda126
-if %CUDA_VER% EQU 128 goto cuda128
 
 echo CUDA %CUDA_VERSION_STR% is not supported
 exit /b 1
@@ -112,33 +111,6 @@ xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
 
 goto cuda_common
 
-:cuda128
-
-set CUDA_INSTALL_EXE=cuda_12.8.0_571.96_windows.exe
-if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
-    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
-    if errorlevel 1 exit /b 1
-    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
-    set "ARGS=cuda_profiler_api_12.8 thrust_12.8 nvcc_12.8 cuobjdump_12.8 nvprune_12.8 nvprof_12.8 cupti_12.8 cublas_12.8 cublas_dev_12.8 cudart_12.8 cufft_12.8 cufft_dev_12.8 curand_12.8 curand_dev_12.8 cusolver_12.8 cusolver_dev_12.8 cusparse_12.8 cusparse_dev_12.8 npp_12.8 npp_dev_12.8 nvrtc_12.8 nvrtc_dev_12.8 nvml_dev_12.8 nvjitlink_12.8 nvtx_12.8"
-)
-
-set CUDNN_FOLDER=cudnn-windows-x86_64-9.7.0.66_cuda12-archive
-set CUDNN_LIB_FOLDER="lib"
-set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
-if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
-    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
-    if errorlevel 1 exit /b 1
-    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
-)
-
-@REM cuDNN 8.3+ required zlib to be installed on the path
-echo Installing ZLIB dlls
-curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
-7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
-xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
-
-goto cuda_common
-
 :cuda_common
 :: NOTE: We only install CUDA if we don't have it installed already.
 :: With GHA runners these should be pre-installed as part of our AMI process

.ci/pytorch/windows/internal/smoke_test.bat

@@ -27,6 +27,7 @@ for /F "delims=" %%i in ('wmic path win32_VideoController get name') do (
 endlocal & set NVIDIA_GPU_EXISTS=%NVIDIA_GPU_EXISTS%
 
 if "%PACKAGE_TYPE%" == "wheel" goto wheel
+if "%PACKAGE_TYPE%" == "conda" goto conda
 if "%PACKAGE_TYPE%" == "libtorch" goto libtorch
 
 echo "unknown package type"
@@ -36,7 +37,6 @@ exit /b 1
 echo "install wheel package"
 
 set PYTHON_INSTALLER_URL=
-if "%DESIRED_PYTHON%" == "3.13t" set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.13.0/python-3.13.0-amd64.exe"
 if "%DESIRED_PYTHON%" == "3.13" set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.13.0/python-3.13.0-amd64.exe"
 if "%DESIRED_PYTHON%" == "3.12" set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.12.0/python-3.12.0-amd64.exe"
 if "%DESIRED_PYTHON%" == "3.11" set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.11.0/python-3.11.0-amd64.exe"
@@ -47,13 +47,6 @@ if "%PYTHON_INSTALLER_URL%" == "" (
     echo Python %DESIRED_PYTHON% not supported yet
 )
 
-set ADDITIONAL_OPTIONS=""
-set PYTHON_EXEC="python"
-if "%DESIRED_PYTHON%" == "3.13t" (
-    set ADDITIONAL_OPTIONS="Include_freethreaded=1"
-    set PYTHON_EXEC="python3.13t"
-)
-
 del python-amd64.exe
 curl --retry 3 -kL "%PYTHON_INSTALLER_URL%" --output python-amd64.exe
 if errorlevel 1 exit /b 1
@@ -62,30 +55,85 @@ if errorlevel 1 exit /b 1
 :: the installed Python to PATH system-wide. Even calling set PATH=%ORIG_PATH% later on won't make
 :: a change. As the builder directory will be removed after the smoke test, all subsequent non-binary
 :: jobs will fail to find any Python executable there
-start /wait "" python-amd64.exe /quiet InstallAllUsers=1 PrependPath=0 Include_test=0 %ADDITIONAL_OPTIONS% TargetDir=%CD%\Python
+start /wait "" python-amd64.exe /quiet InstallAllUsers=1 PrependPath=0 Include_test=0 TargetDir=%CD%\Python
 if errorlevel 1 exit /b 1
 
 set "PATH=%CD%\Python%PYTHON_VERSION%\Scripts;%CD%\Python;%PATH%"
-if "%DESIRED_PYTHON%" == "3.13t" %PYTHON_EXEC% -m pip install --pre numpy==2.2.1 protobuf
-if "%DESIRED_PYTHON%" == "3.13" %PYTHON_EXEC% -m pip install --pre numpy==2.1.2 protobuf
-if "%DESIRED_PYTHON%" == "3.12" %PYTHON_EXEC% -m pip install --pre numpy==2.0.2 protobuf
-if "%DESIRED_PYTHON%" == "3.11" %PYTHON_EXEC% -m pip install --pre numpy==2.0.2 protobuf
-if "%DESIRED_PYTHON%" == "3.10" %PYTHON_EXEC% -m pip install --pre numpy==2.0.2 protobuf
-if "%DESIRED_PYTHON%" == "3.9" %PYTHON_EXEC% -m pip install --pre numpy==2.0.2 protobuf
+
+if "%DESIRED_PYTHON%" == "3.13" pip install -q --pre numpy==2.1.0 protobuf
+if "%DESIRED_PYTHON%" == "3.12" pip install -q --pre numpy==2.0.2 protobuf
+if "%DESIRED_PYTHON%" == "3.11" pip install -q --pre numpy==2.0.2 protobuf
+if "%DESIRED_PYTHON%" == "3.10" pip install -q --pre numpy==2.0.2 protobuf
+if "%DESIRED_PYTHON%" == "3.9" pip install -q --pre numpy==2.0.2 protobuf
+if "%DESIRED_PYTHON%" == "3.8" pip install -q numpy protobuf
 
 if errorlevel 1 exit /b 1
 
-for /F "delims=" %%i in ('where /R "%PYTORCH_FINAL_PACKAGE_DIR:/=\%" *.whl') do %PYTHON_EXEC% -m pip install "%%i"
+for /F "delims=" %%i in ('where /R "%PYTORCH_FINAL_PACKAGE_DIR:/=\%" *.whl') do pip install "%%i"
 if errorlevel 1 exit /b 1
 
 goto smoke_test
 
+:conda
+echo "install conda package"
+
+:: Install Miniconda3
+set "CONDA_HOME=%CD%\conda"
+set "tmp_conda=%CONDA_HOME%"
+set "miniconda_exe=%CD%\miniconda.exe"
+set "CONDA_EXTRA_ARGS=cpuonly -c pytorch-nightly"
+if "%CUDA_VERSION%" == "118" (
+    set "CONDA_EXTRA_ARGS=pytorch-cuda=11.8 -c nvidia -c pytorch-nightly"
+)
+if "%CUDA_VERSION%" == "121" (
+    set "CONDA_EXTRA_ARGS=pytorch-cuda=12.1 -c nvidia -c pytorch-nightly"
+)
+if "%CUDA_VERSION%" == "124" (
+    set "CONDA_EXTRA_ARGS=pytorch-cuda=12.4 -c nvidia -c pytorch-nightly"
+)
+if "%CUDA_VERSION%" == "126" (
+    set "CONDA_EXTRA_ARGS=pytorch-cuda=12.6 -c nvidia -c pytorch-nightly"
+)
+
+rmdir /s /q conda
+del miniconda.exe
+curl -k https://repo.anaconda.com/miniconda/Miniconda3-latest-Windows-x86_64.exe -o "%miniconda_exe%"
+start /wait "" "%miniconda_exe%" /S /InstallationType=JustMe /RegisterPython=0 /AddToPath=0 /D=%tmp_conda%
+if ERRORLEVEL 1 exit /b 1
+
+set "PATH=%CONDA_HOME%;%CONDA_HOME%\scripts;%CONDA_HOME%\Library\bin;%PATH%"
+
+conda create -qyn testenv python=%DESIRED_PYTHON%
+if errorlevel 1 exit /b 1
+call conda install -yq conda-build
+if errorlevel 1 exit /b 1
+call %CONDA_HOME%\condabin\activate.bat testenv
+if errorlevel 1 exit /b 1
+set "NO_ARCH_PATH=%PYTORCH_FINAL_PACKAGE_DIR:/=\%\noarch"
+mkdir %NO_ARCH_PATH%
+for /F "delims=" %%i in ('where /R "%PYTORCH_FINAL_PACKAGE_DIR:/=\%" *') do xcopy "%%i" %NO_ARCH_PATH% /Y
+if ERRORLEVEL 1 exit /b 1
+call conda index %PYTORCH_FINAL_PACKAGE_DIR%
+if errorlevel 1 exit /b 1
+call conda install -yq -c "file:///%PYTORCH_FINAL_PACKAGE_DIR%" pytorch==%PYTORCH_BUILD_VERSION% -c pytorch -c numba/label/dev -c nvidia
+if ERRORLEVEL 1 exit /b 1
+call conda install -yq numpy
+if ERRORLEVEL 1 exit /b 1
+
+set /a CUDA_VER=%CUDA_VERSION%
+set CUDA_VER_MAJOR=%CUDA_VERSION:~0,-1%
+set CUDA_VER_MINOR=%CUDA_VERSION:~-1,1%
+set CUDA_VERSION_STR=%CUDA_VER_MAJOR%.%CUDA_VER_MINOR%
+
+:: Install package we just build
+
+
 :smoke_test
-%PYTHON_EXEC% -c "import torch"
+python -c "import torch"
 if ERRORLEVEL 1 exit /b 1
 
 echo Checking that MKL is available
-%PYTHON_EXEC% -c "import torch; exit(0 if torch.backends.mkl.is_available() else 1)"
+python -c "import torch; exit(0 if torch.backends.mkl.is_available() else 1)"
 if ERRORLEVEL 1 exit /b 1
 
 if "%NVIDIA_GPU_EXISTS%" == "0" (
@@ -94,24 +142,24 @@ if "%NVIDIA_GPU_EXISTS%" == "0" (
 )
 
 echo Checking that CUDA archs are setup correctly
-%PYTHON_EXEC% -c "import torch; torch.randn([3,5]).cuda()"
+python -c "import torch; torch.randn([3,5]).cuda()"
 if ERRORLEVEL 1 exit /b 1
 
 echo Checking that magma is available
-%PYTHON_EXEC% -c "import torch; torch.rand(1).cuda(); exit(0 if torch.cuda.has_magma else 1)"
+python -c "import torch; torch.rand(1).cuda(); exit(0 if torch.cuda.has_magma else 1)"
 if ERRORLEVEL 1 exit /b 1
 
 echo Checking that CuDNN is available
-%PYTHON_EXEC% -c "import torch; exit(0 if torch.backends.cudnn.is_available() else 1)"
+python -c "import torch; exit(0 if torch.backends.cudnn.is_available() else 1)"
 if ERRORLEVEL 1 exit /b 1
 
 echo Checking that basic RNN works
-%PYTHON_EXEC% %PYTORCH_ROOT%\.ci\pytorch\test_example_code\rnn_smoke.py
+python %PYTORCH_ROOT%\.ci\pytorch\test_example_code\rnn_smoke.py
 
 if ERRORLEVEL 1 exit /b 1
 
 echo Checking that basic CNN works
-%PYTHON_EXEC% %PYTORCH_ROOT%\.ci\pytorch\test_example_code\cnn_smoke.py
+python %PYTORCH_ROOT%\.ci\pytorch\test_example_code\cnn_smoke.py
 if ERRORLEVEL 1 exit /b 1
 
 goto end
@@ -119,6 +167,7 @@ goto end
 :libtorch
 echo "install and test libtorch"
 
+if "%VC_YEAR%" == "2019" powershell internal\vs2019_install.ps1
 if "%VC_YEAR%" == "2022" powershell internal\vs2022_install.ps1
 
 if ERRORLEVEL 1 exit /b 1
@@ -130,6 +179,10 @@ pushd tmp\libtorch
 
 set VC_VERSION_LOWER=17
 set VC_VERSION_UPPER=18
+IF "%VC_YEAR%" == "2019" (
+    set VC_VERSION_LOWER=16
+    set VC_VERSION_UPPER=17
+)
 
 for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe" -legacy -products * -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
     if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (

.ci/pytorch/windows/internal/static_lib_test.bat

@@ -70,6 +70,7 @@ echo "install and test libtorch"
 pip install cmake
 echo "installing cmake"
 
+if "%VC_YEAR%" == "2019" powershell internal\vs2019_install.ps1
 if "%VC_YEAR%" == "2022" powershell internal\vs2022_install.ps1
 
 if ERRORLEVEL 1 exit /b 1
@@ -82,6 +83,10 @@ pushd tmp\libtorch
 
 set VC_VERSION_LOWER=17
 set VC_VERSION_UPPER=18
+IF "%VC_YEAR%" == "2019" (
+    set VC_VERSION_LOWER=16
+    set VC_VERSION_UPPER=17
+)
 
 for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe" -legacy -products * -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
     if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (

