Assorted fixes

ghstack-source-id: 1591c4dddcf9d828191ed7bb54ec98669e074816
Pull-Request: https://github.com/pytorch/pytorch/pull/155866

.bazelrc

@@ -2,7 +2,7 @@ build --cxxopt=--std=c++17
 build --copt=-I.
 # Bazel does not support including its cc_library targets as system
 # headers. We work around this for generated code
-# (e.g. torch/headeronly/macros/cmake_macros.h) by making the generated directory a
+# (e.g. c10/macros/cmake_macros.h) by making the generated directory a
 # system include path.
 build --copt=-isystem --copt bazel-out/k8-fastbuild/bin
 build --copt=-isystem --copt bazel-out/darwin-fastbuild/bin

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -4,7 +4,7 @@ set -eux -o pipefail
 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
 if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+    export TORCH_CUDA_ARCH_LIST="9.0;10.0;12.0"
 fi
 
 SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -79,7 +79,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     os.system(f"unzip {wheel_path} -d {folder}/tmp")
     libs_to_copy = [
         "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
-        "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
         "/usr/local/cuda/lib64/libcudnn.so.9",
         "/usr/local/cuda/lib64/libcublas.so.12",
         "/usr/local/cuda/lib64/libcublasLt.so.12",
@@ -89,7 +88,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
         "/usr/local/cuda/lib64/libcusparseLt.so.0",
         "/usr/local/cuda/lib64/libcusolver.so.11",
         "/usr/local/cuda/lib64/libcurand.so.10",
-        "/usr/local/cuda/lib64/libnccl.so.2",
         "/usr/local/cuda/lib64/libnvJitLink.so.12",
         "/usr/local/cuda/lib64/libnvrtc.so.12",
         "/usr/local/cuda/lib64/libcudnn_adv.so.9",

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -438,7 +438,9 @@ def build_torchvision(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -493,7 +495,9 @@ def build_torchdata(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -549,7 +553,9 @@ def build_torchtext(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -607,7 +613,9 @@ def build_torchaudio(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 

.ci/caffe2/test.sh

@@ -5,7 +5,7 @@ source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
 
 if [[ ${BUILD_ENVIRONMENT} == *onnx* ]]; then
   pip install click mock tabulate networkx==2.0
-  pip -q install "file:///var/lib/jenkins/workspace/third_party/onnx#egg=onnx"
+  pip -q install --user "file:///var/lib/jenkins/workspace/third_party/onnx#egg=onnx"
 fi
 
 # Skip tests in environments where they are not built/applicable
@@ -147,8 +147,8 @@ export DNNL_MAX_CPU_ISA=AVX2
 if [[ "${SHARD_NUMBER:-1}" == "1" ]]; then
   # TODO(sdym@meta.com) remove this when the linked issue resolved.
   # py is temporary until https://github.com/Teemu/pytest-sugar/issues/241 is fixed
-  pip install py==1.11.0
-  pip install pytest-sugar
+  pip install --user py==1.11.0
+  pip install --user pytest-sugar
   # NB: Warnings are disabled because they make it harder to see what
   # the actual erroring test is
   "$PYTHON" \

.ci/docker/README.md

@@ -36,104 +36,3 @@ See `build.sh` for valid build environments (it's the giant switch).
 # Set flags (see build.sh) and build image
 sudo bash -c 'TRITON=1 ./build.sh pytorch-linux-bionic-py3.8-gcc9 -t myimage:latest
 ```
-
-## [Guidance] Adding a New Base Docker Image
-
-### Background
-
-The base Docker images in directory `.ci/docker/` are built by the `docker-builds.yml` workflow. Those images are used throughout the PyTorch CI/CD pipeline. You should only create or modify a base Docker image if you need specific environment changes or dependencies before building PyTorch on CI.
-
-1. **Automatic Rebuilding**:
-   - The Docker image building process is triggered automatically when changes are made to files in the `.ci/docker/*` directory
-   - This ensures all images stay up-to-date with the latest dependencies and configurations
-
-2. **Image Reuse in PyTorch Build Workflows** (example: linux-build):
-   - The images generated by `docker-builds.yml` are reused in `_linux-build.yml` through the `calculate-docker-image` step
-   - The `_linux-build.yml` workflow:
-     - Pulls the Docker image determined by the `calculate-docker-image` step
-     - Runs a Docker container with that image
-     - Executes `.ci/pytorch/build.sh` inside the container to build PyTorch
-
-3. **Usage in Test Workflows** (example: linux-test):
-   - The same Docker images are also used in `_linux-test.yml` for running tests
-   - The `_linux-test.yml` workflow follows a similar pattern:
-     - It uses the `calculate-docker-image` step to determine which Docker image to use
-     - It pulls the Docker image and runs a container with that image
-     - It installs the wheels from the artifacts generated by PyTorch build jobs
-     - It executes test scripts (like `.ci/pytorch/test.sh` or `.ci/pytorch/multigpu-test.sh`) inside the container
-
-### Understanding File Purposes
-
-#### `.ci/docker/build.sh` vs `.ci/pytorch/build.sh`
-- **`.ci/docker/build.sh`**:
-  - Used for building base Docker images
-  - Executed by the `docker-builds.yml` workflow to pre-build Docker images for CI
-  - Contains configurations for different Docker build environments
-
-- **`.ci/pytorch/build.sh`**:
-  - Used for building PyTorch inside a Docker container
-  - Called by workflows like `_linux-build.yml` after the Docker container is started
-  - Builds PyTorch wheels and other artifacts
-
-#### `.ci/docker/ci_commit_pins/` vs `.github/ci_commit_pins`
-- **`.ci/docker/ci_commit_pins/`**:
-  - Used for pinning dependency versions during base Docker image building
-  - Ensures consistent environments for building PyTorch
-  - Changes here trigger base Docker image rebuilds
-
-- **`.github/ci_commit_pins`**:
-  - Used for pinning dependency versions during PyTorch building and tests
-  - Ensures consistent dependencies for PyTorch across different builds
-  - Used by build scripts running inside Docker containers
-
-### Step-by-Step Guide for Adding a New Base Docker Image
-
-#### 1. Add Pinned Commits (If Applicable)
-
-We use pinned commits for build stability. The `nightly.yml` workflow checks and updates pinned commits for certain repository dependencies daily.
-
-If your new Docker image needs a library installed from a specific pinned commit or built from source:
-
-1. Add the repository you want to track in `nightly.yml` and `merge-rules.yml`
-2. Add the initial pinned commit in `.ci/docker/ci_commit_pins/`. The text filename should match the one defined in step 1
-
-#### 2. Configure the Base Docker Image
-1. **Add new Base Docker image configuration** (if applicable):
-
-   Add the configuration in `.ci/docker/build.sh`. For example:
-   ```bash
-   pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-new1)
-     CUDA_VERSION=12.8.1
-     ANACONDA_PYTHON_VERSION=3.12
-     GCC_VERSION=11
-     VISION=yes
-     KATEX=yes
-     UCX_COMMIT=${_UCX_COMMIT}
-     UCC_COMMIT=${_UCC_COMMIT}
-     TRITON=yes
-     NEW_ARG_1=yes
-     ;;
-   ```
-
-2. **Add build arguments to Docker build command**:
-
-   If you're introducing a new argument to the Docker build, make sure to add it in the Docker build step in `.ci/docker/build.sh`:
-   ```bash
-   docker build \
-      ....
-      --build-arg "NEW_ARG_1=${NEW_ARG_1}"
-   ```
-
-3. **Update Dockerfile logic**:
-
-   Update the Dockerfile to use the new argument. For example, in `ubuntu/Dockerfile`:
-   ```dockerfile
-   ARG NEW_ARG_1
-   # Set up environment for NEW_ARG_1
-   RUN if [ -n "${NEW_ARG_1}" ]; then bash ./do_something.sh; fi
-   ```
-
-4. **Add the Docker configuration** in `.github/workflows/docker-builds.yml`:
-
-   The `docker-builds.yml` workflow pre-builds the Docker images whenever changes occur in the `.ci/docker/` directory. This includes the
-   pinned commit updates.

.ci/docker/build.sh

@@ -52,8 +52,6 @@ fi
 
 if [[ "$image" == *-jammy* ]]; then
   UBUNTU_VERSION=22.04
-elif [[ "$image" == *-noble* ]]; then
-  UBUNTU_VERSION=24.04
 elif [[ "$image" == *ubuntu* ]]; then
   extract_version_from_image_name ubuntu UBUNTU_VERSION
 fi
@@ -91,18 +89,9 @@ tag=$(echo $image | awk -F':' '{print $2}')
 # configuration, so we hardcode everything here rather than do it
 # from scratch
 case "$tag" in
-  pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11)
-    CUDA_VERSION=12.4
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
-    VISION=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    TRITON=yes
-    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
@@ -113,6 +102,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -124,6 +114,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=9
     VISION=yes
@@ -135,6 +126,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.13
     GCC_VERSION=9
     VISION=yes
@@ -144,18 +136,56 @@ case "$tag" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=11
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
     VISION=yes
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.13
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -176,12 +206,32 @@ case "$tag" in
     VISION=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-jammy-rocm-n-py3-benchmarks | pytorch-linux-noble-rocm-n-py3)
-    if [[ $tag =~ "jammy" ]]; then
-      ANACONDA_PYTHON_VERSION=3.10
-    else
-      ANACONDA_PYTHON_VERSION=3.12
-    fi
+  pytorch-linux-jammy-py3.11-clang12)
+    ANACONDA_PYTHON_VERSION=3.11
+    CLANG_VERSION=12
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3.9-gcc9)
+    ANACONDA_PYTHON_VERSION=3.9
+    GCC_VERSION=9
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-rocm-n-1-py3)
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=11
+    VISION=yes
+    ROCM_VERSION=6.3
+    NINJA_VERSION=1.9.0
+    TRITON=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-rocm-n-py3)