.ci/pytorch/windows/internal/vc_install_helper.bat

@@ -1,8 +1,12 @@
+if "%VC_YEAR%" == "2019" powershell windows/internal/vs2019_install.ps1
 if "%VC_YEAR%" == "2022" powershell windows/internal/vs2022_install.ps1
 
 set VC_VERSION_LOWER=17
 set VC_VERSION_UPPER=18
-
+if "%VC_YEAR%" == "2019" (
+    set VC_VERSION_LOWER=16
+    set VC_VERSION_UPPER=17
+)
 
 for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"  -products Microsoft.VisualStudio.Product.BuildTools -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
     if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (

.ci/pytorch/windows/internal/vs2019_install.ps1

@@ -0,0 +1,48 @@
+# https://developercommunity.visualstudio.com/t/install-specific-version-of-vs-component/1142479
+# https://docs.microsoft.com/en-us/visualstudio/releases/2019/history#release-dates-and-build-numbers
+
+# 16.8.6 BuildTools
+$VS_DOWNLOAD_LINK = "https://ossci-windows.s3.us-east-1.amazonaws.com/vs16.8.6_BuildTools.exe"
+$COLLECT_DOWNLOAD_LINK = "https://aka.ms/vscollect.exe"
+$VS_INSTALL_ARGS = @("--nocache","--quiet","--wait", "--add Microsoft.VisualStudio.Workload.VCTools",
+                                                     "--add Microsoft.Component.MSBuild",
+                                                     "--add Microsoft.VisualStudio.Component.Roslyn.Compiler",
+                                                     "--add Microsoft.VisualStudio.Component.TextTemplating",
+                                                     "--add Microsoft.VisualStudio.Component.VC.CoreIde",
+                                                     "--add Microsoft.VisualStudio.Component.VC.Redist.14.Latest",
+                                                     "--add Microsoft.VisualStudio.ComponentGroup.NativeDesktop.Core",
+                                                     "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64",
+                                                     "--add Microsoft.VisualStudio.ComponentGroup.NativeDesktop.Win81")
+
+curl.exe --retry 3 -kL $VS_DOWNLOAD_LINK --output vs_installer.exe
+if ($LASTEXITCODE -ne 0) {
+    echo "Download of the VS 2019 Version 16.8.5 installer failed"
+    exit 1
+}
+
+if (Test-Path "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe") {
+    $existingPath = & "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe" -products "Microsoft.VisualStudio.Product.BuildTools" -version "[16, 17)" -property installationPath
+    if ($existingPath -ne $null) {
+        if (!${env:CIRCLECI}) {
+            echo "Found correctly versioned existing BuildTools installation in $existingPath"
+            exit 0
+        }
+        echo "Found existing BuildTools installation in $existingPath, keeping it"
+    }
+}
+
+$process = Start-Process "${PWD}\vs_installer.exe" -ArgumentList $VS_INSTALL_ARGS -NoNewWindow -Wait -PassThru
+Remove-Item -Path vs_installer.exe -Force
+$exitCode = $process.ExitCode
+if (($exitCode -ne 0) -and ($exitCode -ne 3010)) {
+    echo "VS 2019 installer exited with code $exitCode, which should be one of [0, 3010]."
+    curl.exe --retry 3 -kL $COLLECT_DOWNLOAD_LINK --output Collect.exe
+    if ($LASTEXITCODE -ne 0) {
+        echo "Download of the VS Collect tool failed."
+        exit 1
+    }
+    Start-Process "${PWD}\Collect.exe" -NoNewWindow -Wait -PassThru
+    New-Item -Path "C:\w\build-results" -ItemType "directory" -Force
+    Copy-Item -Path "C:\Users\${env:USERNAME}\AppData\Local\Temp\vslogs.zip" -Destination "C:\w\build-results\"
+    exit 1
+}

.ci/pytorch/windows/internal/xpu_install.bat

@@ -7,9 +7,6 @@ if not "%CUDA_VERSION%" == "xpu" (
     exit /b 0
 )
 
-set SRC_DIR=%NIGHTLIES_PYTORCH_ROOT%
-if not exist "%SRC_DIR%\temp_build" mkdir "%SRC_DIR%\temp_build"
-
 set XPU_INSTALL_MODE=%~1
 if "%XPU_INSTALL_MODE%"=="" goto xpu_bundle_install_start
 if "%XPU_INSTALL_MODE%"=="bundle" goto xpu_bundle_install_start
@@ -120,14 +117,3 @@ if errorlevel 1 exit /b 1
 del xpu_extra.exe
 
 :xpu_install_end
-
-if not "%XPU_ENABLE_KINETO%"=="1" goto install_end
-:: Install Level Zero SDK
-set XPU_EXTRA_LZ_URL=https://github.com/oneapi-src/level-zero/releases/download/v1.14.0/level-zero-sdk_1.14.0.zip
-curl -k -L %XPU_EXTRA_LZ_URL% --output "%SRC_DIR%\temp_build\level_zero_sdk.zip"
-echo "Installing level zero SDK..."
-7z x "%SRC_DIR%\temp_build\level_zero_sdk.zip" -o"%SRC_DIR%\temp_build\level_zero"
-set "INCLUDE=%SRC_DIR%\temp_build\level_zero\include;%INCLUDE%"
-del "%SRC_DIR%\temp_build\level_zero_sdk.zip"
-
-:install_end

.ci/pytorch/windows/xpu.bat

@@ -28,6 +28,11 @@ call "%XPU_BUNDLE_ROOT%\compiler\latest\env\vars.bat"
 call "%XPU_BUNDLE_ROOT%\ocloc\latest\env\vars.bat"
 IF ERRORLEVEL 1 goto :eof
 
+:: Workaround for https://github.com/pytorch/pytorch/issues/134989
+set CMAKE_SHARED_LINKER_FLAGS=/FORCE:MULTIPLE
+set CMAKE_MODULE_LINKER_FLAGS=/FORCE:MULTIPLE
+set CMAKE_EXE_LINKER_FLAGS=/FORCE:MULTIPLE
+
 if exist "%NIGHTLIES_PYTORCH_ROOT%" cd %NIGHTLIES_PYTORCH_ROOT%\..
 call %~dp0\internal\copy_cpu.bat
 IF ERRORLEVEL 1 goto :eof

.ci/wheel/build_wheel.sh

@@ -130,19 +130,7 @@ export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 SETUPTOOLS_PINNED_VERSION="=46.0.0"
 PYYAML_PINNED_VERSION="=5.3"
 EXTRA_CONDA_INSTALL_FLAGS=""
-CONDA_ENV_CREATE_FLAGS=""
-RENAME_WHEEL=true
 case $desired_python in
-    3.13t)
-        echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=68.0.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="=2.1.0"
-        CONDA_ENV_CREATE_FLAGS="python-freethreading"
-        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-        desired_python="3.13"
-        RENAME_WHEEL=false
-        ;;
     3.13)
         echo "Using 3.13 deps"
         SETUPTOOLS_PINNED_VERSION=">=68.0.0"
@@ -181,15 +169,18 @@ esac
 
 # Install into a fresh env
 tmp_env_name="wheel_py$python_nodot"
-conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
+conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python"
 source activate "$tmp_env_name"
 
-pip install "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing_extensions
-retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
-retry brew install libomp
+pip install -q "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests
+retry pip install -qr "${pytorch_rootdir}/requirements.txt" || true
+# TODO : Remove me later (but in the interim, use Anaconda cmake, to find Anaconda installed OpenMP)
+retry pip uninstall -y cmake
+retry conda install ${EXTRA_CONDA_INSTALL_FLAGS} -yq  llvm-openmp=14.0.6 cmake ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing_extensions
 
-# For USE_DISTRIBUTED=1 on macOS, need libuv, which is build as part of tensorpipe submodule
+# For USE_DISTRIBUTED=1 on macOS, need libuv and pkg-config to find libuv.
 export USE_DISTRIBUTED=1
+retry conda install ${EXTRA_CONDA_INSTALL_FLAGS} -yq libuv pkg-config
 
 if [[ -n "$CROSS_COMPILE_ARM64" ]]; then
     export CMAKE_OSX_ARCHITECTURES=arm64
@@ -231,13 +222,10 @@ echo "The wheel is in $(find $whl_tmp_dir -name '*.whl')"
 wheel_filename_gen=$(find $whl_tmp_dir -name '*.whl' | head -n1 | xargs -I {} basename {})
 popd
 
-if [[ -z "$BUILD_PYTHONLESS" && $RENAME_WHEEL == true  ]]; then
+if [[ -z "$BUILD_PYTHONLESS" ]]; then
     # Copy the whl to a final destination before tests are run
     echo "Renaming Wheel file: $wheel_filename_gen to $wheel_filename_new"
     cp "$whl_tmp_dir/$wheel_filename_gen" "$PYTORCH_FINAL_PACKAGE_DIR/$wheel_filename_new"
-elif [[ $RENAME_WHEEL == false ]]; then
-    echo "Copying Wheel file: $wheel_filename_gen to $PYTORCH_FINAL_PACKAGE_DIR"
-    cp "$whl_tmp_dir/$wheel_filename_gen" "$PYTORCH_FINAL_PACKAGE_DIR/$wheel_filename_gen"
 else
     pushd "$pytorch_rootdir"
 

.circleci/scripts/binary_populate_env.sh

@@ -30,7 +30,9 @@ fi
 # Pick docker image
 export DOCKER_IMAGE=${DOCKER_IMAGE:-}
 if [[ -z "$DOCKER_IMAGE" ]]; then
-  if [[ "$DESIRED_CUDA" == cpu ]]; then
+  if [[ "$PACKAGE_TYPE" == conda ]]; then
+    export DOCKER_IMAGE="pytorch/conda-cuda"
+  elif [[ "$DESIRED_CUDA" == cpu ]]; then
     export DOCKER_IMAGE="pytorch/manylinux:cpu"
   else
     export DOCKER_IMAGE="pytorch/manylinux-builder:${DESIRED_CUDA:2}"
@@ -61,7 +63,7 @@ if tagged_version >/dev/null; then
   # Turns tag v1.6.0-rc1 -> v1.6.0
   BASE_BUILD_VERSION="$(tagged_version | sed -e 's/^v//' -e 's/-.*$//')"
 fi
-if [[ "$(uname)" == 'Darwin' ]]; then
+if [[ "$(uname)" == 'Darwin' ]] || [[ "$PACKAGE_TYPE" == conda ]]; then
   export PYTORCH_BUILD_VERSION="${BASE_BUILD_VERSION}"
 else
   export PYTORCH_BUILD_VERSION="${BASE_BUILD_VERSION}+$DESIRED_CUDA"
@@ -147,6 +149,8 @@ export PYTORCH_EXTRA_INSTALL_REQUIREMENTS="${PYTORCH_EXTRA_INSTALL_REQUIREMENTS:
 
 # TODO: We don't need this anymore IIUC
 export TORCH_PACKAGE_NAME='torch'
+export TORCH_CONDA_BUILD_FOLDER='pytorch-nightly'
+export ANACONDA_USER='pytorch'
 
 export USE_FBGEMM=1
 export PIP_UPLOAD_FOLDER="$PIP_UPLOAD_FOLDER"

.circleci/scripts/binary_upload.sh

@@ -2,7 +2,7 @@
 
 set -euo pipefail
 
-PACKAGE_TYPE=${PACKAGE_TYPE:-wheel}
+PACKAGE_TYPE=${PACKAGE_TYPE:-conda}
 
 PKG_DIR=${PKG_DIR:-/tmp/workspace/final_pkgs}
 
@@ -18,8 +18,10 @@ BUILD_NAME=${BUILD_NAME:-}
 
 DRY_RUN=${DRY_RUN:-enabled}
 # Don't actually do work unless explicit
+ANACONDA="true anaconda"
 AWS_S3_CP="aws s3 cp --dryrun"
 if [[ "${DRY_RUN}" = "disabled" ]]; then
+  ANACONDA="anaconda"
   AWS_S3_CP="aws s3 cp"
 fi
 
@@ -32,6 +34,10 @@ if [[ ${BUILD_NAME} == *-full* ]]; then
   UPLOAD_SUBFOLDER="${UPLOAD_SUBFOLDER}_full"
 fi
 
+# Sleep 2 minutes between retries for conda upload
+retry () {
+  "$@"  || (sleep 5m && "$@") || (sleep 5m && "$@") || (sleep 5m && "$@") || (sleep 5m && "$@")
+}
 
 do_backup() {
   local backup_dir
@@ -43,6 +49,20 @@ do_backup() {
   )
 }
 
+conda_upload() {
+  (
+    set -x
+    retry \
+    ${ANACONDA} \
+    upload  \
+    ${PKG_DIR}/*.tar.bz2 \
+    -u "pytorch-${UPLOAD_CHANNEL}" \
+    --label main \
+    --no-progress \
+    --force
+  )
+}
+
 s3_upload() {
   local extension
   local pkg_type
@@ -58,18 +78,31 @@ s3_upload() {
     for pkg in ${PKG_DIR}/*.${extension}; do
       (
         set -x
-        shm_id=$(sha256sum "${pkg}" | awk '{print $1}')
-        ${AWS_S3_CP} --no-progress --acl public-read "${pkg}" "${s3_upload_dir}" \
-          --metadata "checksum-sha256=${shm_id}"
+        ${AWS_S3_CP} --no-progress --acl public-read "${pkg}" "${s3_upload_dir}"
       )
     done
   )
 }
 