+    ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     ROCM_VERSION=6.4
@@ -190,21 +240,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    if [[ $tag =~ "benchmarks" ]]; then
-      INDUCTOR_BENCHMARKS=yes
-    fi
-    ;;
-  pytorch-linux-noble-rocm-alpha-py3)
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=11
-    VISION=yes
-    ROCM_VERSION=7.0
-    NINJA_VERSION=1.9.0
-    TRITON=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
+    INDUCTOR_BENCHMARKS=yes
     ;;
   pytorch-linux-jammy-xpu-2025.0-py3)
     ANACONDA_PYTHON_VERSION=3.9
@@ -222,7 +258,7 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
+    pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
@@ -234,10 +270,22 @@ case "$tag" in
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
     ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
     ;;
+  pytorch-linux-jammy-py3-clang12-asan)
+    ANACONDA_PYTHON_VERSION=3.9
+    CLANG_VERSION=12
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3-clang15-asan)
+    ANACONDA_PYTHON_VERSION=3.10
+    CLANG_VERSION=15
+    VISION=yes
+    ;;
   pytorch-linux-jammy-py3-clang18-asan)
     ANACONDA_PYTHON_VERSION=3.10
     CLANG_VERSION=18
@@ -285,8 +333,6 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
-    CONDA_CMAKE=yes
-    OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific
     SKIP_LLVM_SRC_BUILD_INSTALL=yes
@@ -296,8 +342,6 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
-    CONDA_CMAKE=yes
-    OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific
     SKIP_LLVM_SRC_BUILD_INSTALL=yes
@@ -312,6 +356,7 @@ case "$tag" in
     fi
     if [[ "$image" == *cuda* ]]; then
       extract_version_from_image_name cuda CUDA_VERSION
+      extract_version_from_image_name cudnn CUDNN_VERSION
     fi
     if [[ "$image" == *rocm* ]]; then
       extract_version_from_image_name rocm ROCM_VERSION
@@ -363,6 +408,9 @@ docker build \
        --build-arg "PYTHON_VERSION=${PYTHON_VERSION}" \
        --build-arg "GCC_VERSION=${GCC_VERSION}" \
        --build-arg "CUDA_VERSION=${CUDA_VERSION}" \
+       --build-arg "CUDNN_VERSION=${CUDNN_VERSION}" \
+       --build-arg "TENSORRT_VERSION=${TENSORRT_VERSION}" \
+       --build-arg "GRADLE_VERSION=${GRADLE_VERSION}" \
        --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
@@ -380,7 +428,6 @@ docker build \
        --build-arg "XPU_VERSION=${XPU_VERSION}" \
        --build-arg "UNINSTALL_DILL=${UNINSTALL_DILL}" \
        --build-arg "ACL=${ACL:-}" \
-       --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
        --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \

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

@@ -1 +1 @@
-v2.27.5-1
+v2.27.3-1

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-f7888497a1eb9e98d4c07537f0d0bcfe180d1363
+c8757738a7418249896224430ce84888e8ecdd79

.ci/docker/common/common_utils.sh

@@ -23,10 +23,6 @@ conda_install() {
   as_jenkins conda install -q -n py_$ANACONDA_PYTHON_VERSION -y python="$ANACONDA_PYTHON_VERSION" $*
 }
 
-conda_install_through_forge() {
-  as_jenkins conda install -c conda-forge -q -n py_$ANACONDA_PYTHON_VERSION -y python="$ANACONDA_PYTHON_VERSION" $*
-}
-
 conda_run() {
   as_jenkins conda run -n py_$ANACONDA_PYTHON_VERSION --no-capture-output $*
 }

.ci/docker/common/install_base.sh

@@ -15,9 +15,6 @@ install_ubuntu() {
   elif [[ "$UBUNTU_VERSION" == "22.04"* ]]; then
     cmake3="cmake=3.22*"
     maybe_libiomp_dev=""
-  elif [[ "$UBUNTU_VERSION" == "24.04"* ]]; then
-    cmake3="cmake=3.28*"
-    maybe_libiomp_dev=""
   else
     cmake3="cmake=3.5*"
     maybe_libiomp_dev="libiomp-dev"

.ci/docker/common/install_conda.sh

@@ -4,8 +4,12 @@ set -ex
 
 # Optionally install conda
 if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
-  BASE_URL="https://github.com/conda-forge/miniforge/releases/latest/download"  # @lint-ignore
-  CONDA_FILE="Miniforge3-Linux-$(uname -m).sh"
+  BASE_URL="https://repo.anaconda.com/miniconda"
+  CONDA_FILE="Miniconda3-latest-Linux-x86_64.sh"
+  if [[ $(uname -m) == "aarch64" ]] || [[ "$BUILD_ENVIRONMENT" == *xpu* ]] || [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
+    BASE_URL="https://github.com/conda-forge/miniforge/releases/latest/download"  # @lint-ignore
+    CONDA_FILE="Miniforge3-Linux-$(uname -m).sh"
+  fi
 
   MAJOR_PYTHON_VERSION=$(echo "$ANACONDA_PYTHON_VERSION" | cut -d . -f 1)
   MINOR_PYTHON_VERSION=$(echo "$ANACONDA_PYTHON_VERSION" | cut -d . -f 2)
@@ -17,6 +21,7 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
       exit 1
       ;;
   esac
+
   mkdir -p /opt/conda
   chown jenkins:jenkins /opt/conda
 
@@ -65,10 +70,10 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
   fi
 
   # Install PyTorch conda deps, as per https://github.com/pytorch/pytorch README
-  if [[ $(uname -m) != "aarch64" ]]; then
-    pip_install mkl==2024.2.0
-    pip_install mkl-static==2024.2.0
-    pip_install mkl-include==2024.2.0
+  if [[ $(uname -m) == "aarch64" ]]; then
+    conda_install "openblas==0.3.29=*openmp*"
+  else
+    conda_install "mkl=2021.4.0 mkl-include=2021.4.0"
   fi
 
   # Install llvm-8 as it is required to compile llvmlite-0.30.0 from source
@@ -82,10 +87,6 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
     conda_run ${SCRIPT_FOLDER}/install_magma_conda.sh $(cut -f1-2 -d'.' <<< ${CUDA_VERSION})
   fi
 
-  if [[ "$UBUNTU_VERSION" == "24.04"* ]] ; then
-    conda_install_through_forge libstdcxx-ng=14
-  fi
-
   # Install some other packages, including those needed for Python test reporting
   pip_install -r /opt/conda/requirements-ci.txt
 

.ci/docker/common/install_cpython.sh

@@ -3,10 +3,11 @@
 set -uex -o pipefail
 
 PYTHON_DOWNLOAD_URL=https://www.python.org/ftp/python
+PYTHON_DOWNLOAD_GITHUB_BRANCH=https://github.com/python/cpython/archive/refs/heads  # @lint-ignore
 GET_PIP_URL=https://bootstrap.pypa.io/get-pip.py
 
 # Python versions to be installed in /opt/$VERSION_NO
-CPYTHON_VERSIONS=${CPYTHON_VERSIONS:-"3.9.0 3.10.1 3.11.0 3.12.0 3.13.0 3.13.0t 3.14.0 3.14.0t"}
+CPYTHON_VERSIONS=${CPYTHON_VERSIONS:-"3.9.0 3.10.1 3.11.0 3.12.0 3.13.0 3.13.0t"}
 
 function check_var {
     if [ -z "$1" ]; then
@@ -23,8 +24,9 @@ function do_cpython_build {
     tar -xzf Python-$py_ver.tgz
 
     local additional_flags=""
-    if [[ "$py_ver" == *"t" ]]; then
+    if [ "$py_ver" == "3.13.0t" ]; then
         additional_flags=" --disable-gil"
+        mv cpython-3.13/ cpython-3.13t/
     fi
 
     pushd $py_folder
@@ -66,29 +68,32 @@ function do_cpython_build {
         ln -s pip3 ${prefix}/bin/pip
     fi
     # install setuptools since python 3.12 is required to use distutils
-    # packaging is needed to create symlink since wheel no longer provides needed information
-    ${prefix}/bin/pip install packaging==25.0 wheel==0.45.1 setuptools==80.9.0
-    local abi_tag=$(${prefix}/bin/python -c "from packaging.tags import interpreter_name, interpreter_version; import sysconfig ; from sysconfig import get_config_var; print('{0}{1}-{0}{1}{2}'.format(interpreter_name(), interpreter_version(), 't' if sysconfig.get_config_var('Py_GIL_DISABLED') else ''))")
+    ${prefix}/bin/pip install wheel==0.34.2 setuptools==68.2.2
+    local abi_tag=$(${prefix}/bin/python -c "from wheel.pep425tags import get_abbr_impl, get_impl_ver, get_abi_tag; print('{0}{1}-{2}'.format(get_abbr_impl(), get_impl_ver(), get_abi_tag()))")
     ln -sf ${prefix} /opt/python/${abi_tag}
 }
 
 function build_cpython {
     local py_ver=$1
     check_var $py_ver
-    local py_suffix=$py_ver
-    local py_folder=$py_ver
+    check_var $PYTHON_DOWNLOAD_URL
+    local py_ver_folder=$py_ver
 
-    # Special handling for nogil
-    if [[ "${py_ver}" == *"t" ]]; then
-        py_suffix=${py_ver::-1}
-        py_folder=$py_suffix
+    if [ "$py_ver" = "3.13.0t" ]; then
+        PY_VER_SHORT="3.13"
+        PYT_VER_SHORT="3.13t"
+        check_var $PYTHON_DOWNLOAD_GITHUB_BRANCH
+        wget $PYTHON_DOWNLOAD_GITHUB_BRANCH/$PY_VER_SHORT.tar.gz -O Python-$py_ver.tgz
+        do_cpython_build $py_ver cpython-$PYT_VER_SHORT
+    elif [ "$py_ver" = "3.13.0" ]; then
+        PY_VER_SHORT="3.13"
+        check_var $PYTHON_DOWNLOAD_GITHUB_BRANCH
+        wget $PYTHON_DOWNLOAD_GITHUB_BRANCH/$PY_VER_SHORT.tar.gz -O Python-$py_ver.tgz
+        do_cpython_build $py_ver cpython-$PY_VER_SHORT
+    else
+        wget -q $PYTHON_DOWNLOAD_URL/$py_ver_folder/Python-$py_ver.tgz
+        do_cpython_build $py_ver Python-$py_ver
     fi
-    # Only b3 is available now
-    if [ "$py_suffix" == "3.14.0" ]; then
-        py_suffix="3.14.0b3"
-    fi
-    wget -q $PYTHON_DOWNLOAD_URL/$py_folder/Python-$py_suffix.tgz -O Python-$py_ver.tgz
-    do_cpython_build $py_ver Python-$py_suffix
 
     rm -f Python-$py_ver.tgz
 }

.ci/docker/common/install_cuda.sh

@@ -10,8 +10,6 @@ else
   arch_path='sbsa'
 fi
 
-NVSHMEM_VERSION=3.3.9
-
 function install_cuda {
   version=$1
   runfile=$2
@@ -42,65 +40,13 @@ function install_cudnn {
   rm -rf tmp_cudnn
 }
 
-function install_nvshmem {
-  cuda_major_version=$1      # e.g. "12"
-  nvshmem_version=$2         # e.g. "3.3.9"
-
-  case "${arch_path}" in
-    sbsa)
-      dl_arch="aarch64"
-      ;;
-    x86_64)
-      dl_arch="x64"
-      ;;
-    *)
-      dl_arch="${arch}"
-      ;;
-  esac
-
-  tmpdir="tmp_nvshmem"
-  mkdir -p "${tmpdir}" && cd "${tmpdir}"
-
-  # nvSHMEM license: https://docs.nvidia.com/nvshmem/api/sla.html
-  filename="libnvshmem_cuda${cuda_major_version}-linux-${arch_path}-${nvshmem_version}"
-  url="https://developer.download.nvidia.com/compute/redist/nvshmem/${nvshmem_version}/builds/cuda${cuda_major_version}/txz/agnostic/${dl_arch}/${filename}.tar.gz"
-
-  # download, unpack, install
-  wget -q "${url}"
-  tar xf "${filename}.tar.gz"
-  cp -a "libnvshmem/include/"* /usr/local/cuda/include/
-  cp -a "libnvshmem/lib/"*     /usr/local/cuda/lib64/
-
-  # cleanup
-  cd ..
-  rm -rf "${tmpdir}"
-
-  echo "nvSHMEM ${nvshmem_version} for CUDA ${cuda_major_version} (${arch_path}) installed."
-}
-
-function install_124 {
-  CUDNN_VERSION=9.1.0.70
-  echo "Installing CUDA 12.4.1 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.6.2"
-  install_cuda 12.4.1 cuda_12.4.1_550.54.15_linux
-
-  install_cudnn 12 $CUDNN_VERSION
-
-  CUDA_VERSION=12.4 bash install_nccl.sh
-
-  CUDA_VERSION=12.4 bash install_cusparselt.sh
-
-  ldconfig
-}
-
 function install_126 {
   CUDNN_VERSION=9.10.2.21
-  echo "Installing CUDA 12.6.3 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
+  echo "Installing CUDA 12.6.3 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.7.1"
   install_cuda 12.6.3 cuda_12.6.3_560.35.05_linux
 
   install_cudnn 12 $CUDNN_VERSION
 
-  install_nvshmem 12 $NVSHMEM_VERSION
-
   CUDA_VERSION=12.6 bash install_nccl.sh
 
   CUDA_VERSION=12.6 bash install_cusparselt.sh
@@ -110,15 +56,13 @@ function install_126 {
 
 function install_129 {
   CUDNN_VERSION=9.10.2.21
-  echo "Installing CUDA 12.9.1 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
+  echo "Installing CUDA 12.9.1 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.7.1"
   # install CUDA 12.9.1 in the same container
   install_cuda 12.9.1 cuda_12.9.1_575.57.08_linux
 
   # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
   install_cudnn 12 $CUDNN_VERSION
 
-  install_nvshmem 12 $NVSHMEM_VERSION
-
   CUDA_VERSION=12.9 bash install_nccl.sh
 
   CUDA_VERSION=12.9 bash install_cusparselt.sh
@@ -126,40 +70,6 @@ function install_129 {
   ldconfig
 }
 
-function prune_124 {
-  echo "Pruning CUDA 12.4"
-  #####################################################################################
-  # CUDA 12.4 prune static libs
-  #####################################################################################
-  export NVPRUNE="/usr/local/cuda-12.4/bin/nvprune"
-  export CUDA_LIB_DIR="/usr/local/cuda-12.4/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
-  if [[ -n "$OVERRIDE_GENCODE_CUDNN" ]]; then
-      export GENCODE_CUDNN=$OVERRIDE_GENCODE_CUDNN
-  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 12.4 prune visual tools
-  #####################################################################################
-  export CUDA_BASE="/usr/local/cuda-12.4/"
-  rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2024.1.0 $CUDA_BASE/nsight-systems-2023.4.4/
-}
-
 function prune_126 {
   echo "Pruning CUDA 12.6"
   #####################################################################################
@@ -196,15 +106,13 @@ function prune_126 {
 
 function install_128 {
   CUDNN_VERSION=9.8.0.87
-  echo "Installing CUDA 12.8.1 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
+  echo "Installing CUDA 12.8.1 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.7.1"
   # install CUDA 12.8.1 in the same container
   install_cuda 12.8.1 cuda_12.8.1_570.124.06_linux
 
   # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
   install_cudnn 12 $CUDNN_VERSION
 
-  install_nvshmem 12 $NVSHMEM_VERSION
-
   CUDA_VERSION=12.8 bash install_nccl.sh
 
   CUDA_VERSION=12.8 bash install_cusparselt.sh
@@ -216,8 +124,6 @@ function install_128 {
 while test $# -gt 0
 do
     case "$1" in
-    12.4) install_124; prune_124
-        ;;
     12.6|12.6.*) install_126; prune_126
         ;;
     12.8|12.8.*) install_128;

.ci/docker/common/install_cudnn.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+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} == "12.9" || ${CUDA_VERSION:0:4} == "12.8" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:4} == "12.6" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:2} == "11" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.1.0.70_cuda11-archive"
+    else
+        print "Unsupported CUDA version ${CUDA_VERSION}"
+        exit 1
+    fi
+    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/${CUDNN_NAME}.tar.xz
+    tar xf ${CUDNN_NAME}.tar.xz
+    cp -a ${CUDNN_NAME}/include/* /usr/local/cuda/include/
+    cp -a ${CUDNN_NAME}/lib/* /usr/local/cuda/lib64/
+    popd
+    rm -rf tmp_cudnn
+    ldconfig
+fi

.ci/docker/common/install_cusparselt.sh

@@ -13,14 +13,6 @@ if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
     fi
     CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.7.1.0-archive"
     curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
-elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
-    arch_path='sbsa'
-    export TARGETARCH=${TARGETARCH:-$(uname -m)}
-    if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
-        arch_path='x86_64'
-    fi
-    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.6.2.3-archive"
-    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
 else
     echo "Not sure which libcusparselt version to install for this ${CUDA_VERSION}"
 fi

.ci/docker/common/install_inductor_benchmark_deps.sh

@@ -15,37 +15,11 @@ function install_timm() {
   commit=$(get_pinned_commit timm)
 
   pip_install "git+https://github.com/huggingface/pytorch-image-models@${commit}"
-}
-
-function install_torchbench() {
-  local commit
-  commit=$(get_pinned_commit torchbench)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  python install.py --continue_on_fail
-
-  # 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
-
-  chown -R jenkins torchbench
+  # Clean up
+  conda_run pip uninstall -y torch torchvision triton
 }
 
 # Pango is needed for weasyprint which is needed for doctr
 conda_install pango
-
-# Stable packages are ok here, just to satisfy TorchBench check
-pip_install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
-
-install_torchbench
 install_huggingface
 install_timm
-
-# Clean up
-conda_run pip uninstall -y torch torchvision torchaudio triton torchao

.ci/docker/common/install_onnx.sh

@@ -20,7 +20,7 @@ pip_install \
 
 pip_install coloredlogs packaging
 pip_install onnxruntime==1.18.1
-pip_install onnxscript==0.3.1
+pip_install onnxscript==0.3.0
 
 # 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/

.ci/docker/common/install_openblas.sh

@@ -4,9 +4,8 @@
 set -ex
 
 cd /
-git clone https://github.com/OpenMathLib/OpenBLAS.git -b "${OPENBLAS_VERSION:-v0.3.30}" --depth 1 --shallow-submodules
+git clone https://github.com/OpenMathLib/OpenBLAS.git -b "${OPENBLAS_VERSION:-v0.3.29}" --depth 1 --shallow-submodules
 
-OPENBLAS_CHECKOUT_DIR="OpenBLAS"
 OPENBLAS_BUILD_FLAGS="
 NUM_THREADS=128
 USE_OPENMP=1
@@ -14,8 +13,9 @@ NO_SHARED=0
 DYNAMIC_ARCH=1
 TARGET=ARMV8
 CFLAGS=-O3
-BUILD_BFLOAT16=1
 "
 
+OPENBLAS_CHECKOUT_DIR="OpenBLAS"
+
 make -j8 ${OPENBLAS_BUILD_FLAGS} -C ${OPENBLAS_CHECKOUT_DIR}
 make -j8 ${OPENBLAS_BUILD_FLAGS} install -C ${OPENBLAS_CHECKOUT_DIR}

.ci/docker/common/install_rocm.sh

@@ -8,11 +8,9 @@ ver() {
 
 install_ubuntu() {
     apt-get update
-    # gpg-agent is not available by default
-    apt-get install -y --no-install-recommends gpg-agent
-    if [[ $(ver $UBUNTU_VERSION) -ge $(ver 22.04) ]]; then
-        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' \
-            | sudo tee /etc/apt/preferences.d/rocm-pin-600
+    if [[ $UBUNTU_VERSION == 20.04 ]]; then
+      # gpg-agent is not available by default on 20.04
+      apt-get install -y --no-install-recommends gpg-agent
     fi
     apt-get install -y kmod
     apt-get install -y wget
@@ -30,25 +28,16 @@ EOF
 
     # we want the patch version of 6.4 instead
     if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-        ROCM_VERSION="${ROCM_VERSION}.2"
-    fi
-
-    # Default url values
-    rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
-    amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
-
-    # Special case for ROCM_VERSION == 7.0
-    if [[ $(ver "$ROCM_VERSION") -eq $(ver 7.0) ]]; then
-        rocm_baseurl="https://repo.radeon.com/rocm/apt/7.0_alpha2"
-        amdgpu_baseurl="https://repo.radeon.com/amdgpu/30.10_alpha2/ubuntu"
+        ROCM_VERSION="${ROCM_VERSION}.1"
     fi
 
     # Add amdgpu repository
     UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
-    echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
+    echo "deb [arch=amd64] https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
 
     # Add rocm repository
     wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
+    local rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
     echo "deb [arch=amd64] ${rocm_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/rocm.list
     apt-get update --allow-insecure-repositories
 
@@ -82,33 +71,29 @@ EOF
     done
 
     # ROCm 6.3 had a regression where initializing static code objects had significant overhead
-    # CI no longer builds for ROCm 6.3, but
     # ROCm 6.4 did not yet fix the regression, also HIP branch names are different
-    if [[ $(ver $ROCM_VERSION) -ge $(ver 6.4) ]] && [[ $(ver $ROCM_VERSION) -lt $(ver 7.0) ]]; then
-        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.2) ]]; then
-            HIP_TAG=rocm-6.4.2
-            CLR_HASH=74d78ba3ac4bac235d02bcb48511c30b5cfdd457  # branch release/rocm-rel-6.4.2-statco-hotfix
-        elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.1) ]]; then
-            HIP_TAG=rocm-6.4.1
-            CLR_HASH=efe6c35790b9206923bfeed1209902feff37f386  # branch release/rocm-rel-6.4.1-statco-hotfix
+    if [[ $(ver $ROCM_VERSION) -ge $(ver 6.3) ]] && [[ $(ver $ROCM_VERSION) -lt $(ver 7.0) ]]; then
+        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.1) ]]; then
+            HIP_BRANCH=release/rocm-rel-6.4
+            VER_STR=6.4
+            VER_PATCH=.1
         elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-            HIP_TAG=rocm-6.4.0
-            CLR_HASH=600f5b0d2baed94d5121e2174a9de0851b040b0c  # branch release/rocm-rel-6.4-statco-hotfix
+            HIP_BRANCH=release/rocm-rel-6.4
+            VER_STR=6.4
+        elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.3) ]]; then
+            HIP_BRANCH=rocm-6.3.x
+            VER_STR=6.3
         fi
         # clr build needs CppHeaderParser but can only find it using conda's python
-        python -m pip install CppHeaderParser
-        git clone https://github.com/ROCm/HIP -b $HIP_TAG
+        /opt/conda/bin/python -m pip install CppHeaderParser
+        git clone https://github.com/ROCm/HIP -b $HIP_BRANCH
         HIP_COMMON_DIR=$(readlink -f HIP)
-        git clone https://github.com/jeffdaily/clr
-        pushd clr
-        git checkout $CLR_HASH
-        popd
+        git clone https://github.com/jeffdaily/clr -b release/rocm-rel-${VER_STR}${VER_PATCH}-statco-hotfix
         mkdir -p clr/build
         pushd clr/build
-        # Need to point CMake to the correct python installation to find CppHeaderParser
-        cmake .. -DPython3_EXECUTABLE=/opt/conda/envs/py_${ANACONDA_PYTHON_VERSION}/bin/python3 -DCLR_BUILD_HIP=ON -DHIP_COMMON_DIR=$HIP_COMMON_DIR
+        cmake .. -DCLR_BUILD_HIP=ON -DHIP_COMMON_DIR=$HIP_COMMON_DIR
         make -j
-        cp hipamd/lib/libamdhip64.so.6.4.* /opt/rocm/lib/libamdhip64.so.6.4.*
+        cp hipamd/lib/libamdhip64.so.${VER_STR}.* /opt/rocm/lib/libamdhip64.so.${VER_STR}.*
         popd
         rm -rf HIP clr
     fi

.ci/docker/common/install_rocm_magma.sh

@@ -5,12 +5,7 @@ set -eou pipefail
 
 function do_install() {
     rocm_version=$1
-    if [[ ${rocm_version} =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
-        # chop off any patch version
-        rocm_version="${rocm_version%.*}"
-    fi
-
-    rocm_version_nodot=${rocm_version//./}
+    rocm_version_nodot=${1//./}
 
     # Version 2.7.2 + ROCm related updates
     MAGMA_VERSION=a1625ff4d9bc362906bd01f805dbbe12612953f6

.ci/docker/common/install_triton.sh

@@ -98,10 +98,6 @@ fi
 if [ -n "${NUMPY_VERSION}" ]; then
   pip_install "numpy==${NUMPY_VERSION}"
 fi
-
-# IMPORTANT: helion needs to be installed without dependencies.