 # Install dependencies (should be a no-op if previously installed)
-pip install -q awscli uv
+conda install -yq anaconda-client
+pip install -q awscli
 
 case "${PACKAGE_TYPE}" in
+  conda)
+    conda_upload
+    for conda_archive in ${PKG_DIR}/*.tar.bz2; do
+      # Fetch  platform (eg. win-64, linux-64, etc.) from index file because
+      # there's no actual conda command to read this
+      subdir=$(\
+        tar -xOf "${conda_archive}" info/index.json \
+          | grep subdir  \
+          | cut -d ':' -f2 \
+          | sed -e 's/[[:space:]]//' -e 's/"//g' -e 's/,//' \
+      )
+      BACKUP_DIR="conda/${subdir}"
+    done
+    ;;
   libtorch)
     s3_upload "zip" "libtorch"
     BACKUP_DIR="libtorch/${UPLOAD_CHANNEL}/${UPLOAD_SUBFOLDER}"

.circleci/scripts/binary_windows_build.sh

@@ -8,9 +8,10 @@ export CUDA_VERSION="${DESIRED_CUDA/cu/}"
 export USE_SCCACHE=1
 export SCCACHE_BUCKET=ossci-compiler-cache
 export SCCACHE_IGNORE_SERVER_IO_ERROR=1
-export VC_YEAR=2022
+export VC_YEAR=2019
 
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
+    export VC_YEAR=2022
     export USE_SCCACHE=0
     export XPU_VERSION=2025.0
 fi

.circleci/scripts/binary_windows_test.sh

@@ -4,9 +4,10 @@ set -eux -o pipefail
 source "${BINARY_ENV_FILE:-/c/w/env}"
 
 export CUDA_VERSION="${DESIRED_CUDA/cu/}"
-export VC_YEAR=2022
+export VC_YEAR=2019
 
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
+    export VC_YEAR=2022
     export XPU_VERSION=2025.0
 fi
 

.clang-tidy

@@ -1,9 +1,8 @@
 ---
 # NOTE there must be no spaces before the '-', so put the comma last.
-# The check bugprone-unchecked-optional-access is also turned on.
-# Note that it can cause clang-tidy to hang randomly. The tracking issue
+# The check bugprone-unchecked-optional-access is also turned off atm
+# because it causes clang-tidy to hang randomly. The tracking issue
 # can be found at https://github.com/llvm/llvm-project/issues/69369.
-# When that happens, we can disable it on the problematic code by NOLINT.
 InheritParentConfig: true
 Checks: '
 bugprone-*,
@@ -13,10 +12,7 @@ bugprone-*,
 -bugprone-lambda-function-name,
 -bugprone-reserved-identifier,
 -bugprone-swapped-arguments,
-clang-analyzer-core.*,
-clang-analyzer-cplusplus.*,
-clang-analyzer-nullability.*,
-clang-analyzer-deadcode.*,
+-bugprone-unchecked-optional-access,
 clang-diagnostic-missing-prototypes,
 cppcoreguidelines-*,
 -cppcoreguidelines-avoid-do-while,

.git-blame-ignore-revs

@@ -24,10 +24,6 @@ e3900d2ba5c9f91a24a9ce34520794c8366d5c54
 2e26976ad3b06ce95dd6afccfdbe124802edf28f
 # 2021-06-07 Strictly typed everything in `.github` and `tools`
 737d920b21db9b4292d056ee1329945990656304
-# 2021-08-12 [codemod][lint][fbcode/c*] Enable BLACK by default
-b0043072529b81276a69df29e00555333117646c
-# 2021-08-25 Reformat run_test.py
-67d8e7b659b19e1ee68208b28bfa7dba73375dbc
 # 2022-06-09 Apply clang-format to ATen headers
 95b15c266baaf989ef7b6bbd7c23a2d90bacf687
 # 2022-06-11 [lint] autoformat test/cpp and torch/csrc
@@ -48,57 +44,3 @@ a53cda1ddc15336dc1ff0ce1eff2a49cdc5f882e
 d80939e5e9337e8078f11489afefec59fd42f93b
 # 2024-06-28 enable UFMT in `torch.utils.data`
 7cf0b90e49689d45be91aa539fdf54cf2ea8a9a3
-# 2024-07-03 Enable UFMT on test/test_public_bindings.py (#128389)
-fe5424d0f8604f6e66d827ae9f94b05cb7119d55
-# 2024-07-03 Enable UFMT on test/test_public_bindings.py (#128389)
-c686304277f7cd72331f685605325498cff94a0b
-# 2024-07-15 Enable UFMT on all of torch/sparse (#130545)
-535016967ae65a6027f83d6b935a985996223d49
-# 2024-07-15 [BE][Easy][1/19] enforce style for empty lines in import segments (#129752)
-a3abfa5cb57203b6a8ba7dff763f4057db8282a8
-# 2024-07-15 [BE][Easy][2/19] enforce style for empty lines in import segments in `.ci/` and `.github/` (#129753)
-ba48cf653541e9160dfdefa7bfea885c22e48608
-# 2024-07-16 [BE][Easy][5/19] enforce style for empty lines in import segments in `tools/` and `torchgen/` (#129756)
-f6838d521a243dbedc50ae96575720bf2cc8a8ad
-# 2024-07-17 [BE][Easy][9/19] enforce style for empty lines in import segments in `test/[e-h]*/` (#129760)
-76169cf69184bd462b9add40f893f57675f8a057
-# 2024-07-16 [BE][Easy][3/19] enforce style for empty lines in import segments in `benchmarks/` (#129754)
-c0ed38e644aed812d76b0ec85fae2f6019bf462b
-# 2024-07-16 [BE][Easy][4/19] enforce style for empty lines in import segments in `functorch/` (#129755)
-740fb229660f388feddc288c127ab12c82e67d36
-# 2024-07-17 [BE][Easy][12/19] enforce style for empty lines in import segments in `test/i*/` (#129763)
-aecc746fccc4495313167e3a7f94210daf457e1d
-# 2024-07-18 Revert "[BE][Easy][12/19] enforce style for empty lines in import segments in `test/i*/` (#129763)"
-b732b52f1e4378f8486ceb5e7026be3321c2651c
-# 2024-07-18 [BE][Easy][12/19] enforce style for empty lines in import segments in `test/i*/` (#129763)
-134bc4fc34bb02795aa694e66b132dcea5dde1e1
-# 2024-07-26 [BE][Easy][8/19] enforce style for empty lines in import segments in `test/[k-p]*/` (#129759)
-fbe6f42dcf1834213e0baa87b87529161df3c4d7
-# 2024-07-31 [BE][Easy][14/19] enforce style for empty lines in import segments in `torch/_[a-c]*/` and `torch/_[e-h]*/` and `torch/_[j-z]*/` (#129765)
-e7eeee473c6cb45942e87de5a616b0eb635513d6
-# 2024-07-31 Fix lint after PR #130572 (#132316)
-d72e863b3ecd3de4c8ea00518e110da964583f4f
-# 2024-07-31 [BE][Easy][15/19] enforce style for empty lines in import segments in `torch/_d*/` (#129767)
-e74ba1b34a476b46e76b4e32afe2d481f97e9a47
-# 2024-07-31 [BE][Easy][18/19] enforce style for empty lines in import segments in `torch/d*/` (#129770)
-b25ef91bf158ce459d8654e33c50c8e6ed8db716
-# 2024-07-20 [BE][Easy][13/19] enforce style for empty lines in import segments in `test/j*/` (#129764)
-6ff1e43a416c43cd82b210e22ac47384494c172e
-# 2024-11-01 [Lint] Clang-format all metal kernels (#139530)
-b3ad45733bd908b7358959ca1e1f8d026f4507eb
-# 2024-11-17 [BE][MPS] Apply clang-format to mps headers (#140906)
-99014a297c179862af38ee86bac2051434d3db41
-# 2024-11-27 Apply clang-format for ATen/core/boxing headers (#141105)
-19d01a1ef0c0d65768eb0a5c97a25328eec57fbd
-# 2024-12-05 fix the lint from D66795414 (#142122)
-65c2086d452ae6966ce9d7fb3cb2eef2fd0d2add
-# 2024-12-20 Apply clang-format for ATen/core/dispatch headers (#143620)
-cee06e74eeb54994b97000a02b715a4e63a97951
-# 2024-12-22 Better fix for f-strings in set_linter for py3.12 (#143725)
-eebc93d41eeffb936cbf20c9052e1e813d0cc052
-# 2025-01-04 [mps/BE] Fix linter warning/advice. (#144199)
-0dc1e6be192b260f1c072d70e1b06a3ac8e109fa
-# 2025-01-07 Fix lint in `test_provenance_tracing.py` (#144296)
-61c0a3d1cbaf6420e40ab0f9c9019daa21145e69
-# 2025-01-09 [BE] fix ruff rule E226: add missing whitespace around operator in f-strings (#144415)
-dcc3cf7066b4d8cab63ecb73daf1e36b01220a4e

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -5,7 +5,7 @@ body:
 - type: markdown
   attributes:
     value: >
-      #### Before submitting a bug, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/pytorch/pytorch/issues?q=is%3Aissue+sort%3Acreated-desc+). Note: Please write your bug report in English to ensure it can be understood and addressed by the development team. If you are filing a bug for torch.compile, please use the [torch.compile issue template](https://github.com/pytorch/pytorch/issues/new?q=sort%3Aupdated-desc+is%3Aissue+is%3Aopen&template=pt2-bug-report.yml).
+      #### Before submitting a bug, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/pytorch/pytorch/issues?q=is%3Aissue+sort%3Acreated-desc+).
 - type: textarea
   attributes:
     label: 🐛 Describe the bug

.github/ISSUE_TEMPLATE/disable-ci-jobs.md

@@ -5,7 +5,7 @@ title: "DISABLED [WORKFLOW_NAME] / [PLATFORM_NAME] / [JOB_NAME]"
 labels: "module: ci"
 ---
 
-> For example, DISABLED pull / win-vs2022-cpu-py3 / test (default). Once
+> For example, DISABLED pull / win-vs2019-cpu-py3 / test (default). Once
 > created, the job will be disabled within 15 minutes. You can check the
 > list of disabled jobs at https://ossci-metrics.s3.amazonaws.com/disabled-jobs.json
 

.github/ISSUE_TEMPLATE/documentation.yml

@@ -2,10 +2,6 @@ name: 📚 Documentation
 description: Report an issue related to https://pytorch.org/docs/stable/index.html
 
 body:
-- type: markdown
-  attributes:
-    value: >
-      #### Note: Please report your documentation issue in English to ensure it can be understood and addressed by the development team.
 - type: textarea
   attributes:
     label: 📚 The doc issue

.github/ISSUE_TEMPLATE/feature-request.yml

@@ -2,10 +2,6 @@ name: 🚀 Feature request
 description: Submit a proposal/request for a new PyTorch feature
 
 body:
-- type: markdown
-  attributes:
-    value: >
-      #### Note: Please write your feature request in English to ensure it can be understood and addressed by the development team.
 - type: textarea
   attributes:
     label: 🚀 The feature, motivation and pitch

.github/ISSUE_TEMPLATE/pt2-bug-report.yml

@@ -3,10 +3,6 @@ description: Create a report to help us reproduce and fix the bug
 labels: ["oncall: pt2"]
 
 body:
-  - type: markdown
-    attributes:
-      value: >
-        #### Note: Please write your bug report in English to ensure it can be understood and addressed by the development team.
   - type: markdown
     attributes:
       value: >
@@ -22,8 +18,6 @@ body:
 
         - If comparing eager and torch.compile at fp16/bf16, you should use fp32 as baseline
 
-        - Ensure rng state used to compare results is equivalent. Use `torch._inductor.config.fallback_random=True` and reset the torch rng seed between comparisons
-
         If the above requirements are met, add the label "topic: fuzzer" to your issue.
 
   - type: textarea

.github/actions/diskspace-cleanup/action.yml

@@ -17,10 +17,6 @@ runs:
         set -ex
         diskspace_cutoff=${{ inputs.diskspace-cutoff }}
         docker_root_dir=$(docker info -f '{{.DockerRootDir}}')
-        if [ ! -d "$docker_root_dir" ]; then
-            echo "Docker root directory ($docker_root_dir) does not exist. Skipping disk space check."
-            exit 0
-        fi
         diskspace=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
         msg="Please file an issue on pytorch/pytorch reporting the faulty runner. Include a link to the runner logs so the runner can be identified"
         if [[ "$diskspace" -ge "$diskspace_cutoff" ]] ; then

.github/actions/setup-rocm/action.yml

@@ -5,6 +5,20 @@ description: Set up ROCm host for CI
 runs:
   using: composite
   steps:
+    - name: Set DOCKER_HOST
+      shell: bash
+      run: echo "DOCKER_HOST=unix:///run/user/$(id -u)/docker.sock" >> "${GITHUB_ENV}"
+
+    - name: Remove leftover Docker config file
+      shell: bash
+      continue-on-error: true
+      run: |
+        set -ex
+
+        cat ~/.docker/config.json || true
+        # https://stackoverflow.com/questions/64455468/error-when-logging-into-ecr-with-docker-login-error-saving-credentials-not
+        rm -f ~/.docker/config.json
+
     - name: Stop all running docker containers
       if: always()
       shell: bash
@@ -24,12 +38,6 @@ runs:
         cat /opt/rocm/.info/version || true
         whoami
 
-    - name: Runner health check amdgpu info
-      if: always()
-      shell: bash
-      run: |
-        dpkg -l | grep -E "  amdgpu"
-
     - name: Runner health check rocm-smi
       if: always()
       shell: bash
@@ -97,16 +105,8 @@ runs:
       shell: bash
       run: |
         # All GPUs are visible to the runner; visibility, if needed, will be set by run_test.py.
-        # Add render group for container creation.
-        render_gid=`cat /etc/group | grep render | cut -d: -f3`
-        # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
-        if [ -f "/etc/podinfo/gha-render-devices" ]; then
-          DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-        else
-          DEVICE_FLAG="--device /dev/dri"
-        fi
         # The --group-add daemon and --group-add bin are needed in the Ubuntu 24.04 and Almalinux OSs respectively.
         # This is due to the device files (/dev/kfd & /dev/dri) being owned by video group on bare metal.
         # This video group ID maps to subgid 1 inside the docker image due to the /etc/subgid entries.
         # The group name corresponding to group ID 1 can change depending on the OS, so both are necessary.
-        echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd $DEVICE_FLAG --group-add video --group-add $render_gid --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host" >> "${GITHUB_ENV}"
+        echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon --group-add bin" >> "${GITHUB_ENV}"

.github/actions/upload-utilization-stats/action.yml

@@ -1,56 +0,0 @@
-name: upload-utilization-stats
-
-description: Upload utilization stats to artifacts
-
-inputs:
-    workflow_run_id:
-      type: string
-      description: 'workflow (run) id of the workflow the test is running'
-      required: True
-    workflow_attempt:
-      type: string
-      description: 'the workflow (run) attempt'
-      required: True
-    workflow_name:
-      description: 'name of the workflow'
-      type: string
-      required: True
-    job_id:
-      type: string
-      description: 'the job (run) id for the test'
-      required: True
-    job_name:
-      type: string
-      description: 'the job name of the test'
-      required: True
-
-runs:
-  using: composite
-  steps:
-    - name: Print Inputs
-      shell: bash
-      run: |
-        echo "workflow_id: ${{inputs.workflow_run_id}}"
-        echo "workflow_attempt: ${{inputs.workflow_attempt}}"
-        echo "workflow_Name: ${{inputs.workflow_name}}"
-        echo "job_id: ${{inputs.job_id}}"
-        echo "job_name:  ${{inputs.job_name}}"
-    - uses: nick-fields/retry@v3.0.0
-      name: Setup dependencies
-      with:
-        shell: bash
-        timeout_minutes: 5
-        max_attempts: 5
-        retry_wait_seconds: 30
-        command: |
-          set -eu
-          python3 -m pip install python-dateutil==2.8.2 boto3==1.35.42 pandas==2.1.3
-    - name: Upload utilizatoin stats to s3
-      shell: bash
-      run: |
-        python3 -m tools.stats.upload_utilization_stats.upload_utilization_stats \
-          --workflow-run-id "${{inputs.workflow_run_id}}" \
-          --workflow-name "${{inputs.workflow_name}}" \
-          --workflow-run-attempt "${{inputs.workflow_attempt}}" \
-          --job-id "${{inputs.job_id}}" \
-          --job-name "${{inputs.job_name}}"

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-f084f34bbb743fada85f66b0ed8041387565e69c
+b6d4675c7aedc53ba04f3f55786aac1de32be6b4

.github/ci_commit_pins/fbgemm_rocm.txt

@@ -1 +0,0 @@
-5fb5024118e9bb9decf96c2b0b1a8f0010bf56be

.github/ci_commit_pins/torchbench.txt

@@ -1 +1 @@
-373ffb19dc470f4423a3176a4133f8f4b3cdb5bd
+766a5e3a189384659fd35a68c3b17b88c761aaac

.github/labeler.yml

@@ -30,9 +30,9 @@
 - torch/fx/experimental/sym_node.py
 - torch/fx/experimental/validator.py
 - torch/fx/experimental/proxy_tensor.py
-- test/distributed/tensor/test_dtensor_compile.py
+- test/distributed/_tensor/test_dtensor_compile.py
 - test/distributed/tensor/parallel/test_fsdp_2d_parallel.py
-- torch/distributed/tensor/**
+- torch/distributed/_tensor/**
 - torch/distributed/fsdp/**
 - torch/csrc/inductor/**
 - torch/csrc/dynamo/**
@@ -107,8 +107,3 @@
 - torch/csrc/dynamo/compiled_autograd.h
 - torch/_dynamo/compiled_autograd.py
 - torch/inductor/test_compiled_autograd.py
-
-"ciflow/xpu":
-- torch/csrc/inductor/aoti_include/xpu.h
-- torch/csrc/inductor/cpp_wrapper/device_internal/xpu.h
-- torch/csrc/inductor/cpp_wrapper/xpu.h

.github/merge_rules.yaml

@@ -79,6 +79,7 @@
   - .ci/docker/ci_commit_pins/triton.txt
   approved_by:
   - pytorchbot
+  ignore_flaky_failures: false
   mandatory_checks_name:
   - EasyCLA
   - Lint
@@ -90,6 +91,7 @@
   - test/slow_tests.json
   approved_by:
   - pytorchbot
+  ignore_flaky_failures: false
   mandatory_checks_name:
   - EasyCLA
   - Lint
@@ -101,10 +103,12 @@
   - .ci/docker/ci_commit_pins/executorch.txt
   approved_by:
   - pytorchbot
+  ignore_flaky_failures: false
   mandatory_checks_name:
   - EasyCLA
   - Lint
-  - pull
+  - pull / linux-jammy-py3-clang12-executorch / build
+  - pull / linux-jammy-py3-clang12-executorch / test (executorch, 1, 1, linux.2xlarge)
 
 - name: OSS CI / pytorchbot / XLA
   patterns:
@@ -115,7 +119,8 @@
   mandatory_checks_name:
   - EasyCLA
   - Lint
-  - pull
+  - pull / linux-focal-py3_9-clang9-xla / build
+  - pull / linux-focal-py3_9-clang9-xla / test (xla, 1, 1, linux.12xlarge)
 
 - name: Documentation
   patterns:
@@ -242,6 +247,25 @@
   - Lint
   - pull
 
+- name: XPU ATen
+  patterns:
+  - aten/src/ATen/xpu/**
+  - c10/xpu/**
+  - torch/csrc/xpu/**
+  - torch/xpu/**
+  - test/xpu/**
+  - test/test_xpu.py
+  - third_party/xpu.txt
+  - .ci/docker/ci_commit_pins/triton-xpu.txt
+  approved_by:
+  - EikanWang
+  - jgong5
+  - gujinghui
+  mandatory_checks_name:
+  - EasyCLA
+  - Lint
+  - pull
+
 - name: Distributions
   patterns:
   - torch/distributions/**
@@ -495,19 +519,6 @@
   - Lint
   - pull
 
-- name: XPU
-  patterns:
-  - '**xpu**'
-  - '**sycl**'
-  approved_by:
-  - EikanWang
-  - jgong5
-  - gujinghui
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
 - name: superuser
   patterns:
   - '*'

.github/nitpicks.yml

@@ -3,10 +3,3 @@
     If you are adding a new function or defaulted argument to native_functions.yaml, you cannot use it from pre-existing Python frontend code until our FC window passes (two weeks).  Split your PR into two PRs, one which adds the new C++ functionality, and one that makes use of it from Python, and land them two weeks apart.  See https://github.com/pytorch/pytorch/wiki/PyTorch's-Python-Frontend-Backward-and-Forward-Compatibility-Policy#forwards-compatibility-fc for more info.
   pathFilter:
     - 'aten/src/ATen/native/native_functions.yaml'
-
-- markdown: |
-    ## Attention! PyTorch one of the C-stable API file was changed
-    You MUST NOT change existing function declarations in this, as this header defines a stable C ABI.  If you need to change the signature for a function, introduce a new v2 version of the function and modify code generation to target the new version of the function.
-  pathFilter:
-    - 'torch/csrc/inductor/aoti_torch/c/*'
-    - 'torch/csrc/inductor/aoti_torch/generated/*'

.github/requirements/pip-requirements-macOS.txt

@@ -19,7 +19,8 @@ pytest-rerunfailures==10.3
 pytest-flakefinder==1.1.0
 pytest-subtests==0.13.1
 scipy==1.10.1
-sympy==1.13.3
+sympy==1.12.1 ; python_version == "3.8"
+sympy==1.13.1 ; python_version >= "3.9"
 unittest-xml-reporting<=3.2.0,>=2.0.0
 xdoctest==1.1.0
 filelock==3.6.0

.github/scripts/build_triton_wheel.py

@@ -52,6 +52,7 @@ def build_triton(
     *,
     version: str,
     commit_hash: str,
+    build_conda: bool = False,
     device: str = "cuda",
     py_version: Optional[str] = None,
     release: bool = False,
@@ -82,6 +83,55 @@ def build_triton(
         else:
             check_call(["git", "checkout", commit_hash], cwd=triton_basedir)
 
+        if build_conda:
+            with open(triton_basedir / "meta.yaml", "w") as meta:
+                print(
+                    f"package:\n  name: torchtriton\n  version: {version}\n",
+                    file=meta,
+                )
+                print("source:\n  path: .\n", file=meta)
+                print(
+                    "build:\n  string: py{{py}}\n  number: 1\n  script: cd python; "
+                    "python setup.py install --record=record.txt\n",
+                    " script_env:\n   - MAX_JOBS\n",
+                    file=meta,
+                )
+                print(
+                    "requirements:\n  host:\n    - python\n    - setuptools\n    - pybind11\n"
+                    "  run:\n    - python\n    - filelock\n    - pytorch\n",
+                    file=meta,
+                )
+                print(
+                    "about:\n  home: https://github.com/openai/triton\n  license: MIT\n  summary:"
+                    " 'A language and compiler for custom Deep Learning operation'",
+                    file=meta,
+                )
+
+            patch_init_py(
+                triton_pythondir / "triton" / "__init__.py",
+                version=f"{version}",
+            )
+            if py_version is None:
+                py_version = f"{sys.version_info.major}.{sys.version_info.minor}"
+            check_call(
+                [
+                    "conda",
+                    "build",
+                    "--python",
+                    py_version,
+                    "-c",
+                    "pytorch-nightly",
+                    "--output-folder",
+                    tmpdir,
+                    ".",
+                ],
+                cwd=triton_basedir,
+                env=env,
+            )
+            conda_path = next(iter(Path(tmpdir).glob("linux-64/torchtriton*.bz2")))
+            shutil.copy(conda_path, Path.cwd())
+            return Path.cwd() / conda_path.name
+
         # change built wheel name and version
         env["TRITON_WHEEL_NAME"] = triton_pkg_name
         if with_clang_ldd:
@@ -122,6 +172,7 @@ def main() -> None:
 
     parser = ArgumentParser("Build Triton binaries")
     parser.add_argument("--release", action="store_true")
+    parser.add_argument("--build-conda", action="store_true")
     parser.add_argument(
         "--device", type=str, default="cuda", choices=["cuda", "rocm", "xpu"]
     )
@@ -137,6 +188,7 @@ def main() -> None:
             args.commit_hash if args.commit_hash else read_triton_pin(args.device)
         ),
         version=args.triton_version,
+        build_conda=args.build_conda,
         py_version=args.py_version,
         release=args.release,
         with_clang_ldd=args.with_clang_ldd,

.github/scripts/cherry_pick.py

@@ -3,7 +3,7 @@
 import json
 import os
 import re
-from typing import Any, cast, Optional
+from typing import Any, cast, Dict, List, Optional
 from urllib.error import HTTPError
 
 from github_utils import gh_fetch_url, gh_post_pr_comment, gh_query_issues_by_labels
@@ -67,7 +67,7 @@ def get_release_version(onto_branch: str) -> Optional[str]:
 
 def get_tracker_issues(
     org: str, project: str, onto_branch: str
-) -> list[dict[str, Any]]:
+) -> List[Dict[str, Any]]:
     """
     Find the tracker issue from the repo. The tracker issue needs to have the title
     like [VERSION] Release Tracker following the convention on PyTorch
@@ -117,7 +117,7 @@ def cherry_pick(
                 continue
 
             res = cast(
-                dict[str, Any],
+                Dict[str, Any],
                 post_tracker_issue_comment(
                     org,
                     project,
@@ -220,7 +220,7 @@ def submit_pr(
 
 def post_pr_comment(
     org: str, project: str, pr_num: int, msg: str, dry_run: bool = False
-) -> list[dict[str, Any]]:
+) -> List[Dict[str, Any]]:
     """
     Post a comment on the PR itself to point to the cherry picking PR when success
     or print the error when failure
@@ -255,7 +255,7 @@ def post_tracker_issue_comment(
     classification: str,
     fixes: str,
     dry_run: bool = False,
-) -> list[dict[str, Any]]:
+) -> List[Dict[str, Any]]:
     """
     Post a comment on the tracker issue (if any) to record the cherry pick
     """