-# It depends on torch and triton. We don't want to install
-# triton and torch from production on Docker CI images
 if [[ "$ANACONDA_PYTHON_VERSION" != 3.9* ]]; then
-  pip_install helion --no-deps
+  pip_install helion
 fi

.ci/docker/common/install_xpu.sh

@@ -34,27 +34,18 @@ function install_ubuntu() {
 
     # The xpu-smi packages
     apt-get install -y flex bison xpu-smi
-
-    if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-        # Compute and Media Runtimes
-        apt-get install -y \
-            intel-opencl-icd intel-level-zero-gpu level-zero \
-            intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
-        # Development Packages
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
-    else # rolling driver
-        apt-get install -y \
-            intel-opencl-icd libze-intel-gpu1 libze1 \
-            intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev
+    # Compute and Media Runtimes
+    apt-get install -y \
+        intel-opencl-icd intel-level-zero-gpu level-zero \
+        intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
+        libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
+        libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
+        mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
+    if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+        apt-get install -y intel-ocloc
     fi
-
+    # Development Packages
+    apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
     # Install Intel Support Packages
     apt-get install -y ${XPU_PACKAGES}
 
@@ -65,10 +56,14 @@ function install_ubuntu() {
 
 function install_rhel() {
     . /etc/os-release
-
-    if [[ ! " 8.8 8.10 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
-        echo "RHEL version ${VERSION_ID} not supported"
-        exit
+    if [[ "${ID}" == "rhel" ]]; then
+        if [[ ! " 8.8 8.9 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
+            echo "RHEL version ${VERSION_ID} not supported"
+            exit
+        fi
+    elif [[ "${ID}" == "almalinux" ]]; then
+        # Workaround for almalinux8 which used by quay.io/pypa/manylinux_2_28_x86_64
+        VERSION_ID="8.8"
     fi
 
     dnf install -y 'dnf-command(config-manager)'
@@ -139,11 +134,11 @@ function install_sles() {
 
 }
 
-# Default use GPU driver rolling releases
-XPU_DRIVER_VERSION=""
-if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-    # Use GPU driver LTS releases
-    XPU_DRIVER_VERSION="/lts/2350"
+# Default use GPU driver LTS releases
+XPU_DRIVER_VERSION="/lts/2350"
+if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+    # Use GPU driver rolling releases
+    XPU_DRIVER_VERSION=""
 fi
 
 # Default use Intel® oneAPI Deep Learning Essentials 2025.0

.ci/docker/libtorch/build.sh

@@ -39,10 +39,6 @@ case ${DOCKER_TAG_PREFIX} in
         DOCKER_GPU_BUILD_ARG=""
         ;;
     rocm*)
-        # we want the patch version of 6.4 instead
-        if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
-        fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete
         PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"

.ci/docker/linter/Dockerfile

@@ -27,7 +27,5 @@ COPY ./common/install_linter.sh install_linter.sh
 RUN bash ./install_linter.sh
 RUN rm install_linter.sh
 
-RUN chown -R jenkins:jenkins /var/lib/jenkins/ci_env
-
 USER jenkins
 CMD ["bash"]

.ci/docker/manywheel/Dockerfile_s390x

@@ -131,8 +131,6 @@ RUN pip3 install flatbuffers && \
   git clone https://github.com/microsoft/onnxruntime && \
   cd onnxruntime && git checkout v1.21.0 && \
   git submodule update --init --recursive && \
-  wget https://github.com/microsoft/onnxruntime/commit/f57db79743c4d1a3553aa05cf95bcd10966030e6.patch && \
-  patch -p1 < f57db79743c4d1a3553aa05cf95bcd10966030e6.patch && \
   ./build.sh --config Release --parallel 0 --enable_pybind \
   --build_wheel --enable_training --enable_training_apis \
   --enable_training_ops --skip_tests --allow_running_as_root \

.ci/docker/manywheel/build.sh

@@ -41,7 +41,7 @@ case ${image} in
         GPU_IMAGE=arm64v8/almalinux:8
         DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13 --build-arg NINJA_VERSION=1.12.1"
         MANY_LINUX_VERSION="2_28_aarch64"
-        OPENBLAS_VERSION="v0.3.30"
+        OPENBLAS_VERSION="v0.3.29"
         ;;
     manylinuxcxx11-abi-builder:cpu-cxx11-abi)
         TARGET=final
@@ -75,10 +75,6 @@ case ${image} in
         DOCKERFILE_SUFFIX="_cuda_aarch64"
         ;;
     manylinux2_28-builder:rocm*)
-        # we want the patch version of 6.4 instead
-        if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
-        fi
         TARGET=rocm_final
         MANY_LINUX_VERSION="2_28"
         DEVTOOLSET_VERSION="11"

.ci/docker/requirements-ci.txt

@@ -16,7 +16,6 @@ click
 #test that import:
 
 coremltools==5.0b5 ; python_version < "3.12"
-coremltools==8.3 ; python_version == "3.12"
 #Description: Apple framework for ML integration
 #Pinned versions: 5.0b5
 #test that import:
@@ -50,7 +49,7 @@ flatbuffers==24.12.23
 hypothesis==5.35.1
 # Pin hypothesis to avoid flakiness: https://github.com/pytorch/pytorch/issues/31136
 #Description: advanced library for generating parametrized tests
-#Pinned versions: 5.35.1
+#Pinned versions: 3.44.6, 4.53.2
 #test that import: test_xnnpack_integration.py, test_pruning_op.py, test_nn.py
 
 junitparser==2.1.1
@@ -63,12 +62,10 @@ lark==0.12.0
 #Pinned versions: 0.12.0
 #test that import:
 
-librosa>=0.6.2 ; python_version < "3.11" and platform_machine != "s390x"
-librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
+librosa>=0.6.2 ; python_version < "3.11"
 #Description: A python package for music and audio analysis
 #Pinned versions: >=0.6.2
 #test that import: test_spectral_ops.py
-#librosa depends on numba; disable it for s390x while numba is disabled too
 
 #mkl #this breaks linux-bionic-rocm4.5-py3.7
 #Description: Intel oneAPI Math Kernel Library
@@ -111,15 +108,13 @@ ninja==1.11.1.3
 #Pinned versions: 1.11.1.3
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
 
-numba==0.49.0 ; python_version < "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
-numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
+numba==0.49.0 ; python_version < "3.9"
+numba==0.55.2 ; python_version == "3.9"
+numba==0.55.2 ; python_version == "3.10"
 #Description: Just-In-Time Compiler for Numerical Functions
 #Pinned versions: 0.54.1, 0.49.0, <=0.49.1
 #test that import: test_numba_integration.py
 #For numba issue see https://github.com/pytorch/pytorch/issues/51511
-#Need release > 0.61.2 for s390x due to https://github.com/numba/numba/pull/10073
 
 #numpy
 #Description: Provides N-dimensional arrays and linear algebra
@@ -223,9 +218,9 @@ pygments==2.15.0
 #Pinned versions: 2.12.0
 #test that import: the doctests
 
-#pyyaml
+#PyYAML
 #Description: data serialization format
-#Pinned versions: 6.0.2
+#Pinned versions:
 #test that import:
 
 #requests
@@ -235,7 +230,7 @@ pygments==2.15.0
 
 #rich
 #Description: rich text and beautiful formatting in the terminal
-#Pinned versions: 14.1.0
+#Pinned versions: 10.9.0
 #test that import:
 
 scikit-image==0.19.3 ; python_version < "3.10"
@@ -309,7 +304,7 @@ pytest-cpp==2.3.0
 #Pinned versions: 2.3.0
 #test that import:
 
-z3-solver==4.15.1.0 ; platform_machine != "s390x"
+z3-solver==4.12.6.0
 #Description: The Z3 Theorem Prover Project
 #Pinned versions:
 #test that import:
@@ -344,7 +339,7 @@ onnx==1.18.0
 #Pinned versions:
 #test that import:
 
-onnxscript==0.3.1
+onnxscript==0.2.6
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -363,11 +358,12 @@ pwlf==2.2.1
 #Pinned versions: 2.2.1
 #test that import: test_sac_estimator.py
 
+
 # To build PyTorch itself
-pyyaml
+astunparse
+PyYAML
 pyzstd
-setuptools>=70.1.0
-six
+setuptools
 
 scons==4.5.2 ; platform_machine == "aarch64"
 
@@ -387,12 +383,6 @@ cmake==4.0.0
 tlparse==0.3.30
 #Description: required for log parsing
 
-cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"
+cuda-bindings>=12.0,<13.0
 #Description: required for testing CUDAGraph::raw_cuda_graph(). See https://nvidia.github.io/cuda-python/cuda-bindings/latest/support.html for how this version was chosen. Note "Any fix in the latest bindings would be backported to the prior major version" means that only the newest version of cuda-bindings will get fixes. Depending on the latest version of 12.x is okay because all 12.y versions will be supported via "CUDA minor version compatibility". Pytorch builds against 13.z versions of cuda toolkit work with 12.x versions of cuda-bindings as well because newer drivers work with old toolkits.
 #test that import: test_cuda.py
-
-setuptools-git-versioning==2.1.0
-scikit-build==0.18.1
-pyre-extensions==0.0.32
-tabulate==0.9.0
-#Description: These package are needed to build FBGEMM and torchrec on PyTorch CI

.ci/docker/requirements-docs.txt

@@ -1,11 +1,11 @@
 sphinx==5.3.0
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 5.3.0
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git@722b7e6f9ca512fcc526ad07d62b3d28c50bb6cd#egg=pytorch_sphinx_theme2
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
 
 # TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
-# but it doesn't seem to work and hangs around idly. The initial thought that it is probably
-# something related to Docker setup. We can investigate this later.
+# but it doesn't seem to work and hangs around idly. The initial thought is probably
+# something related to Docker setup. We can investigate this later
 
 sphinxcontrib.katex==0.8.6
 #Description: This is used to generate PyTorch docs
@@ -19,10 +19,9 @@ sphinx_sitemap==2.6.0
 #Description: This is used to generate sitemap for PyTorch docs
 #Pinned versions: 2.6.0
 
-matplotlib==3.5.3 ; python_version < "3.13"
-matplotlib==3.6.3 ; python_version >= "3.13"
+matplotlib==3.5.3
 #Description: This is used to generate PyTorch docs
-#Pinned versions: 3.6.3 if python > 3.12. Otherwise 3.5.3.
+#Pinned versions: 3.5.3
 
 tensorboard==2.13.0 ; python_version < "3.13"
 tensorboard==2.18.0 ; python_version >= "3.13"
@@ -50,8 +49,8 @@ IPython==8.12.0
 #Pinned versions: 8.12.0
 
 myst-nb==0.17.2
-#Description: This is used to generate PyTorch functorch and torch.compile docs.
-#Pinned versions: 0.17.2
+#Description: This is used to generate PyTorch functorch docs
+#Pinned versions: 0.13.2
 
 # The following are required to build torch.distributed.elastic.rendezvous.etcd* docs
 python-etcd==0.4.5

.ci/docker/triton_version.txt

@@ -1 +1 @@
-3.4.0
+3.3.1

.ci/docker/ubuntu-rocm/Dockerfile

@@ -98,9 +98,8 @@ COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.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 torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -98,9 +98,8 @@ COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.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 torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 ARG TRITON
 ARG TRITON_CPU
@@ -148,12 +147,6 @@ RUN if [ -n "${ACL}" ]; then bash ./install_acl.sh; fi
 RUN rm install_acl.sh
 ENV INSTALLED_ACL ${ACL}
 
-ARG OPENBLAS
-COPY ./common/install_openblas.sh install_openblas.sh
-RUN if [ -n "${OPENBLAS}" ]; then bash ./install_openblas.sh; fi
-RUN rm install_openblas.sh
-ENV INSTALLED_OPENBLAS ${OPENBLAS}
-
 # Install ccache/sccache (do this last, so we get priority in PATH)
 ARG SKIP_SCCACHE_INSTALL
 COPY ./common/install_cache.sh install_cache.sh

.ci/manywheel/build_common.sh

@@ -97,7 +97,7 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
     exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -qUr requirements-build.txt
+retry pip install -q cmake
 python setup.py clean
 retry pip install -qr requirements.txt
 case ${DESIRED_PYTHON} in
@@ -138,11 +138,28 @@ fi
 
 echo "Calling setup.py bdist at $(date)"
 
-time CMAKE_ARGS=${CMAKE_ARGS[@]} \
-    EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 \
     BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
     USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
     python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 \
+    BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+    USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+    CMAKE_FRESH=1 python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    time CMAKE_ARGS=${CMAKE_ARGS[@]} \
+        EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+        BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+        USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+        python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+fi
 echo "Finished setup.py bdist at $(date)"
 
 # Build libtorch packages
@@ -255,6 +272,10 @@ ls /tmp/$WHEELHOUSE_DIR
 mkdir -p "/$WHEELHOUSE_DIR"
 mv /tmp/$WHEELHOUSE_DIR/torch*linux*.whl /$WHEELHOUSE_DIR/
 
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    mv /tmp/$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/ || true
+fi
+
 if [[ -n "$BUILD_PYTHONLESS" ]]; then
     mkdir -p /$LIBTORCH_HOUSE_DIR
     mv /tmp/$LIBTORCH_HOUSE_DIR/*.zip /$LIBTORCH_HOUSE_DIR
@@ -431,8 +452,16 @@ if [[ -z "$BUILD_PYTHONLESS" ]]; then
   pushd $PYTORCH_ROOT/test
 
   # Install the wheel for this Python version
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip uninstall -y "$TORCH_NO_PYTHON_PACKAGE_NAME" || true
+  fi
+
   pip uninstall -y "$TORCH_PACKAGE_NAME"
 
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip install "$TORCH_NO_PYTHON_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
+  fi
+
   pip install "$TORCH_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
 
   # Print info on the libraries installed in this wheel

.ci/manywheel/build_cuda.sh

@@ -51,23 +51,20 @@ else
 fi
 
 cuda_version_nodot=$(echo $CUDA_VERSION | tr -d '.')
-EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
 
+TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6"
 case ${CUDA_VERSION} in
-    #removing sm_50-sm_60 as these architectures are deprecated in CUDA 12.8/9 and will be removed in future releases
-    #however we would like to keep sm_70 architecture see: https://github.com/pytorch/pytorch/issues/157517
-    12.8)
-        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0"
-        ;;
-    12.9)
-        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0+PTX"
+    12.8|12.9)
+        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;12.0+PTX" #removing sm_50-sm_70 as these architectures are deprecated in CUDA 12.8/9 and will be removed in future releases
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         # WAR to resolve the ld error in libtorch build with CUDA 12.9
-        if [[ "$PACKAGE_TYPE" == "libtorch" ]]; then
+        if [[ "$DESIRED_CUDA" == "cu129" && "$PACKAGE_TYPE" == "libtorch" ]]; then
             TORCH_CUDA_ARCH_LIST="7.5;8.0;9.0;10.0;12.0+PTX"
         fi
         ;;
     12.6)
-        TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6;9.0"
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         ;;
     *)
         echo "unknown cuda version $CUDA_VERSION"
@@ -134,8 +131,6 @@ if [[ $CUDA_VERSION == 12* ]]; then
             "/usr/local/cuda/lib64/libnvrtc-builtins.so"
             "/usr/local/cuda/lib64/libcufile.so.0"
             "/usr/local/cuda/lib64/libcufile_rdma.so.1"
-            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12"
-            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
         DEPS_SONAME+=(
             "libcudnn_adv.so.9"
@@ -154,8 +149,6 @@ if [[ $CUDA_VERSION == 12* ]]; then
             "libnvrtc-builtins.so"
             "libcufile.so.0"
             "libcufile_rdma.so.1"
-            "libcupti.so.12"
-            "libnvperf_host.so"
         )
         # Add libnvToolsExt only if CUDA version is not 12.9
         if [[ $CUDA_VERSION != 12.9* ]]; then

.ci/manywheel/build_libtorch.sh

@@ -92,7 +92,7 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
     exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -qUr requirements-build.txt
+retry pip install -q cmake
 python setup.py clean
 retry pip install -qr requirements.txt
 retry pip install -q numpy==2.0.1
@@ -104,7 +104,7 @@ if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
     export ROCclr_DIR=/opt/rocm/rocclr/lib/cmake/rocclr
 fi
 
-echo "Calling 'python -m pip install .' at $(date)"
+echo "Calling setup.py install at $(date)"
 
 if [[ $LIBTORCH_VARIANT = *"static"* ]]; then
     STATIC_CMAKE_FLAG="-DTORCH_STATIC=1"
@@ -120,7 +120,7 @@ fi
         # TODO: Remove this flag once https://github.com/pytorch/pytorch/issues/55952 is closed
         CFLAGS='-Wno-deprecated-declarations' \
         BUILD_LIBTORCH_CPU_WITH_DEBUG=1 \
-        python -m pip install --no-build-isolation -v .
+        python setup.py install
 
     mkdir -p libtorch/{lib,bin,include,share}
 

.ci/manywheel/build_rocm.sh

@@ -194,7 +194,7 @@ ROCBLAS_LIB_SRC=$ROCM_HOME/lib/rocblas/library
 ROCBLAS_LIB_DST=lib/rocblas/library
 ROCBLAS_ARCH_SPECIFIC_FILES=$(ls $ROCBLAS_LIB_SRC | grep -E $ARCH)
 ROCBLAS_OTHER_FILES=$(ls $ROCBLAS_LIB_SRC | grep -v gfx)
-ROCBLAS_LIB_FILES=($ROCBLAS_ARCH_SPECIFIC_FILES $ROCBLAS_OTHER_FILES)
+ROCBLAS_LIB_FILES=($ROCBLAS_ARCH_SPECIFIC_FILES $OTHER_FILES)
 
 # hipblaslt library files
 HIPBLASLT_LIB_SRC=$ROCM_HOME/lib/hipblaslt/library

.ci/onnx/test.sh

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

.ci/pytorch/build-mobile.sh

@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+# DO NOT ADD 'set -x' not to reveal CircleCI secret context environment variables
+set -eu -o pipefail
+
+# This script uses linux host toolchain + mobile build options in order to
+# build & test mobile libtorch without having to setup Android/iOS
+# toolchain/simulator.
+
+# shellcheck source=./common.sh
+source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
+# shellcheck source=./common-build.sh
+source "$(dirname "${BASH_SOURCE[0]}")/common-build.sh"
+
+# Install torch & torchvision - used to download & trace test model.
+# Ideally we should use the libtorch built on the PR so that backward
+# incompatible changes won't break this script - but it will significantly slow
+# down mobile CI jobs.
+# Here we install nightly instead of stable so that we have an option to
+# temporarily skip mobile CI jobs on BC-breaking PRs until they are in nightly.
+retry pip install --pre torch torchvision \
+  -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html \
+  --progress-bar off
+
+# Run end-to-end process of building mobile library, linking into the predictor
+# binary, and running forward pass with a real model.
+if [[ "$BUILD_ENVIRONMENT" == *-mobile-custom-build-static* ]]; then
+  TEST_CUSTOM_BUILD_STATIC=1 test/mobile/custom_build/build.sh
+elif [[ "$BUILD_ENVIRONMENT" == *-mobile-lightweight-dispatch* ]]; then
+  test/mobile/lightweight_dispatch/build.sh
+else
+  TEST_DEFAULT_BUILD=1 test/mobile/custom_build/build.sh
+fi
+
+print_sccache_stats

.ci/pytorch/build.sh

@@ -11,6 +11,10 @@ source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
 # shellcheck source=./common-build.sh
 source "$(dirname "${BASH_SOURCE[0]}")/common-build.sh"
 
+if [[ "$BUILD_ENVIRONMENT" == *-mobile-*build* ]]; then
+  exec "$(dirname "${BASH_SOURCE[0]}")/build-mobile.sh" "$@"
+fi
+
 echo "Python version:"
 python --version
 
@@ -50,6 +54,9 @@ if [[ ${BUILD_ENVIRONMENT} == *"parallelnative"* ]]; then
   export ATEN_THREADING=NATIVE
 fi
 
+# Enable LLVM dependency for TensorExpr testing
+export USE_LLVM=/opt/llvm
+export LLVM_DIR=/opt/llvm/lib/cmake/llvm
 
 if ! which conda; then
   # In ROCm CIs, we are doing cross compilation on build machines with
@@ -117,8 +124,26 @@ if [[ "$BUILD_ENVIRONMENT" == *libtorch* ]]; then
 fi
 
 # Use special scripts for Android builds
+if [[ "${BUILD_ENVIRONMENT}" == *-android* ]]; then
+  export ANDROID_NDK=/opt/ndk
+  build_args=()
+  if [[ "${BUILD_ENVIRONMENT}" == *-arm-v7a* ]]; then
+    build_args+=("-DANDROID_ABI=armeabi-v7a")
+  elif [[ "${BUILD_ENVIRONMENT}" == *-arm-v8a* ]]; then
+    build_args+=("-DANDROID_ABI=arm64-v8a")
+  elif [[ "${BUILD_ENVIRONMENT}" == *-x86_32* ]]; then
+    build_args+=("-DANDROID_ABI=x86")
+  elif [[ "${BUILD_ENVIRONMENT}" == *-x86_64* ]]; then
+    build_args+=("-DANDROID_ABI=x86_64")
+  fi
+  if [[ "${BUILD_ENVIRONMENT}" == *vulkan* ]]; then
+    build_args+=("-DUSE_VULKAN=ON")
+  fi
+  build_args+=("-DUSE_LITE_INTERPRETER_PROFILER=OFF")
+  exec ./scripts/build_android.sh "${build_args[@]}" "$@"
+fi
 
-if [[ "$BUILD_ENVIRONMENT" == *vulkan* ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *android* && "$BUILD_ENVIRONMENT" == *vulkan* ]]; then
   export USE_VULKAN=1
   # shellcheck disable=SC1091
   source /var/lib/jenkins/vulkansdk/setup-env.sh
@@ -173,8 +198,10 @@ fi
 
 # We only build FlashAttention files for CUDA 8.0+, and they require large amounts of
 # memory to build and will OOM
-if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && echo "${TORCH_CUDA_ARCH_LIST}" | tr ' ' '\n' | sed 's/$/>= 8.0/' | bc | grep -q 1; then
-  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
+if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]] && [ -z "$MAX_JOBS_OVERRIDE" ]; then
+  echo "WARNING: FlashAttention files require large amounts of memory to build and will OOM"
+  echo "Setting MAX_JOBS=(nproc-2)/3 to reduce memory usage"
+  export MAX_JOBS="$(( $(nproc --ignore=2) / 3 ))"
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *clang* ]]; then
@@ -189,6 +216,7 @@ if [[ "$BUILD_ENVIRONMENT" == *-clang*-asan* ]]; then
   export USE_ASAN=1
   export REL_WITH_DEB_INFO=1
   export UBSAN_FLAGS="-fno-sanitize-recover=all"
+  unset USE_LLVM
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *no-ops* ]]; then
@@ -199,7 +227,7 @@ if [[ "${BUILD_ENVIRONMENT}" == *-pch* ]]; then
     export USE_PRECOMPILED_HEADERS=1
 fi
 
-if [[ "${BUILD_ENVIRONMENT}" != *cuda* ]]; then
+if [[ "${BUILD_ENVIRONMENT}" != *android* && "${BUILD_ENVIRONMENT}" != *cuda* ]]; then
   export BUILD_STATIC_RUNTIME_BENCHMARK=ON
 fi
 
@@ -229,7 +257,6 @@ if [[ "$BUILD_ENVIRONMENT" == *-bazel-* ]]; then
   set -e -o pipefail
 
   get_bazel
-  python3 tools/optional_submodules.py checkout_eigen
 
   # Leave 1 CPU free and use only up to 80% of memory to reduce the change of crashing
   # the runner
@@ -261,32 +288,25 @@ else
 
       WERROR=1 python setup.py clean
 
-      WERROR=1 python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        python3 tools/packaging/split_wheel.py bdist_wheel
+      else
+        WERROR=1 python setup.py bdist_wheel
+      fi
     else
       python setup.py clean
       if [[ "$BUILD_ENVIRONMENT" == *xla* ]]; then
         source .ci/pytorch/install_cache_xla.sh
       fi
-      python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        echo "USE_SPLIT_BUILD cannot be used with xla or rocm"
+        exit 1
+      else
+        python setup.py bdist_wheel
+      fi
     fi
     pip_install_whl "$(echo dist/*.whl)"
 
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *vision* ]]; then
-      install_torchvision
-    fi
-
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *audio* ]]; then
-      install_torchaudio
-    fi
-
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *torchrec* || "${BUILD_ADDITIONAL_PACKAGES:-}" == *fbgemm* ]]; then
-      install_torchrec_and_fbgemm
-    fi
-
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *torchao* ]]; then
-      install_torchao
-    fi
-
     if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
       echo "Checking that xpu is compiled"
       pushd dist/
@@ -374,8 +394,10 @@ else
     # This is an attempt to mitigate flaky libtorch build OOM error. By default, the build parallelization
     # is set to be the number of CPU minus 2. So, let's try a more conservative value here. A 4xlarge has
     # 16 CPUs
-    MAX_JOBS=$(nproc --ignore=4)
-    export MAX_JOBS
+    if [ -z "$MAX_JOBS_OVERRIDE" ]; then
+      MAX_JOBS=$(nproc --ignore=4)
+      export MAX_JOBS
+    fi
 