.github/scripts/close_nonexistent_disable_issues.py

@@ -6,7 +6,7 @@ import re
 import sys
 import tempfile
 from pathlib import Path
-from typing import Any
+from typing import Any, Dict, List, Tuple
 
 import requests
 from gitutils import retries_decorator
@@ -76,7 +76,7 @@ DISABLED_TESTS_JSON = (
 
 
 @retries_decorator()
-def query_db(query: str, params: dict[str, Any]) -> list[dict[str, Any]]:
+def query_db(query: str, params: Dict[str, Any]) -> List[Dict[str, Any]]:
     return query_clickhouse(query, params)
 
 
@@ -97,7 +97,7 @@ def download_log_worker(temp_dir: str, id: int, name: str) -> None:
         f.write(data)
 
 
-def printer(item: tuple[str, tuple[int, str, list[Any]]], extra: str) -> None:
+def printer(item: Tuple[str, Tuple[int, str, List[Any]]], extra: str) -> None:
     test, (_, link, _) = item
     print(f"{link:<55} {test:<120} {extra}")
 
@@ -107,25 +107,21 @@ def close_issue(num: int) -> None:
         "Accept": "application/vnd.github.v3+json",
         "Authorization": f"token {os.environ['GITHUB_TOKEN']}",
     }
-    response = requests.post(
+    requests.post(
         f"https://api.github.com/repos/pytorch/pytorch/issues/{num}/comments",
         data=json.dumps({"body": CLOSING_COMMENT}),
         headers=headers,
     )
-    if response.status_code != 201:
-        raise RuntimeError(f"Failed to comment on issue {num}: {response.text}")
-    response = requests.patch(
+    requests.patch(
         f"https://api.github.com/repos/pytorch/pytorch/issues/{num}",
         data=json.dumps({"state": "closed"}),
         headers=headers,
     )
-    if response.status_code != 200:
-        raise RuntimeError(f"Failed to close issue {num}: {response.text}")
 
 
 def check_if_exists(
-    item: tuple[str, tuple[int, str, list[str]]], all_logs: list[str]
-) -> tuple[bool, str]:
+    item: Tuple[str, Tuple[int, str, List[str]]], all_logs: List[str]
+) -> Tuple[bool, str]:
     test, (_, link, _) = item
     # Test names should look like `test_a (module.path.classname)`
     reg = re.match(r"(\S+) \((\S*)\)", test)
@@ -194,13 +190,6 @@ if __name__ == "__main__":
     if args.dry_run:
         print("dry run, not actually closing")
     else:
-        failed = False
         for item in to_be_closed:
             _, (num, _, _) = item
-            try:
-                close_issue(num)
-            except RuntimeError as e:
-                print(e)
-                failed = True
-        if failed:
-            sys.exit(1)
+            close_issue(num)

.github/scripts/collect_ciflow_labels.py

@@ -2,7 +2,7 @@
 
 import sys
 from pathlib import Path
-from typing import Any, cast
+from typing import Any, cast, Dict, List, Set
 
 import yaml
 
@@ -10,9 +10,9 @@ import yaml
 GITHUB_DIR = Path(__file__).parent.parent
 
 
-def get_workflows_push_tags() -> set[str]:
+def get_workflows_push_tags() -> Set[str]:
     "Extract all known push tags from workflows"
-    rc: set[str] = set()
+    rc: Set[str] = set()
     for fname in (GITHUB_DIR / "workflows").glob("*.yml"):
         with fname.open("r") as f:
             wf_yml = yaml.safe_load(f)
@@ -25,19 +25,19 @@ def get_workflows_push_tags() -> set[str]:
     return rc
 
 
-def filter_ciflow_tags(tags: set[str]) -> list[str]:
+def filter_ciflow_tags(tags: Set[str]) -> List[str]:
     "Return sorted list of ciflow tags"
     return sorted(
         tag[:-2] for tag in tags if tag.startswith("ciflow/") and tag.endswith("/*")
     )
 
 
-def read_probot_config() -> dict[str, Any]:
+def read_probot_config() -> Dict[str, Any]:
     with (GITHUB_DIR / "pytorch-probot.yml").open("r") as f:
-        return cast(dict[str, Any], yaml.safe_load(f))
+        return cast(Dict[str, Any], yaml.safe_load(f))
 
 
-def update_probot_config(labels: set[str]) -> None:
+def update_probot_config(labels: Set[str]) -> None:
     orig = read_probot_config()
     orig["ciflow_push_tags"] = filter_ciflow_tags(labels)
     with (GITHUB_DIR / "pytorch-probot.yml").open("w") as f:

.github/scripts/delete_old_branches.py

@@ -4,7 +4,7 @@ import re
 from datetime import datetime
 from functools import lru_cache
 from pathlib import Path
-from typing import Any, Callable
+from typing import Any, Callable, Dict, List, Set
 
 from github_utils import gh_fetch_json_dict, gh_graphql
 from gitutils import GitRepo
@@ -112,7 +112,7 @@ def convert_gh_timestamp(date: str) -> float:
     return datetime.strptime(date, "%Y-%m-%dT%H:%M:%SZ").timestamp()
 
 
-def get_branches(repo: GitRepo) -> dict[str, Any]:
+def get_branches(repo: GitRepo) -> Dict[str, Any]:
     # Query locally for branches, group by branch base name (e.g. gh/blah/base -> gh/blah), and get the most recent branch
     git_response = repo._run_git(
         "for-each-ref",
@@ -120,7 +120,7 @@ def get_branches(repo: GitRepo) -> dict[str, Any]:
         "--format=%(refname) %(committerdate:iso-strict)",
         "refs/remotes/origin",
     )
-    branches_by_base_name: dict[str, Any] = {}
+    branches_by_base_name: Dict[str, Any] = {}
     for line in git_response.splitlines():
         branch, date = line.split(" ")
         re_branch = re.match(r"refs/remotes/origin/(.*)", branch)
@@ -140,14 +140,14 @@ def get_branches(repo: GitRepo) -> dict[str, Any]:
 
 def paginate_graphql(
     query: str,
-    kwargs: dict[str, Any],
-    termination_func: Callable[[list[dict[str, Any]]], bool],
-    get_data: Callable[[dict[str, Any]], list[dict[str, Any]]],
-    get_page_info: Callable[[dict[str, Any]], dict[str, Any]],
-) -> list[Any]:
+    kwargs: Dict[str, Any],
+    termination_func: Callable[[List[Dict[str, Any]]], bool],
+    get_data: Callable[[Dict[str, Any]], List[Dict[str, Any]]],
+    get_page_info: Callable[[Dict[str, Any]], Dict[str, Any]],
+) -> List[Any]:
     hasNextPage = True
     endCursor = None
-    data: list[dict[str, Any]] = []
+    data: List[Dict[str, Any]] = []
     while hasNextPage:
         ESTIMATED_TOKENS[0] += 1
         res = gh_graphql(query, cursor=endCursor, **kwargs)
@@ -159,11 +159,11 @@ def paginate_graphql(
     return data
 
 
-def get_recent_prs() -> dict[str, Any]:
+def get_recent_prs() -> Dict[str, Any]:
     now = datetime.now().timestamp()
 
     # Grab all PRs updated in last CLOSED_PR_RETENTION days
-    pr_infos: list[dict[str, Any]] = paginate_graphql(
+    pr_infos: List[Dict[str, Any]] = paginate_graphql(
         GRAPHQL_ALL_PRS_BY_UPDATED_AT,
         {"owner": "pytorch", "repo": "pytorch"},
         lambda data: (
@@ -190,7 +190,7 @@ def get_recent_prs() -> dict[str, Any]:
 
 
 @lru_cache(maxsize=1)
-def get_open_prs() -> list[dict[str, Any]]:
+def get_open_prs() -> List[Dict[str, Any]]:
     return paginate_graphql(
         GRAPHQL_OPEN_PRS,
         {"owner": "pytorch", "repo": "pytorch"},
@@ -200,8 +200,8 @@ def get_open_prs() -> list[dict[str, Any]]:
     )
 
 
-def get_branches_with_magic_label_or_open_pr() -> set[str]:
-    pr_infos: list[dict[str, Any]] = paginate_graphql(
+def get_branches_with_magic_label_or_open_pr() -> Set[str]:
+    pr_infos: List[Dict[str, Any]] = paginate_graphql(
         GRAPHQL_NO_DELETE_BRANCH_LABEL,
         {"owner": "pytorch", "repo": "pytorch"},
         lambda data: False,

.github/scripts/file_io_utils.py

@@ -2,7 +2,7 @@ import json
 import re
 import shutil
 from pathlib import Path
-from typing import Any
+from typing import Any, List
 
 import boto3  # type: ignore[import]
 
@@ -77,7 +77,7 @@ def upload_file_to_s3(file_name: Path, bucket: str, key: str) -> None:
 
 def download_s3_objects_with_prefix(
     bucket_name: str, prefix: str, download_folder: Path
-) -> list[Path]:
+) -> List[Path]:
     s3 = boto3.resource("s3")
     bucket = s3.Bucket(bucket_name)
 

.github/scripts/filter_test_configs.py

@@ -8,9 +8,9 @@ import subprocess
 import sys
 import warnings
 from enum import Enum
-from functools import cache
+from functools import lru_cache
 from logging import info
-from typing import Any, Callable, Optional
+from typing import Any, Callable, Dict, List, Optional, Set
 from urllib.request import Request, urlopen
 
 import yaml
@@ -32,7 +32,7 @@ def is_cuda_or_rocm_job(job_name: Optional[str]) -> bool:
 
 # Supported modes when running periodically. Only applying the mode when
 # its lambda condition returns true
-SUPPORTED_PERIODICAL_MODES: dict[str, Callable[[Optional[str]], bool]] = {
+SUPPORTED_PERIODICAL_MODES: Dict[str, Callable[[Optional[str]], bool]] = {
     # Memory leak check is only needed for CUDA and ROCm jobs which utilize GPU memory
     "mem_leak_check": is_cuda_or_rocm_job,
     "rerun_disabled_tests": lambda job_name: True,
@@ -102,8 +102,8 @@ def parse_args() -> Any:
     return parser.parse_args()
 
 
-@cache
-def get_pr_info(pr_number: int) -> dict[str, Any]:
+@lru_cache(maxsize=None)
+def get_pr_info(pr_number: int) -> Dict[str, Any]:
     """
     Dynamically get PR information
     """
@@ -116,7 +116,7 @@ def get_pr_info(pr_number: int) -> dict[str, Any]:
         "Accept": "application/vnd.github.v3+json",
         "Authorization": f"token {github_token}",
     }
-    json_response: dict[str, Any] = download_json(
+    json_response: Dict[str, Any] = download_json(
         url=f"{pytorch_github_api}/issues/{pr_number}",
         headers=headers,
     )
@@ -128,7 +128,7 @@ def get_pr_info(pr_number: int) -> dict[str, Any]:
     return json_response
 
 
-def get_labels(pr_number: int) -> set[str]:
+def get_labels(pr_number: int) -> Set[str]:
     """
     Dynamically get the latest list of labels from the pull request
     """
@@ -138,14 +138,14 @@ def get_labels(pr_number: int) -> set[str]:
     }
 
 
-def filter_labels(labels: set[str], label_regex: Any) -> set[str]:
+def filter_labels(labels: Set[str], label_regex: Any) -> Set[str]:
     """
     Return the list of matching labels
     """
     return {l for l in labels if re.match(label_regex, l)}
 
 
-def filter(test_matrix: dict[str, list[Any]], labels: set[str]) -> dict[str, list[Any]]:
+def filter(test_matrix: Dict[str, List[Any]], labels: Set[str]) -> Dict[str, List[Any]]:
     """
     Select the list of test config to run from the test matrix. The logic works
     as follows:
@@ -157,7 +157,7 @@ def filter(test_matrix: dict[str, list[Any]], labels: set[str]) -> dict[str, lis
 
     If the PR has none of the test-config label, all tests are run as usual.
     """
-    filtered_test_matrix: dict[str, list[Any]] = {"include": []}
+    filtered_test_matrix: Dict[str, List[Any]] = {"include": []}
 
     for entry in test_matrix.get("include", []):
         config_name = entry.get("config", "")
@@ -185,8 +185,8 @@ def filter(test_matrix: dict[str, list[Any]], labels: set[str]) -> dict[str, lis
 
 
 def filter_selected_test_configs(
-    test_matrix: dict[str, list[Any]], selected_test_configs: set[str]
-) -> dict[str, list[Any]]:
+    test_matrix: Dict[str, List[Any]], selected_test_configs: Set[str]
+) -> Dict[str, List[Any]]:
     """
     Keep only the selected configs if the list if not empty. Otherwise, keep all test configs.
     This filter is used when the workflow is dispatched manually.
@@ -194,7 +194,7 @@ def filter_selected_test_configs(
     if not selected_test_configs:
         return test_matrix
 
-    filtered_test_matrix: dict[str, list[Any]] = {"include": []}
+    filtered_test_matrix: Dict[str, List[Any]] = {"include": []}
     for entry in test_matrix.get("include", []):
         config_name = entry.get("config", "")
         if not config_name:
@@ -207,12 +207,12 @@ def filter_selected_test_configs(
 
 
 def set_periodic_modes(
-    test_matrix: dict[str, list[Any]], job_name: Optional[str]
-) -> dict[str, list[Any]]:
+    test_matrix: Dict[str, List[Any]], job_name: Optional[str]
+) -> Dict[str, List[Any]]:
     """
     Apply all periodic modes when running under a schedule
     """
-    scheduled_test_matrix: dict[str, list[Any]] = {
+    scheduled_test_matrix: Dict[str, List[Any]] = {
         "include": [],
     }
 
@@ -229,8 +229,8 @@ def set_periodic_modes(
 
 
 def mark_unstable_jobs(
-    workflow: str, job_name: str, test_matrix: dict[str, list[Any]]
-) -> dict[str, list[Any]]:
+    workflow: str, job_name: str, test_matrix: Dict[str, List[Any]]
+) -> Dict[str, List[Any]]:
     """
     Check the list of unstable jobs and mark them accordingly. Note that if a job
     is unstable, all its dependents will also be marked accordingly
@@ -245,8 +245,8 @@ def mark_unstable_jobs(
 
 
 def remove_disabled_jobs(
-    workflow: str, job_name: str, test_matrix: dict[str, list[Any]]
-) -> dict[str, list[Any]]:
+    workflow: str, job_name: str, test_matrix: Dict[str, List[Any]]
+) -> Dict[str, List[Any]]:
     """
     Check the list of disabled jobs, remove the current job and all its dependents
     if it exists in the list
@@ -261,15 +261,15 @@ def remove_disabled_jobs(
 
 
 def _filter_jobs(
-    test_matrix: dict[str, list[Any]],
+    test_matrix: Dict[str, List[Any]],
     issue_type: IssueType,
     target_cfg: Optional[str] = None,
-) -> dict[str, list[Any]]:
+) -> Dict[str, List[Any]]:
     """
     An utility function used to actually apply the job filter
     """
     # The result will be stored here
-    filtered_test_matrix: dict[str, list[Any]] = {"include": []}
+    filtered_test_matrix: Dict[str, List[Any]] = {"include": []}
 
     # This is an issue to disable a CI job
     if issue_type == IssueType.DISABLED:
@@ -302,10 +302,10 @@ def _filter_jobs(
 def process_jobs(
     workflow: str,
     job_name: str,
-    test_matrix: dict[str, list[Any]],
+    test_matrix: Dict[str, List[Any]],
     issue_type: IssueType,
     url: str,
-) -> dict[str, list[Any]]:
+) -> Dict[str, List[Any]]:
     """
     Both disabled and unstable jobs are in the following format:
 
@@ -441,7 +441,7 @@ def process_jobs(
     return test_matrix
 
 
-def download_json(url: str, headers: dict[str, str], num_retries: int = 3) -> Any:
+def download_json(url: str, headers: Dict[str, str], num_retries: int = 3) -> Any:
     for _ in range(num_retries):
         try:
             req = Request(url=url, headers=headers)
@@ -462,7 +462,7 @@ def set_output(name: str, val: Any) -> None:
         print(f"::set-output name={name}::{val}")
 
 
-def parse_reenabled_issues(s: Optional[str]) -> list[str]:
+def parse_reenabled_issues(s: Optional[str]) -> List[str]:
     # NB: When the PR body is empty, GitHub API returns a None value, which is
     # passed into this function
     if not s:
@@ -477,7 +477,7 @@ def parse_reenabled_issues(s: Optional[str]) -> list[str]:
     return issue_numbers
 
 
-def get_reenabled_issues(pr_body: str = "") -> list[str]:
+def get_reenabled_issues(pr_body: str = "") -> List[str]:
     default_branch = f"origin/{os.environ.get('GIT_DEFAULT_BRANCH', 'main')}"
     try:
         commit_messages = subprocess.check_output(
@@ -489,12 +489,12 @@ def get_reenabled_issues(pr_body: str = "") -> list[str]:
     return parse_reenabled_issues(pr_body) + parse_reenabled_issues(commit_messages)
 
 
-def check_for_setting(labels: set[str], body: str, setting: str) -> bool:
+def check_for_setting(labels: Set[str], body: str, setting: str) -> bool:
     return setting in labels or f"[{setting}]" in body
 
 
 def perform_misc_tasks(
-    labels: set[str], test_matrix: dict[str, list[Any]], job_name: str, pr_body: str
+    labels: Set[str], test_matrix: Dict[str, List[Any]], job_name: str, pr_body: str
 ) -> None:
     """
     In addition to apply the filter logic, the script also does the following

.github/scripts/generate_binary_build_matrix.py

@@ -12,23 +12,13 @@ architectures:
 """
 
 import os
-from typing import Optional
+from typing import Dict, List, Optional, Tuple
 
 
 # NOTE: Also update the CUDA sources in tools/nightly.py when changing this list
-CUDA_ARCHES = ["11.8", "12.4", "12.6", "12.8"]
-CUDA_ARCHES_FULL_VERSION = {
-    "11.8": "11.8.0",
-    "12.4": "12.4.1",
-    "12.6": "12.6.3",
-    "12.8": "12.8.0",
-}
-CUDA_ARCHES_CUDNN_VERSION = {
-    "11.8": "9",
-    "12.4": "9",
-    "12.6": "9",
-    "12.8": "9",
-}
+CUDA_ARCHES = ["11.8", "12.4", "12.6"]
+CUDA_ARCHES_FULL_VERSION = {"11.8": "11.8.0", "12.4": "12.4.1", "12.6": "12.6.3"}
+CUDA_ARCHES_CUDNN_VERSION = {"11.8": "9", "12.4": "9", "12.6": "9"}
 
 # NOTE: Also update the ROCm sources in tools/nightly.py when changing this list
 ROCM_ARCHES = ["6.2.4", "6.3"]
@@ -41,7 +31,7 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64"]
+CUDA_AARCH64_ARCH = ["cuda-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
@@ -69,7 +59,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusolver-cu12==11.6.1.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparse-cu12==12.3.1.170; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparselt-cu12==0.6.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.25.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvtx-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvjitlink-cu12==12.4.127; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
@@ -84,26 +74,9 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparselt-cu12==0.6.3; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.25.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.21.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'"
-    ),
-    "12.8": (
-        "nvidia-cuda-nvrtc-cu12==12.8.61; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-runtime-cu12==12.8.57; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-cupti-cu12==12.8.57; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cudnn-cu12==9.7.1.26; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cublas-cu12==12.8.3.14; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufft-cu12==11.3.3.41; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-curand-cu12==10.3.9.55; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusolver-cu12==11.7.2.55; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparse-cu12==12.5.7.53; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparselt-cu12==0.6.3; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.25.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvtx-cu12==12.8.55; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvjitlink-cu12==12.8.61; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufile-cu12==1.13.0.11; platform_system == 'Linux' and platform_machine == 'x86_64'"
+        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "xpu": (
         "intel-cmplr-lib-rt==2025.0.2 | "
@@ -112,11 +85,37 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "intel-sycl-rt==2025.0.2 | "
         "tcmlib==1.2.0 | "
         "umf==0.9.1 | "
-        "intel-pti==0.10.0"
+        "intel-pti==0.10.0; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
 }
 
 
+def get_nccl_submodule_version() -> str:
+    from pathlib import Path
+
+    nccl_version_mk = (
+        Path(__file__).absolute().parents[2]
+        / "third_party"
+        / "nccl"
+        / "nccl"
+        / "makefiles"
+        / "version.mk"
+    )
+    if not nccl_version_mk.exists():
+        raise RuntimeError(
+            "Please make sure that nccl submodule is checked out when importing this script"
+        )
+    with nccl_version_mk.open("r") as f:
+        content = f.read()
+    d = {}
+    for l in content.split("\n"):
+        if not l.startswith("NCCL_"):
+            continue
+        (k, v) = l.split(":=")
+        d[k.strip()] = v.strip()
+    return f"{d['NCCL_MAJOR']}.{d['NCCL_MINOR']}.{d['NCCL_PATCH']}"
+
+
 def get_nccl_wheel_version(arch_version: str) -> str:
     import re
 
@@ -128,26 +127,12 @@ def get_nccl_wheel_version(arch_version: str) -> str:
     ]
 
 
-def read_nccl_pin(arch_version: str) -> str:
-    from pathlib import Path
-
-    nccl_pin_path = os.path.join(
-        Path(__file__).absolute().parents[2],
-        ".ci",
-        "docker",
-        "ci_commit_pins",
-        f"nccl-cu{arch_version[:2]}.txt",
-    )
-    with open(nccl_pin_path) as f:
-        return f.read().strip()
-
-
 def validate_nccl_dep_consistency(arch_version: str) -> None:
-    nccl_release_tag = read_nccl_pin(arch_version)
     wheel_ver = get_nccl_wheel_version(arch_version)
-    if not nccl_release_tag.startswith(f"v{wheel_ver}"):
+    submodule_ver = get_nccl_submodule_version()
+    if wheel_ver != submodule_ver:
         raise RuntimeError(
-            f"{arch_version} NCCL release tag version {nccl_release_tag} does not correspond to wheel version {wheel_ver}"
+            f"NCCL submodule version {submodule_ver} differs from wheel version {wheel_ver}"
         )
 
 
@@ -164,7 +149,7 @@ def arch_type(arch_version: str) -> str:
         return "cpu-aarch64"
     elif arch_version in CPU_S390X_ARCH:
         return "cpu-s390x"
-    elif arch_version in CUDA_AARCH64_ARCHES:
+    elif arch_version in CUDA_AARCH64_ARCH:
         return "cuda-aarch64"
     else:  # arch_version should always be "cpu" in this case
         return "cpu"
@@ -178,10 +163,6 @@ WHEEL_CONTAINER_IMAGES = {
         gpu_arch: f"pytorch/manylinux2_28-builder:cuda{gpu_arch}-{DEFAULT_TAG}"
         for gpu_arch in CUDA_ARCHES
     },
-    **{
-        gpu_arch: f"pytorch/manylinuxaarch64-builder:cuda{gpu_arch.replace('-aarch64', '')}-{DEFAULT_TAG}"
-        for gpu_arch in CUDA_AARCH64_ARCHES
-    },
     **{
         gpu_arch: f"pytorch/manylinux2_28-builder:rocm{gpu_arch}-{DEFAULT_TAG}"
         for gpu_arch in ROCM_ARCHES
@@ -191,13 +172,23 @@ WHEEL_CONTAINER_IMAGES = {
     "cpu-cxx11-abi": f"pytorch/manylinuxcxx11-abi-builder:cpu-cxx11-abi-{DEFAULT_TAG}",
     "cpu-aarch64": f"pytorch/manylinux2_28_aarch64-builder:cpu-aarch64-{DEFAULT_TAG}",
     "cpu-s390x": f"pytorch/manylinuxs390x-builder:cpu-s390x-{DEFAULT_TAG}",
+    "cuda-aarch64": f"pytorch/manylinuxaarch64-builder:cuda12.6-{DEFAULT_TAG}",
 }
 
+
+PRE_CXX11_ABI = "pre-cxx11"
 CXX11_ABI = "cxx11-abi"
 RELEASE = "release"
 DEBUG = "debug"
 
-LIBTORCH_CONTAINER_IMAGES: dict[tuple[str, str], str] = {
+LIBTORCH_CONTAINER_IMAGES: Dict[Tuple[str, str], str] = {
+    **{
+        (
+            gpu_arch,
+            PRE_CXX11_ABI,
+        ): f"pytorch/manylinux-builder:cuda{gpu_arch}-{DEFAULT_TAG}"
+        for gpu_arch in CUDA_ARCHES
+    },
     **{
         (
             gpu_arch,
@@ -212,10 +203,11 @@ LIBTORCH_CONTAINER_IMAGES: dict[tuple[str, str], str] = {
         ): f"pytorch/libtorch-cxx11-builder:rocm{gpu_arch}-{DEFAULT_TAG}"
         for gpu_arch in ROCM_ARCHES
     },
+    ("cpu", PRE_CXX11_ABI): f"pytorch/manylinux-builder:cpu-{DEFAULT_TAG}",
     ("cpu", CXX11_ABI): f"pytorch/libtorch-cxx11-builder:cpu-{DEFAULT_TAG}",
 }
 
-FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t"]
+FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12"]
 
 
 def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
@@ -225,35 +217,29 @@ def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
         "cpu-cxx11-abi": "cpu-cxx11-abi",
         "cpu-s390x": "cpu",
         "cuda": f"cu{gpu_arch_version.replace('.', '')}",
-        "cuda-aarch64": f"cu{gpu_arch_version.replace('-aarch64', '').replace('.', '')}",
+        "cuda-aarch64": "cu126",
         "rocm": f"rocm{gpu_arch_version}",
         "xpu": "xpu",
     }.get(gpu_arch_type, gpu_arch_version)
 