     # NB: Install outside of source directory (at the same level as the root
     # pytorch folder) so that it doesn't get cleaned away prior to docker push.

.ci/pytorch/common-build.sh

@@ -13,13 +13,6 @@ if [[ "$BUILD_ENVIRONMENT" != *win-* ]]; then
     fi
 
     if which sccache > /dev/null; then
-        # Clear SCCACHE_BUCKET and SCCACHE_REGION if they are empty, otherwise
-        # sccache will complain about invalid bucket configuration
-        if [[ -z "${SCCACHE_BUCKET:-}" ]]; then
-          unset SCCACHE_BUCKET
-          unset SCCACHE_REGION
-        fi
-
         # Save sccache logs to file
         sccache --stop-server > /dev/null  2>&1 || true
         rm -f ~/sccache_error.log || true

.ci/pytorch/common_utils.sh

@@ -78,34 +78,6 @@ function pip_install_whl() {
   fi
 }
 
-function pip_build_and_install() {
-  local build_target=$1
-  local wheel_dir=$2
-
-  local found_whl=0
-  for file in "${wheel_dir}"/*.whl
-  do
-    if [[ -f "${file}" ]]; then
-      found_whl=1
-      break
-    fi
-  done
-
-  # Build the wheel if it doesn't exist
-  if [ "${found_whl}" == "0" ]; then
-    python3 -m pip wheel \
-      --no-build-isolation \
-      --no-deps \
-      --no-use-pep517 \
-      -w "${wheel_dir}" \
-      "${build_target}"
-  fi
-
-  for file in "${wheel_dir}"/*.whl
-  do
-    pip_install_whl "${file}"
-  done
-}
 
 function pip_install() {
   # retry 3 times
@@ -152,7 +124,14 @@ function get_pinned_commit() {
 function install_torchaudio() {
   local commit
   commit=$(get_pinned_commit audio)
-  pip_build_and_install "git+https://github.com/pytorch/audio.git@${commit}" dist/audio
+  if [[ "$1" == "cuda" ]]; then
+    # TODO: This is better to be passed as a parameter from _linux-test workflow
+    # so that it can be consistent with what is set in build
+    TORCH_CUDA_ARCH_LIST="8.0;8.6" pip_install --no-use-pep517 --user "git+https://github.com/pytorch/audio.git@${commit}"
+  else
+    pip_install --no-use-pep517 --user "git+https://github.com/pytorch/audio.git@${commit}"
+  fi
+
 }
 
 function install_torchtext() {
@@ -160,8 +139,8 @@ function install_torchtext() {
   local text_commit
   data_commit=$(get_pinned_commit data)
   text_commit=$(get_pinned_commit text)
-  pip_build_and_install "git+https://github.com/pytorch/data.git@${data_commit}" dist/data
-  pip_build_and_install "git+https://github.com/pytorch/text.git@${text_commit}" dist/text
+  pip_install --no-use-pep517 --user "git+https://github.com/pytorch/data.git@${data_commit}"
+  pip_install --no-use-pep517 --user "git+https://github.com/pytorch/text.git@${text_commit}"
 }
 
 function install_torchvision() {
@@ -174,14 +153,7 @@ function install_torchvision() {
     echo 'char* dlerror(void) { return "";}'|gcc -fpic -shared -o "${HOME}/dlerror.so" -x c -
     LD_PRELOAD=${orig_preload}:${HOME}/dlerror.so
   fi
-
-  if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]]; then
-    # Not sure if both are needed, but why not
-    export FORCE_CUDA=1
-    export WITH_CUDA=1
-  fi
-  pip_build_and_install "git+https://github.com/pytorch/vision.git@${commit}" dist/vision
-
+  pip_install --no-use-pep517 --user "git+https://github.com/pytorch/vision.git@${commit}"
   if [ -n "${LD_PRELOAD}" ]; then
     LD_PRELOAD=${orig_preload}
   fi
@@ -201,71 +173,25 @@ function install_torchrec_and_fbgemm() {
 
   if [[ "$BUILD_ENVIRONMENT" == *rocm* ]] ; then
     # install torchrec first because it installs fbgemm nightly on top of rocm fbgemm
-    pip_build_and_install "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}" dist/torchrec
+    pip_install --no-use-pep517 --user "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}"
     pip_uninstall fbgemm-gpu-nightly
 
-    # Set ROCM_HOME isn't available, use ROCM_PATH if set or /opt/rocm
-    ROCM_HOME="${ROCM_HOME:-${ROCM_PATH:-/opt/rocm}}"
-
-    # Find rocm_version.h header file for ROCm version extract
-    rocm_version_h="${ROCM_HOME}/include/rocm-core/rocm_version.h"
-    if [ ! -f "$rocm_version_h" ]; then
-        rocm_version_h="${ROCM_HOME}/include/rocm_version.h"
-    fi
-
-    # Error out if rocm_version.h not found
-    if [ ! -f "$rocm_version_h" ]; then
-        echo "Error: rocm_version.h not found in expected locations." >&2
-        exit 1
-    fi
-
-    # Extract major, minor and patch ROCm version numbers
-    MAJOR_VERSION=$(grep 'ROCM_VERSION_MAJOR' "$rocm_version_h" | awk '{print $3}')
-    MINOR_VERSION=$(grep 'ROCM_VERSION_MINOR' "$rocm_version_h" | awk '{print $3}')
-    PATCH_VERSION=$(grep 'ROCM_VERSION_PATCH' "$rocm_version_h" | awk '{print $3}')
-    ROCM_INT=$((MAJOR_VERSION * 10000 + MINOR_VERSION * 100 + PATCH_VERSION))
-    echo "ROCm version: $ROCM_INT"
-    export BUILD_ROCM_VERSION="$MAJOR_VERSION.$MINOR_VERSION"
-
     pip_install tabulate  # needed for newer fbgemm
     pip_install patchelf  # needed for rocm fbgemm
-
-    local wheel_dir=dist/fbgemm_gpu
-    local found_whl=0
-    for file in "${wheel_dir}"/*.whl
-    do
-      if [[ -f "${file}" ]]; then
-        found_whl=1
-        break
-      fi
-    done
-
-    # Build the wheel if it doesn't exist
-    if [ "${found_whl}" == "0" ]; then
-      git clone --recursive https://github.com/pytorch/fbgemm
-      pushd fbgemm/fbgemm_gpu
-      git checkout "${fbgemm_commit}" --recurse-submodules
-      python setup.py bdist_wheel \
-        --build-variant=rocm \
-        -DHIP_ROOT_DIR="${ROCM_PATH}" \
-        -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
-        -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
-      popd
-
-      # Save the wheel before cleaning up
-      mkdir -p dist/fbgemm_gpu
-      cp fbgemm/fbgemm_gpu/dist/*.whl dist/fbgemm_gpu
-    fi
-
-    for file in "${wheel_dir}"/*.whl
-    do
-      pip_install_whl "${file}"
-    done
-
+    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
-    pip_build_and_install "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}" dist/torchrec
-    pip_build_and_install "git+https://github.com/pytorch/FBGEMM.git@${fbgemm_commit}#subdirectory=fbgemm_gpu" dist/fbgemm_gpu
+    # 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}"
   fi
 }
 
@@ -281,10 +207,34 @@ function clone_pytorch_xla() {
   fi
 }
 
+function checkout_install_torchbench() {
+  local commit
+  commit=$(get_pinned_commit torchbench)
+  git clone https://github.com/pytorch/benchmark torchbench
+  pushd torchbench
+  git checkout "$commit"
+
+  if [ "$1" ]; then
+    python install.py --continue_on_fail models "$@"
+  else
+    # Occasionally the installation may fail on one model but it is ok to continue
+    # 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
+}
+
 function install_torchao() {
   local commit
   commit=$(get_pinned_commit torchao)
-  pip_build_and_install "git+https://github.com/pytorch/ao.git@${commit}" dist/ao
+  pip_install --no-use-pep517 --user "git+https://github.com/pytorch/ao.git@${commit}"
 }
 
 function print_sccache_stats() {

.ci/pytorch/create_test_cert.py

@@ -0,0 +1,123 @@
+from datetime import datetime, timedelta, timezone
+from tempfile import mkdtemp
+
+from cryptography import x509
+from cryptography.hazmat.primitives import hashes, serialization
+from cryptography.hazmat.primitives.asymmetric import rsa
+from cryptography.x509.oid import NameOID
+
+
+temp_dir = mkdtemp()
+print(temp_dir)
+
+
+def genrsa(path):
+    key = rsa.generate_private_key(
+        public_exponent=65537,
+        key_size=2048,
+    )
+    with open(path, "wb") as f:
+        f.write(
+            key.private_bytes(
+                encoding=serialization.Encoding.PEM,
+                format=serialization.PrivateFormat.TraditionalOpenSSL,
+                encryption_algorithm=serialization.NoEncryption(),
+            )
+        )
+    return key
+
+
+def create_cert(path, C, ST, L, O, key):
+    subject = issuer = x509.Name(
+        [
+            x509.NameAttribute(NameOID.COUNTRY_NAME, C),
+            x509.NameAttribute(NameOID.STATE_OR_PROVINCE_NAME, ST),
+            x509.NameAttribute(NameOID.LOCALITY_NAME, L),
+            x509.NameAttribute(NameOID.ORGANIZATION_NAME, O),
+        ]
+    )
+    cert = (
+        x509.CertificateBuilder()
+        .subject_name(subject)
+        .issuer_name(issuer)
+        .public_key(key.public_key())
+        .serial_number(x509.random_serial_number())
+        .not_valid_before(datetime.now(timezone.utc))
+        .not_valid_after(
+            # Our certificate will be valid for 10 days
+            datetime.now(timezone.utc) + timedelta(days=10)
+        )
+        .add_extension(
+            x509.BasicConstraints(ca=True, path_length=None),
+            critical=True,
+        )
+        .sign(key, hashes.SHA256())
+    )
+    # Write our certificate out to disk.
+    with open(path, "wb") as f:
+        f.write(cert.public_bytes(serialization.Encoding.PEM))
+    return cert
+
+
+def create_req(path, C, ST, L, O, key):
+    csr = (
+        x509.CertificateSigningRequestBuilder()
+        .subject_name(
+            x509.Name(
+                [
+                    # Provide various details about who we are.
+                    x509.NameAttribute(NameOID.COUNTRY_NAME, C),
+                    x509.NameAttribute(NameOID.STATE_OR_PROVINCE_NAME, ST),
+                    x509.NameAttribute(NameOID.LOCALITY_NAME, L),
+                    x509.NameAttribute(NameOID.ORGANIZATION_NAME, O),
+                ]
+            )
+        )
+        .sign(key, hashes.SHA256())
+    )
+    with open(path, "wb") as f:
+        f.write(csr.public_bytes(serialization.Encoding.PEM))
+    return csr
+
+
+def sign_certificate_request(path, csr_cert, ca_cert, private_ca_key):
+    cert = (
+        x509.CertificateBuilder()
+        .subject_name(csr_cert.subject)
+        .issuer_name(ca_cert.subject)
+        .public_key(csr_cert.public_key())
+        .serial_number(x509.random_serial_number())
+        .not_valid_before(datetime.now(timezone.utc))
+        .not_valid_after(
+            # Our certificate will be valid for 10 days
+            datetime.now(timezone.utc) + timedelta(days=10)
+            # Sign our certificate with our private key
+        )
+        .sign(private_ca_key, hashes.SHA256())
+    )
+    with open(path, "wb") as f:
+        f.write(cert.public_bytes(serialization.Encoding.PEM))
+    return cert
+
+
+ca_key = genrsa(temp_dir + "/ca.key")
+ca_cert = create_cert(
+    temp_dir + "/ca.pem",
+    "US",
+    "New York",
+    "New York",
+    "Gloo Certificate Authority",
+    ca_key,
+)
+
+pkey = genrsa(temp_dir + "/pkey.key")
+csr = create_req(
+    temp_dir + "/csr.csr",
+    "US",
+    "California",
+    "San Francisco",
+    "Gloo Testing Company",
+    pkey,
+)
+
+cert = sign_certificate_request(temp_dir + "/cert.pem", csr, ca_cert, ca_key)

.ci/pytorch/macos-test.sh

@@ -5,6 +5,11 @@ set -x
 # shellcheck source=./macos-common.sh
 source "$(dirname "${BASH_SOURCE[0]}")/macos-common.sh"
 
+if [[ -n "$CONDA_ENV" ]]; then
+  # Use binaries under conda environment
+  export PATH="$CONDA_ENV/bin":$PATH
+fi
+
 # Test that OpenMP is enabled
 pushd test
 if [[ ! $(python -c "import torch; print(int(torch.backends.openmp.is_available()))") == "1" ]]; then
@@ -157,29 +162,6 @@ test_jit_hooks() {
   assert_git_not_dirty
 }
 
-# Shellcheck doesn't like it when you pass no arguments to a function
-# that can take args. See https://www.shellcheck.net/wiki/SC2120
-# shellcheck disable=SC2120
-checkout_install_torchbench() {
-  local commit
-  commit=$(cat .ci/docker/ci_commit_pins/torchbench.txt)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  if [ "$1" ]; then
-    python install.py --continue_on_fail models "$@"
-  else
-    # Occasionally the installation may fail on one model but it is ok to continue
-    # to install and test other models
-    python install.py --continue_on_fail
-  fi
-
-  echo "Print all dependencies after TorchBench is installed"
-  python -mpip freeze
-  popd
-}
-
 torchbench_setup_macos() {
   git clone --recursive https://github.com/pytorch/vision torchvision
   git clone --recursive https://github.com/pytorch/audio torchaudio
@@ -202,11 +184,13 @@ torchbench_setup_macos() {
   USE_OPENMP=0 python setup.py develop
   popd
 
+  # Shellcheck doesn't like it when you pass no arguments to a function that can take args. See https://www.shellcheck.net/wiki/SC2120
+  # shellcheck disable=SC2119,SC2120
   checkout_install_torchbench
 }
 
 pip_benchmark_deps() {
-  python -mpip install --no-input requests cython scikit-learn six
+  python -mpip install --no-input astunparse requests cython scikit-learn
 }
 
 

.ci/pytorch/run_glootls_test.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+CREATE_TEST_CERT="$(dirname "${BASH_SOURCE[0]}")/create_test_cert.py"
+TMP_CERT_DIR=$(python "$CREATE_TEST_CERT")
+
+openssl verify -CAfile "${TMP_CERT_DIR}/ca.pem" "${TMP_CERT_DIR}/cert.pem"
+
+export GLOO_DEVICE_TRANSPORT=TCP_TLS
+export GLOO_DEVICE_TRANSPORT_TCP_TLS_PKEY=${TMP_CERT_DIR}/pkey.key
+export GLOO_DEVICE_TRANSPORT_TCP_TLS_CERT=${TMP_CERT_DIR}/cert.pem
+export GLOO_DEVICE_TRANSPORT_TCP_TLS_CA_FILE=${TMP_CERT_DIR}/ca.pem
+
+time python test/run_test.py --include distributed/test_c10d_gloo --verbose -- ProcessGroupGlooTest
+
+unset GLOO_DEVICE_TRANSPORT
+unset GLOO_DEVICE_TRANSPORT_TCP_TLS_PKEY
+unset GLOO_DEVICE_TRANSPORT_TCP_TLS_CERT
+unset GLOO_DEVICE_TRANSPORT_TCP_TLS_CA_FILE

.ci/pytorch/run_tests.sh

@@ -74,13 +74,12 @@ else
 fi
 
 # Environment initialization
-retry pip install -qUr requirements-build.txt
 if [[ "$(uname)" == Darwin ]]; then
     # Install the testing dependencies
-    retry pip install -q future hypothesis ${NUMPY_PACKAGE} ${PROTOBUF_PACKAGE} pytest
+    retry pip install -q future hypothesis ${NUMPY_PACKAGE} ${PROTOBUF_PACKAGE} pytest setuptools six typing_extensions pyyaml
 else
     retry pip install -qr requirements.txt || true
-    retry pip install -q hypothesis protobuf pytest || true
+    retry pip install -q hypothesis protobuf pytest setuptools || true
     numpy_ver=1.15
     case "$(python --version 2>&1)" in
       *2* | *3.5* | *3.6*)

.ci/pytorch/smoke_test/smoke_test.py

@@ -385,29 +385,6 @@ def smoke_test_compile(device: str = "cpu") -> None:
     x_pt2 = torch.compile(model, mode="max-autotune")(x)
 
 
-def smoke_test_nvshmem() -> None:
-    if not torch.cuda.is_available():
-        print("CUDA is not available, skipping NVSHMEM test")
-        return
-
-    # Check if NVSHMEM is compiled in current build
-    try:
-        from torch._C._distributed_c10d import _is_nvshmem_available
-    except ImportError:
-        # Not built with NVSHMEM support.
-        # torch is not compiled with NVSHMEM prior to 2.9
-        if torch.__version__ < "2.9":
-            return
-        else:
-            # After 2.9: NVSHMEM is expected to be compiled in current build
-            raise RuntimeError("torch not compiled with NVSHMEM") from None
-
-    print("torch compiled with NVSHMEM")
-
-    # Check if NVSHMEM is available on current system.
-    print(f"NVSHMEM available at run time: {_is_nvshmem_available()}")
-
-
 def smoke_test_modules():
     cwd = os.getcwd()
     for module in MODULES:
@@ -502,8 +479,6 @@ def main() -> None:
         options.pypi_pkg_check,
     )
 
-    smoke_test_nvshmem()
-
 
 if __name__ == "__main__":
     main()

.ci/pytorch/test.sh

@@ -11,8 +11,6 @@ export TERM=vt100
 
 # shellcheck source=./common.sh
 source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
-# shellcheck source=./common-build.sh
-source "$(dirname "${BASH_SOURCE[0]}")/common-build.sh"
 
 # Do not change workspace permissions for ROCm and s390x CI jobs
 # as it can leave workspace with bad permissions for cancelled jobs
@@ -165,6 +163,8 @@ elif [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   export PYTORCH_TESTING_DEVICE_ONLY_FOR="xpu"
   # setting PYTHON_TEST_EXTRA_OPTION
   export PYTHON_TEST_EXTRA_OPTION="--xpu"
+  # Disable sccache for xpu test due to flaky issue https://github.com/pytorch/pytorch/issues/143585
+  sudo rm -rf /opt/cache
 fi
 
 if [[ "$TEST_CONFIG" == *crossref* ]]; then
@@ -201,7 +201,7 @@ fi
 
 if [[ "$BUILD_ENVIRONMENT" != *-bazel-* ]] ; then
   # JIT C++ extensions require ninja.
-  pip_install "ninja==1.10.2"
+  pip_install --user "ninja==1.10.2"
   # ninja is installed in $HOME/.local/bin, e.g., /var/lib/jenkins/.local/bin for CI user jenkins
   # but this script should be runnable by any user, including root
   export PATH="$HOME/.local/bin:$PATH"
@@ -289,12 +289,6 @@ elif [[ $TEST_CONFIG == 'nogpu_AVX512' ]]; then
   export ATEN_CPU_CAPABILITY=avx2
 fi
 
-if [[ "${TEST_CONFIG}" == "legacy_nvidia_driver" ]]; then
-  # Make sure that CUDA can be initialized
-  (cd test && python -c "import torch; torch.rand(2, 2, device='cuda')")
-  export USE_LEGACY_DRIVER=1
-fi
-
 test_python_legacy_jit() {
   time python test/run_test.py --include test_jit_legacy test_jit_fuser_legacy --verbose
   assert_git_not_dirty
@@ -336,21 +330,6 @@ test_h100_distributed() {
   assert_git_not_dirty
 }
 
-test_h100_symm_mem() {
-  # symmetric memory test
-  time python test/run_test.py --include distributed/test_symmetric_memory.py  $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  time python test/run_test.py --include distributed/test_nvshmem.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  time python test/run_test.py --include distributed/test_nvshmem_triton.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  time python test/run_test.py --include distributed/test_nccl.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
-test_h100_cutlass_backend() {
-  # cutlass backend tests for H100
-  TORCHINDUCTOR_CUTLASS_DIR=$(realpath "./third_party/cutlass") python test/run_test.py --include inductor/test_cutlass_backend -k "not addmm" $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  TORCHINDUCTOR_CUTLASS_DIR=$(realpath "./third_party/cutlass") python test/run_test.py --include inductor/test_cutlass_evt $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-}
-
 test_lazy_tensor_meta_reference_disabled() {
   export TORCH_DISABLE_FUNCTIONALIZATION_META_REFERENCE=1
   echo "Testing lazy tensor operations without meta reference"
@@ -379,24 +358,12 @@ test_dynamo_wrapped_shard() {
   assert_git_not_dirty
 }
 
-test_einops() {
-  pip install einops==0.6.1
-  time python test/run_test.py --einops --verbose --upload-artifacts-while-running
-  pip install einops==0.7.0
-  time python test/run_test.py --einops --verbose --upload-artifacts-while-running
-  pip install einops==0.8.1
-  time python test/run_test.py --einops --verbose --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
-
 test_inductor_distributed() {
   # Smuggle a few multi-gpu tests here so that we don't have to request another large node
   echo "Testing multi_gpu tests in test_torchinductor"
+  python test/run_test.py -i inductor/test_torchinductor.py -k test_multi_gpu --verbose
+  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 inductor/test_aot_inductor.py -k test_on_gpu_device1 --verbose
-  python test/run_test.py -i inductor/test_aot_inductor.py -k test_non_default_gpu_device --verbose
-  python test/run_test.py -i inductor/test_aot_inductor.py -k test_load_package_multiple_gpus --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/parallel/test_micro_pipeline_tp.py --verbose
@@ -448,21 +415,14 @@ test_inductor_aoti() {
     python3 tools/amd_build/build_amd.py
   fi
   if [[ "$BUILD_ENVIRONMENT" == *sm86* ]]; then
-    BUILD_COMMAND=(TORCH_CUDA_ARCH_LIST=8.6 USE_FLASH_ATTENTION=OFF python -m pip install --no-build-isolation -v -e .)
+    BUILD_AOT_INDUCTOR_TEST=1 TORCH_CUDA_ARCH_LIST=8.6 USE_FLASH_ATTENTION=OFF python setup.py develop
     # TODO: Replace me completely, as one should not use conda libstdc++, nor need special path to TORCH_LIB
-    TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="/opt/conda/envs/py_3.10/lib:${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}")
+    LD_LIBRARY_PATH=/opt/conda/envs/py_3.10/lib/:${TORCH_LIB_DIR}:$LD_LIBRARY_PATH
+    CPP_TESTS_DIR="${BUILD_BIN_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference -dist=loadfile
   else
-    BUILD_COMMAND=(python -m pip install --no-build-isolation -v -e .)
-    TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}")
+    BUILD_AOT_INDUCTOR_TEST=1 python setup.py develop
+    CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference -dist=loadfile
   fi
-
-  # aoti cmake custom command requires `torch` to be installed
-  # initialize the cmake build cache and install torch
-  /usr/bin/env "${BUILD_COMMAND[@]}"
-  # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
-  /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${BUILD_COMMAND[@]}"
-
-  /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
 test_inductor_cpp_wrapper_shard() {
@@ -475,26 +435,47 @@ test_inductor_cpp_wrapper_shard() {
   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.  Run all CPU tests, as well as specific GPU tests triggering past
+    # bugs, for now.
+    python test/run_test.py \
+      --include inductor/test_torchinductor_opinfo \
+      -k 'linalg or to_sparse or TestInductorOpInfoCPU' \
+      --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.
-  #
-  # TODO: I'm pretty sure that "TestInductorOpInfoCPU" is not a valid filter,
-  # but change that in another PR to more accurately monitor the increased CI
-  # usage.
-  python test/run_test.py \
-    --include inductor/test_torchinductor_opinfo \
-    -k 'linalg or to_sparse or TestInductorOpInfoCPU' \
-    --shard "$1" "$NUM_TEST_SHARDS" \
-    --verbose
   python test/run_test.py \
     --include inductor/test_torchinductor inductor/test_max_autotune inductor/test_cpu_repro \
-    --shard "$1" "$NUM_TEST_SHARDS" \
-    --verbose
-  python test/run_test.py --inductor \
-    --include test_torch \
-    -k 'take' \
-    --shard "$1" "$NUM_TEST_SHARDS" \
     --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.