 
-def list_without(in_list: list[str], without: list[str]) -> list[str]:
+def list_without(in_list: List[str], without: List[str]) -> List[str]:
     return [item for item in in_list if item not in without]
 
 
 def generate_libtorch_matrix(
     os: str,
     abi_version: str,
-    arches: Optional[list[str]] = None,
-    libtorch_variants: Optional[list[str]] = None,
-) -> list[dict[str, str]]:
+    arches: Optional[List[str]] = None,
+    libtorch_variants: Optional[List[str]] = None,
+) -> List[Dict[str, str]]:
     if arches is None:
         arches = ["cpu"]
         if os == "linux":
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
-            # skip CUDA 12.8 builds for libtorch
-            if "12.8" in arches:
-                arches.remove("12.8")
         elif os == "windows":
             arches += CUDA_ARCHES
-            # skip CUDA 12.8 builds on Windows
-            if "12.8" in arches:
-                arches.remove("12.8")
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -262,7 +248,7 @@ def generate_libtorch_matrix(
             "static-without-deps",
         ]
 
-    ret: list[dict[str, str]] = []
+    ret: List[Dict[str, str]] = []
     for arch_version in arches:
         for libtorch_variant in libtorch_variants:
             # one of the values in the following list must be exactly
@@ -271,7 +257,9 @@ def generate_libtorch_matrix(
             gpu_arch_type = arch_type(arch_version)
             gpu_arch_version = "" if arch_version == "cpu" else arch_version
             # ROCm builds without-deps failed even in ROCm runners; skip for now
-            if gpu_arch_type == "rocm" and ("without-deps" in libtorch_variant):
+            if gpu_arch_type == "rocm" and (
+                "without-deps" in libtorch_variant or "pre-cxx11" in abi_version
+            ):
                 continue
             ret.append(
                 {
@@ -299,17 +287,17 @@ def generate_libtorch_matrix(
 
 def generate_wheels_matrix(
     os: str,
-    arches: Optional[list[str]] = None,
-    python_versions: Optional[list[str]] = None,
+    arches: Optional[List[str]] = None,
+    python_versions: Optional[List[str]] = None,
     use_split_build: bool = False,
-) -> list[dict[str, str]]:
+) -> List[Dict[str, str]]:
     package_type = "wheel"
     if os == "linux" or os == "linux-aarch64" or os == "linux-s390x":
         # NOTE: We only build manywheel packages for x86_64 and aarch64 and s390x linux
         package_type = "manywheel"
 
     if python_versions is None:
-        python_versions = FULL_PYTHON_VERSIONS
+        python_versions = FULL_PYTHON_VERSIONS + ["3.13", "3.13t"]
 
     if arches is None:
         # Define default compute archivectures
@@ -318,19 +306,16 @@ def generate_wheels_matrix(
             arches += CPU_CXX11_ABI_ARCH + CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
             arches += CUDA_ARCHES + XPU_ARCHES
-            # skip CUDA 12.8 builds on Windows until available
-            if "12.8" in arches:
-                arches.remove("12.8")
         elif os == "linux-aarch64":
-            # Separate new if as the CPU type is different and
+            # Only want the one arch as the CPU type is different and
             # uses different build/test scripts
-            arches = CPU_AARCH64_ARCH + CUDA_AARCH64_ARCHES
+            arches = ["cpu-aarch64", "cuda-aarch64"]
         elif os == "linux-s390x":
             # Only want the one arch as the CPU type is different and
             # uses different build/test scripts
             arches = ["cpu-s390x"]
 
-    ret: list[dict[str, str]] = []
+    ret: List[Dict[str, str]] = []
     for python_version in python_versions:
         for arch_version in arches:
             gpu_arch_type = arch_type(arch_version)
@@ -340,19 +325,38 @@ def generate_wheels_matrix(
                 or arch_version == "cpu-cxx11-abi"
                 or arch_version == "cpu-aarch64"
                 or arch_version == "cpu-s390x"
+                or arch_version == "cuda-aarch64"
                 or arch_version == "xpu"
                 else arch_version
             )
 
-            # TODO: Enable python 3.13t on cpu-s390x
-            if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
+            # TODO: Enable python 3.13 on aarch64, windows
+            if (
+                os
+                not in [
+                    "linux",
+                    "linux-s390x",
+                    "linux-aarch64",
+                    "macos-arm64",
+                    "windows",
+                ]
+            ) and python_version in ["3.13", "3.13t"]:
+                continue
+
+            # TODO: Enable python 3.13t on xpu and cpu-s390x or MacOS or Windows
+            if (
+                gpu_arch_type in ["xpu", "cpu-s390x"]
+                or os == "macos-arm64"
+                or os == "linux-aarch64"
+                or os == "windows"
+            ) and python_version == "3.13t":
                 continue
 
             if use_split_build and (
                 arch_version not in ["12.6", "12.4", "11.8", "cpu"] or os != "linux"
             ):
                 raise RuntimeError(
-                    "Split build is only supported on linux with cuda 12*, 11.8, and cpu.\n"
+                    "Split build is only supported on linux with cuda 12.6, 12.4, 11.8, and cpu.\n"
                     f"Currently attempting to build on arch version {arch_version} and os {os}.\n"
                     "Please modify the matrix generation to exclude this combination."
                 )
@@ -360,9 +364,9 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["12.8", "12.6", "12.4", "11.8"]
+                arch_version in ["12.6", "12.4", "11.8"]
                 and os == "linux"
-                or arch_version in CUDA_AARCH64_ARCHES
+                or arch_version == "cuda-aarch64"
             ):
                 ret.append(
                     {
@@ -381,12 +385,9 @@ def generate_wheels_matrix(
                             if os != "linux-aarch64"
                             else ""
                         ),
-                        "build_name": (
-                            f"{package_type}-py{python_version}-{gpu_arch_type}"
-                            f"{'-' if 'aarch64' in gpu_arch_type else ''}{gpu_arch_version.replace('-aarch64', '')}".replace(
-                                ".", "_"
-                            )
-                        ),  # include special case for aarch64 build, remove the -aarch64 postfix
+                        "build_name": f"{package_type}-py{python_version}-{gpu_arch_type}{gpu_arch_version}".replace(  # noqa: B950
+                            ".", "_"
+                        ),
                     }
                 )
                 # Special build building to use on Colab. Python 3.11 for 12.4 CUDA
@@ -443,7 +444,6 @@ def generate_wheels_matrix(
     return ret
 
 
-validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")
 validate_nccl_dep_consistency("12.4")
 validate_nccl_dep_consistency("11.8")

.github/scripts/generate_ci_workflows.py

@@ -2,10 +2,9 @@
 
 import os
 import sys
-from collections.abc import Iterable
 from dataclasses import asdict, dataclass, field
 from pathlib import Path
-from typing import Literal
+from typing import Dict, Iterable, List, Literal, Set
 from typing_extensions import TypedDict  # Python 3.11+
 
 import generate_binary_build_matrix  # type: ignore[import]
@@ -28,7 +27,7 @@ LABEL_CIFLOW_BINARIES_WHEEL = "ciflow/binaries_wheel"
 class CIFlowConfig:
     # For use to enable workflows to run on pytorch/pytorch-canary
     run_on_canary: bool = False
-    labels: set[str] = field(default_factory=set)
+    labels: Set[str] = field(default_factory=set)
     # Certain jobs might not want to be part of the ciflow/[all,trunk] workflow
     isolated_workflow: bool = False
     unstable: bool = False
@@ -49,7 +48,7 @@ class Config(TypedDict):
 @dataclass
 class BinaryBuildWorkflow:
     os: str
-    build_configs: list[dict[str, str]]
+    build_configs: List[Dict[str, str]]
     package_type: str
 
     # Optional fields
@@ -143,6 +142,20 @@ LINUX_BINARY_BUILD_WORFKLOWS = [
             isolated_workflow=True,
         ),
     ),
+    BinaryBuildWorkflow(
+        os=OperatingSystem.LINUX,
+        package_type="libtorch",
+        abi_version=generate_binary_build_matrix.PRE_CXX11_ABI,
+        build_configs=generate_binary_build_matrix.generate_libtorch_matrix(
+            OperatingSystem.LINUX,
+            generate_binary_build_matrix.PRE_CXX11_ABI,
+            libtorch_variants=["shared-with-deps"],
+        ),
+        ciflow_config=CIFlowConfig(
+            labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_LIBTORCH},
+            isolated_workflow=True,
+        ),
+    ),
 ]
 
 LINUX_BINARY_SMOKE_WORKFLOWS = [
@@ -151,7 +164,7 @@ LINUX_BINARY_SMOKE_WORKFLOWS = [
         package_type="manywheel",
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
-            arches=["11.8", "12.4", "12.6", "12.8"],
+            arches=["11.8", "12.4", "12.6"],
             python_versions=["3.9"],
         ),
         branches="main",
@@ -184,6 +197,18 @@ LINUX_BINARY_SMOKE_WORKFLOWS = [
         ),
         branches="main",
     ),
+    BinaryBuildWorkflow(
+        os=OperatingSystem.LINUX,
+        package_type="libtorch",
+        abi_version=generate_binary_build_matrix.PRE_CXX11_ABI,
+        build_configs=generate_binary_build_matrix.generate_libtorch_matrix(
+            OperatingSystem.LINUX,
+            generate_binary_build_matrix.PRE_CXX11_ABI,
+            arches=["cpu"],
+            libtorch_variants=["shared-with-deps"],
+        ),
+        branches="main",
+    ),
 ]
 
 WINDOWS_BINARY_BUILD_WORKFLOWS = [

.github/scripts/generate_docker_release_matrix.py

@@ -12,6 +12,7 @@ Will output a condensed version of the matrix. Will include fllowing:
 """
 
 import json
+from typing import Dict, List
 
 import generate_binary_build_matrix
 
@@ -19,8 +20,8 @@ import generate_binary_build_matrix
 DOCKER_IMAGE_TYPES = ["runtime", "devel"]
 
 
-def generate_docker_matrix() -> dict[str, list[dict[str, str]]]:
-    ret: list[dict[str, str]] = []
+def generate_docker_matrix() -> Dict[str, List[Dict[str, str]]]:
+    ret: List[Dict[str, str]] = []
     # CUDA amd64 Docker images are available as both runtime and devel while
     # CPU arm64 image is only available as runtime.
     for cuda, version in generate_binary_build_matrix.CUDA_ARCHES_FULL_VERSION.items():

.github/scripts/get_workflow_job_id.py

@@ -11,11 +11,11 @@ import sys
 import time
 import urllib
 import urllib.parse
-from typing import Any, Callable, Optional
+from typing import Any, Callable, Dict, List, Optional, Tuple
 from urllib.request import Request, urlopen
 
 
-def parse_json_and_links(conn: Any) -> tuple[Any, dict[str, dict[str, str]]]:
+def parse_json_and_links(conn: Any) -> Tuple[Any, Dict[str, Dict[str, str]]]:
     links = {}
     # Extract links which GH uses for pagination
     # see https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Link
@@ -42,7 +42,7 @@ def parse_json_and_links(conn: Any) -> tuple[Any, dict[str, dict[str, str]]]:
 def fetch_url(
     url: str,
     *,
-    headers: Optional[dict[str, str]] = None,
+    headers: Optional[Dict[str, str]] = None,
     reader: Callable[[Any], Any] = lambda x: x.read(),
     retries: Optional[int] = 3,
     backoff_timeout: float = 0.5,
@@ -83,7 +83,7 @@ def parse_args() -> Any:
     return parser.parse_args()
 
 
-def fetch_jobs(url: str, headers: dict[str, str]) -> list[dict[str, str]]:
+def fetch_jobs(url: str, headers: Dict[str, str]) -> List[Dict[str, str]]:
     response, links = fetch_url(url, headers=headers, reader=parse_json_and_links)
     jobs = response["jobs"]
     assert type(jobs) is list
@@ -111,7 +111,7 @@ def fetch_jobs(url: str, headers: dict[str, str]) -> list[dict[str, str]]:
 # running.
 
 
-def find_job_id_name(args: Any) -> tuple[str, str]:
+def find_job_id_name(args: Any) -> Tuple[str, str]:
     # From https://docs.github.com/en/actions/learn-github-actions/environment-variables
     PYTORCH_REPO = os.environ.get("GITHUB_REPOSITORY", "pytorch/pytorch")
     PYTORCH_GITHUB_API = f"https://api.github.com/repos/{PYTORCH_REPO}"

.github/scripts/github_utils.py

@@ -4,7 +4,7 @@ import json
 import os
 import warnings
 from dataclasses import dataclass
-from typing import Any, Callable, cast, Optional, Union
+from typing import Any, Callable, cast, Dict, List, Optional, Tuple, Union
 from urllib.error import HTTPError
 from urllib.parse import quote
 from urllib.request import Request, urlopen
@@ -27,11 +27,11 @@ class GitHubComment:
 def gh_fetch_url_and_headers(
     url: str,
     *,
-    headers: Optional[dict[str, str]] = None,
-    data: Union[Optional[dict[str, Any]], str] = None,
+    headers: Optional[Dict[str, str]] = None,
+    data: Union[Optional[Dict[str, Any]], str] = None,
     method: Optional[str] = None,
     reader: Callable[[Any], Any] = lambda x: x.read(),
-) -> tuple[Any, Any]:
+) -> Tuple[Any, Any]:
     if headers is None:
         headers = {}
     token = os.environ.get("GITHUB_TOKEN")
@@ -70,8 +70,8 @@ def gh_fetch_url_and_headers(
 def gh_fetch_url(
     url: str,
     *,
-    headers: Optional[dict[str, str]] = None,
-    data: Union[Optional[dict[str, Any]], str] = None,
+    headers: Optional[Dict[str, str]] = None,
+    data: Union[Optional[Dict[str, Any]], str] = None,
     method: Optional[str] = None,
     reader: Callable[[Any], Any] = json.load,
 ) -> Any:
@@ -82,25 +82,25 @@ def gh_fetch_url(
 
 def gh_fetch_json(
     url: str,
-    params: Optional[dict[str, Any]] = None,
-    data: Optional[dict[str, Any]] = None,
+    params: Optional[Dict[str, Any]] = None,
+    data: Optional[Dict[str, Any]] = None,
     method: Optional[str] = None,
-) -> list[dict[str, Any]]:
+) -> List[Dict[str, Any]]:
     headers = {"Accept": "application/vnd.github.v3+json"}
     if params is not None and len(params) > 0:
         url += "?" + "&".join(
             f"{name}={quote(str(val))}" for name, val in params.items()
         )
     return cast(
-        list[dict[str, Any]],
+        List[Dict[str, Any]],
         gh_fetch_url(url, headers=headers, data=data, reader=json.load, method=method),
     )
 
 
 def _gh_fetch_json_any(
     url: str,
-    params: Optional[dict[str, Any]] = None,
-    data: Optional[dict[str, Any]] = None,
+    params: Optional[Dict[str, Any]] = None,
+    data: Optional[Dict[str, Any]] = None,
 ) -> Any:
     headers = {"Accept": "application/vnd.github.v3+json"}
     if params is not None and len(params) > 0:
@@ -112,21 +112,21 @@ def _gh_fetch_json_any(
 
 def gh_fetch_json_list(
     url: str,
-    params: Optional[dict[str, Any]] = None,
-    data: Optional[dict[str, Any]] = None,
-) -> list[dict[str, Any]]:
-    return cast(list[dict[str, Any]], _gh_fetch_json_any(url, params, data))
+    params: Optional[Dict[str, Any]] = None,
+    data: Optional[Dict[str, Any]] = None,
+) -> List[Dict[str, Any]]:
+    return cast(List[Dict[str, Any]], _gh_fetch_json_any(url, params, data))
 
 
 def gh_fetch_json_dict(
     url: str,
-    params: Optional[dict[str, Any]] = None,
-    data: Optional[dict[str, Any]] = None,
-) -> dict[str, Any]:
-    return cast(dict[str, Any], _gh_fetch_json_any(url, params, data))
+    params: Optional[Dict[str, Any]] = None,
+    data: Optional[Dict[str, Any]] = None,
+) -> Dict[str, Any]:
+    return cast(Dict[str, Any], _gh_fetch_json_any(url, params, data))
 
 
-def gh_graphql(query: str, **kwargs: Any) -> dict[str, Any]:
+def gh_graphql(query: str, **kwargs: Any) -> Dict[str, Any]:
     rc = gh_fetch_url(
         "https://api.github.com/graphql",
         data={"query": query, "variables": kwargs},
@@ -136,12 +136,12 @@ def gh_graphql(query: str, **kwargs: Any) -> dict[str, Any]:
         raise RuntimeError(
             f"GraphQL query {query}, args {kwargs} failed: {rc['errors']}"
         )
-    return cast(dict[str, Any], rc)
+    return cast(Dict[str, Any], rc)
 
 
 def _gh_post_comment(
     url: str, comment: str, dry_run: bool = False
-) -> list[dict[str, Any]]:
+) -> List[Dict[str, Any]]:
     if dry_run:
         print(comment)
         return []
@@ -150,7 +150,7 @@ def _gh_post_comment(
 
 def gh_post_pr_comment(
     org: str, repo: str, pr_num: int, comment: str, dry_run: bool = False
-) -> list[dict[str, Any]]:
+) -> List[Dict[str, Any]]:
     return _gh_post_comment(
         f"{GITHUB_API_URL}/repos/{org}/{repo}/issues/{pr_num}/comments",
         comment,
@@ -160,7 +160,7 @@ def gh_post_pr_comment(
 
 def gh_post_commit_comment(
     org: str, repo: str, sha: str, comment: str, dry_run: bool = False
-) -> list[dict[str, Any]]:
+) -> List[Dict[str, Any]]:
     return _gh_post_comment(
         f"{GITHUB_API_URL}/repos/{org}/{repo}/commits/{sha}/comments",
         comment,
@@ -220,8 +220,8 @@ def gh_update_pr_state(org: str, repo: str, pr_num: int, state: str = "open") ->
 
 
 def gh_query_issues_by_labels(
-    org: str, repo: str, labels: list[str], state: str = "open"
-) -> list[dict[str, Any]]:
+    org: str, repo: str, labels: List[str], state: str = "open"
+) -> List[Dict[str, Any]]:
     url = f"{GITHUB_API_URL}/repos/{org}/{repo}/issues"
     return gh_fetch_json(
         url, method="GET", params={"labels": ",".join(labels), "state": state}

.github/scripts/gitutils.py

@@ -4,10 +4,20 @@ import os
 import re
 import tempfile
 from collections import defaultdict
-from collections.abc import Iterator
 from datetime import datetime
 from functools import wraps
-from typing import Any, Callable, cast, Optional, TypeVar, Union
+from typing import (
+    Any,
+    Callable,
+    cast,
+    Dict,
+    Iterator,
+    List,
+    Optional,
+    Tuple,
+    TypeVar,
+    Union,
+)
 
 
 T = TypeVar("T")
@@ -25,17 +35,17 @@ def get_git_repo_dir() -> str:
     return os.getenv("GIT_REPO_DIR", str(Path(__file__).resolve().parents[2]))
 
 
-def fuzzy_list_to_dict(items: list[tuple[str, str]]) -> dict[str, list[str]]:
+def fuzzy_list_to_dict(items: List[Tuple[str, str]]) -> Dict[str, List[str]]:
     """
     Converts list to dict preserving elements with duplicate keys
     """
-    rc: dict[str, list[str]] = defaultdict(list)
+    rc: Dict[str, List[str]] = defaultdict(list)
     for key, val in items:
         rc[key].append(val)
     return dict(rc)
 
 
-def _check_output(items: list[str], encoding: str = "utf-8") -> str:
+def _check_output(items: List[str], encoding: str = "utf-8") -> str:
     from subprocess import CalledProcessError, check_output, STDOUT
 
     try:
@@ -85,7 +95,7 @@ class GitCommit:
         return item in self.body or item in self.title
 
 
-def parse_fuller_format(lines: Union[str, list[str]]) -> GitCommit:
+def parse_fuller_format(lines: Union[str, List[str]]) -> GitCommit:
     """
     Expect commit message generated using `--format=fuller --date=unix` format, i.e.:
         commit <sha1>
@@ -132,13 +142,13 @@ class GitRepo:
             print(f"+ git -C {self.repo_dir} {' '.join(args)}")
         return _check_output(["git", "-C", self.repo_dir] + list(args))
 
-    def revlist(self, revision_range: str) -> list[str]:
+    def revlist(self, revision_range: str) -> List[str]:
         rc = self._run_git("rev-list", revision_range, "--", ".").strip()
         return rc.split("\n") if len(rc) > 0 else []
 
     def branches_containing_ref(
         self, ref: str, *, include_remote: bool = True
-    ) -> list[str]:
+    ) -> List[str]:
         rc = (
             self._run_git("branch", "--remote", "--contains", ref)
             if include_remote
@@ -179,7 +189,7 @@ class GitRepo:
     def get_merge_base(self, from_ref: str, to_ref: str) -> str:
         return self._run_git("merge-base", from_ref, to_ref).strip()
 
-    def patch_id(self, ref: Union[str, list[str]]) -> list[tuple[str, str]]:
+    def patch_id(self, ref: Union[str, List[str]]) -> List[Tuple[str, str]]:
         is_list = isinstance(ref, list)
         if is_list:
             if len(ref) == 0:
@@ -188,9 +198,9 @@ class GitRepo:
         rc = _check_output(
             ["sh", "-c", f"git -C {self.repo_dir} show {ref}|git patch-id --stable"]
         ).strip()
-        return [cast(tuple[str, str], x.split(" ", 1)) for x in rc.split("\n")]
+        return [cast(Tuple[str, str], x.split(" ", 1)) for x in rc.split("\n")]
 
-    def commits_resolving_gh_pr(self, pr_num: int) -> list[str]:
+    def commits_resolving_gh_pr(self, pr_num: int) -> List[str]:
         owner, name = self.gh_owner_and_name()
         msg = f"Pull Request resolved: https://github.com/{owner}/{name}/pull/{pr_num}"
         rc = self._run_git("log", "--format=%H", "--grep", msg).strip()
@@ -209,7 +219,7 @@ class GitRepo:
 
     def compute_branch_diffs(
         self, from_branch: str, to_branch: str
-    ) -> tuple[list[str], list[str]]:
+    ) -> Tuple[List[str], List[str]]:
         """
         Returns list of commmits that are missing in each other branch since their merge base
         Might be slow if merge base is between two branches is pretty far off
@@ -301,14 +311,14 @@ class GitRepo:
     def remote_url(self) -> str:
         return self._run_git("remote", "get-url", self.remote)
 
-    def gh_owner_and_name(self) -> tuple[str, str]:
+    def gh_owner_and_name(self) -> Tuple[str, str]:
         url = os.getenv("GIT_REMOTE_URL", None)
         if url is None:
             url = self.remote_url()
         rc = RE_GITHUB_URL_MATCH.match(url)
         if rc is None:
             raise RuntimeError(f"Unexpected url format {url}")
-        return cast(tuple[str, str], rc.groups())
+        return cast(Tuple[str, str], rc.groups())
 
     def commit_message(self, ref: str) -> str:
         return self._run_git("log", "-1", "--format=%B", ref)
@@ -356,7 +366,7 @@ class PeekableIterator(Iterator[str]):
         return rc
 
 
-def patterns_to_regex(allowed_patterns: list[str]) -> Any:
+def patterns_to_regex(allowed_patterns: List[str]) -> Any:
     """
     pattern is glob-like, i.e. the only special sequences it has are:
       - ? - matches single character
@@ -427,7 +437,7 @@ def retries_decorator(
 ) -> Callable[[Callable[..., T]], Callable[..., T]]:
     def decorator(f: Callable[..., T]) -> Callable[..., T]:
         @wraps(f)
-        def wrapper(*args: list[Any], **kwargs: dict[str, Any]) -> T:
+        def wrapper(*args: List[Any], **kwargs: Dict[str, Any]) -> T:
             for idx in range(num_retries):
                 try:
                     return f(*args, **kwargs)

.github/scripts/label_utils.py

@@ -2,7 +2,7 @@
 
 import json
 from functools import lru_cache
-from typing import Any, TYPE_CHECKING, Union
+from typing import Any, List, Tuple, TYPE_CHECKING, Union
 
 from github_utils import gh_fetch_url_and_headers, GitHubComment
 
@@ -28,14 +28,14 @@ https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for
 """
 
 
-def request_for_labels(url: str) -> tuple[Any, Any]:
+def request_for_labels(url: str) -> Tuple[Any, Any]:
     headers = {"Accept": "application/vnd.github.v3+json"}
     return gh_fetch_url_and_headers(
         url, headers=headers, reader=lambda x: x.read().decode("utf-8")
     )
 
 
-def update_labels(labels: list[str], info: str) -> None:
+def update_labels(labels: List[str], info: str) -> None:
     labels_json = json.loads(info)
     labels.extend([x["name"] for x in labels_json])
 
@@ -56,10 +56,10 @@ def get_last_page_num_from_header(header: Any) -> int:
 
 
 @lru_cache
-def gh_get_labels(org: str, repo: str) -> list[str]:
+def gh_get_labels(org: str, repo: str) -> List[str]:
     prefix = f"https://api.github.com/repos/{org}/{repo}/labels?per_page=100"
     header, info = request_for_labels(prefix + "&page=1")
-    labels: list[str] = []
+    labels: List[str] = []
     update_labels(labels, info)
 
     last_page = get_last_page_num_from_header(header)
@@ -74,7 +74,7 @@ def gh_get_labels(org: str, repo: str) -> list[str]:
 
 
 def gh_add_labels(
-    org: str, repo: str, pr_num: int, labels: Union[str, list[str]], dry_run: bool
+    org: str, repo: str, pr_num: int, labels: Union[str, List[str]], dry_run: bool
 ) -> None:
     if dry_run:
         print(f"Dryrun: Adding labels {labels} to PR {pr_num}")
@@ -97,7 +97,7 @@ def gh_remove_label(
     )
 
 
-def get_release_notes_labels(org: str, repo: str) -> list[str]:
+def get_release_notes_labels(org: str, repo: str) -> List[str]:
     return [
         label
         for label in gh_get_labels(org, repo)