+  if [[ "${PYTORCH_TEST_RERUN_DISABLED_TESTS}" == "1" ]]; then
+    echo "skip dynamo benchmark tests for rerun-disabled-test"
+  else
+    echo "run dynamo benchmark tests with cpp wrapper"
+    python benchmarks/dynamo/timm_models.py --device cuda --accuracy --amp \
+    --training --inductor --disable-cudagraphs --only vit_base_patch16_224 \
+    --output "$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"
+
+    python benchmarks/dynamo/torchbench.py --device cuda --accuracy \
+      --bfloat16 --inference --inductor --only hf_T5 --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv"
+    python benchmarks/dynamo/torchbench.py --device cuda --accuracy \
+      --bfloat16 --inference --inductor --only llama --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv"
+    python benchmarks/dynamo/torchbench.py --device cuda --accuracy \
+      --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"
+  fi
 }
 
 # "Global" flags for inductor benchmarking controlled by TEST_CONFIG
@@ -507,7 +488,7 @@ DYNAMO_BENCHMARK_FLAGS=()
 
 pr_time_benchmarks() {
 
-  pip_install "fbscribelogger"
+  pip_install --user "fbscribelogger"
 
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
@@ -615,8 +596,8 @@ test_perf_for_dashboard() {
 
   local device=cuda
   if [[ "${TEST_CONFIG}" == *cpu* ]]; then
-    if [[ "${TEST_CONFIG}" == *cpu_x86_zen* ]]; then
-      device=cpu_x86_zen
+    if [[ "${TEST_CONFIG}" == *zen_cpu_x86* ]]; then
+      device=zen_cpu_x86
     elif [[ "${TEST_CONFIG}" == *cpu_x86* ]]; then
       device=cpu_x86
     elif [[ "${TEST_CONFIG}" == *cpu_aarch64* ]]; then
@@ -627,19 +608,13 @@ test_perf_for_dashboard() {
     device=cuda_a10g
   elif [[ "${TEST_CONFIG}" == *h100* ]]; then
     device=cuda_h100
-  elif [[ "${TEST_CONFIG}" == *b200* ]]; then
-    device=cuda_b200
   elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
     device=rocm
   fi
 
   for mode in "${modes[@]}"; do
     if [[ "$mode" == "inference" ]]; then
-      if [[ "$device" == "cpu_x86" ]]; then
-        dtype=amp
-      else
-        dtype=bfloat16
-      fi
+      dtype=bfloat16
     elif [[ "$mode" == "training" ]]; then
       dtype=amp
     fi
@@ -651,10 +626,6 @@ test_perf_for_dashboard() {
         target_flag+=( --no-translation-validation)
       fi
 
-      if [[ "$DASHBOARD_TAG" == *freezing-true* ]]; then
-        target_flag+=( --freezing)
-      fi
-
       if [[ "$DASHBOARD_TAG" == *default-true* ]]; then
         $TASKSET python "benchmarks/dynamo/$suite.py" \
             "${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" --disable-cudagraphs "$@" \
@@ -803,16 +774,6 @@ test_dynamo_benchmark() {
   if [[ "${TEST_CONFIG}" == *perf_compare* ]]; then
     test_single_dynamo_benchmark "training" "$suite" "$shard_id" --training --amp "$@"
   elif [[ "${TEST_CONFIG}" == *perf* ]]; then
-    # TODO (huydhn): Just smoke test some sample models
-    if [[ "${TEST_CONFIG}" == *b200* ]]; then
-      if [[ "${suite}" == "huggingface" ]]; then
-        export TORCHBENCH_ONLY_MODELS="DistillGPT2"
-      elif [[ "${suite}" == "timm_models" ]]; then
-        export TORCHBENCH_ONLY_MODELS="inception_v3"
-      elif [[ "${suite}" == "torchbench" ]]; then
-        export TORCHBENCH_ONLY_MODELS="hf_Bert"
-      fi
-    fi
     test_single_dynamo_benchmark "dashboard" "$suite" "$shard_id" "$@"
   else
     if [[ "${TEST_CONFIG}" == *cpu* ]]; then
@@ -940,6 +901,12 @@ test_torchbench_gcp_smoketest(){
   popd
 }
 
+test_python_gloo_with_tls() {
+  source "$(dirname "${BASH_SOURCE[0]}")/run_glootls_test.sh"
+  assert_git_not_dirty
+}
+
+
 test_aten() {
   # Test ATen
   # The following test(s) of ATen have already been skipped by caffe2 in rocm environment:
@@ -986,8 +953,6 @@ test_without_numpy() {
   if [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then
     python -c "import sys;sys.path.insert(0, 'fake_numpy');import torch;torch.compile(lambda x:print(x))('Hello World')"
   fi
-  # Regression test for https://github.com/pytorch/pytorch/pull/157734 (torch.onnx should be importable without numpy)
-  python -c "import sys;sys.path.insert(0, 'fake_numpy');import torch; import torch.onnx"
   popd
 }
 
@@ -1051,10 +1016,20 @@ test_libtorch_api() {
     mkdir -p $TEST_REPORTS_DIR
 
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" "$TORCH_BIN_DIR"/test_api --gtest_filter='-IMethodTest.*' --gtest_output=xml:$TEST_REPORTS_DIR/test_api.xml
+    "$TORCH_BIN_DIR"/test_tensorexpr --gtest_output=xml:$TEST_REPORTS_DIR/test_tensorexpr.xml
   else
     # Exclude IMethodTest that relies on torch::deploy, which will instead be ran in test_deploy
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${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
   fi
 
   # quantization is not fully supported on s390x yet
@@ -1322,13 +1297,10 @@ EOF
 
   # Step 2. Make sure that the public API test "test_correct_module_names" fails when an existing
   # file is modified to introduce an invalid public API function.
-  # The filepath here must not have __all__ defined in it, otherwise the test will pass.
-  # If your PR introduces __all__ to torch/cuda/streams.py please point this to another file
-  # that does not have __all__ defined.
-  EXISTING_FILEPATH="${TORCH_INSTALL_DIR}/cuda/streams.py"
+  EXISTING_FILEPATH="${TORCH_INSTALL_DIR}/nn/parameter.py"
   cp -v "${EXISTING_FILEPATH}" "${EXISTING_FILEPATH}.orig"
   echo "${BAD_PUBLIC_FUNC}" >> "${EXISTING_FILEPATH}"
-  invalid_api="torch.cuda.streams.new_public_func"
+  invalid_api="torch.nn.parameter.new_public_func"
   echo "Appended an invalid public API function to existing file ${EXISTING_FILEPATH}..."
 
   check_public_api_test_fails \
@@ -1483,8 +1455,8 @@ test_bazel() {
 
 test_benchmarks() {
   if [[ "$BUILD_ENVIRONMENT" == *cuda* && $TEST_CONFIG != *nogpu* ]]; then
-    pip_install "pytest-benchmark==3.2.3"
-    pip_install "requests"
+    pip_install --user "pytest-benchmark==3.2.3"
+    pip_install --user "requests"
     BENCHMARK_DATA="benchmarks/.data"
     mkdir -p ${BENCHMARK_DATA}
     pytest benchmarks/fastrnns/test_bench.py --benchmark-sort=Name --benchmark-json=${BENCHMARK_DATA}/fastrnns_default.json --fuser=default --executor=default
@@ -1562,7 +1534,7 @@ 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_ops \
+        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops test_cpp_extensions_open_device_registration \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1592,7 +1564,7 @@ test_operator_benchmark() {
   test_inductor_set_cpu_affinity
 
   cd benchmarks/operator_benchmark/pt_extension
-  python -m pip install .
+  python setup.py install
 
   cd "${TEST_DIR}"/benchmarks/operator_benchmark
   $TASKSET python -m benchmark_all_test --device "$1" --tag-filter "$2" \
@@ -1612,13 +1584,7 @@ if ! [[ "${BUILD_ENVIRONMENT}" == *libtorch* || "${BUILD_ENVIRONMENT}" == *-baze
 fi
 if [[ "${TEST_CONFIG}" == *numpy_2* ]]; then
   # Install numpy-2.0.2 and compatible scipy & numba versions
-  # Force re-install of pandas to avoid error where pandas checks numpy version from initial install and fails upon import
-  TMP_PANDAS_VERSION=$(python -c "import pandas; print(pandas.__version__)" 2>/dev/null)
-  if [ -n "$TMP_PANDAS_VERSION" ]; then
-    python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0 pandas=="$TMP_PANDAS_VERSION" --force-reinstall
-  else
-    python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
-  fi
+  python -mpip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
   python test/run_test.py --include dynamo/test_functions.py dynamo/test_unspec.py test_binary_ufuncs.py test_fake_tensor.py test_linalg.py test_numpy_interop.py test_tensor_creation_ops.py test_torch.py torch_np/test_basic.py
 elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
   test_linux_aarch64
@@ -1672,40 +1638,52 @@ elif [[ "${TEST_CONFIG}" == *timm* ]]; then
   id=$((SHARD_NUMBER-1))
   test_dynamo_benchmark timm_models "$id"
 elif [[ "${TEST_CONFIG}" == cachebench ]]; then
-  install_torchaudio
+  install_torchaudio cuda
   install_torchvision
-  PYTHONPATH=/torchbench test_cachebench
+  checkout_install_torchbench nanogpt BERT_pytorch resnet50 hf_T5 llama moco
+  PYTHONPATH=$(pwd)/torchbench test_cachebench
 elif [[ "${TEST_CONFIG}" == verify_cachebench ]]; then
-  install_torchaudio
+  install_torchaudio cpu
   install_torchvision
-  PYTHONPATH=/torchbench test_verify_cachebench
+  checkout_install_torchbench nanogpt
+  PYTHONPATH=$(pwd)/torchbench test_verify_cachebench
 elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
-  install_torchaudio
+  if [[ "${TEST_CONFIG}" == *cpu* ]]; then
+    install_torchaudio cpu
+  else
+    install_torchaudio cuda
+  fi
   install_torchvision
-  install_torchao
+  TORCH_CUDA_ARCH_LIST="8.0;8.6" install_torchao
   id=$((SHARD_NUMBER-1))
   # https://github.com/opencv/opencv-python/issues/885
   pip_install opencv-python==4.8.0.74
   if [[ "${TEST_CONFIG}" == *inductor_torchbench_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_smoketest_perf
+    checkout_install_torchbench hf_Bert hf_Albert timm_vision_transformer
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *inductor_torchbench_cpu_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_cpu_smoketest_perf
+    checkout_install_torchbench timm_vision_transformer phlippe_densenet basic_gnn_edgecnn \
+      llama_v2_7b_16h resnet50 timm_efficientnet mobilenet_v3_large timm_resnest \
+      functorch_maml_omniglot yolov3 mobilenet_v2 resnext50_32x4d densenet121 mnasnet1_0
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_cpu_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *torchbench_gcp_smoketest* ]]; then
-    TORCHBENCHPATH=/torchbench test_torchbench_gcp_smoketest
+    checkout_install_torchbench
+    TORCHBENCHPATH=$(pwd)/torchbench test_torchbench_gcp_smoketest
   else
+    checkout_install_torchbench
     # Do this after checkout_install_torchbench to ensure we clobber any
     # nightlies that torchbench may pull in
     if [[ "${TEST_CONFIG}" != *cpu* ]]; then
       install_torchrec_and_fbgemm
     fi
-    PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
+    PYTHONPATH=$(pwd)/torchbench test_dynamo_benchmark torchbench "$id"
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
+  install_torchaudio cuda
   install_torchvision
-  PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
-  if [[ "$SHARD_NUMBER" -eq "1" ]]; then
-    test_inductor_aoti
-  fi
+  checkout_install_torchbench hf_T5 llama moco
+  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
+  test_inductor_aoti
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   test_inductor_shard "${SHARD_NUMBER}"
@@ -1714,8 +1692,6 @@ elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
       test_inductor_distributed
     fi
   fi
-elif [[ "${TEST_CONFIG}" == *einops* ]]; then
-  test_einops
 elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then
   install_torchvision
   test_dynamo_wrapped_shard "${SHARD_NUMBER}"
@@ -1765,10 +1741,6 @@ elif [[ "${TEST_CONFIG}" == smoke ]]; then
   test_python_smoke
 elif [[ "${TEST_CONFIG}" == h100_distributed ]]; then
   test_h100_distributed
-elif [[ "${TEST_CONFIG}" == "h100-symm-mem" ]]; then
-  test_h100_symm_mem
-elif [[ "${TEST_CONFIG}" == h100_cutlass_backend ]]; then
-  test_h100_cutlass_backend
 else
   install_torchvision
   install_monkeytype

.ci/pytorch/win-arm64-build.ps1

@@ -1,34 +0,0 @@
-# If you want to rebuild, run this with $env:REBUILD=1
-# If you want to build with CUDA, run this with $env:USE_CUDA=1
-# If you want to build without CUDA, run this with $env:USE_CUDA=0
-
-# Check for setup.py in the current directory
-if (-not (Test-Path "setup.py")) {
-    Write-Host "ERROR: Please run this build script from PyTorch root directory."
-    exit 1
-}
-
-# Get the script's parent directory
-$ScriptParentDir = Split-Path -Parent $MyInvocation.MyCommand.Definition
-
-# Set TMP_DIR and convert to Windows path
-$env:TMP_DIR = Join-Path (Get-Location) "build\win_tmp"
-$env:TMP_DIR_WIN = $env:TMP_DIR  # Already in Windows format, no cygpath needed
-
-# Set final package directory with default fallback
-if (-not $env:PYTORCH_FINAL_PACKAGE_DIR) {
-    $env:PYTORCH_FINAL_PACKAGE_DIR = "C:\w\build-results"
-}
-
-# Create the final package directory if it doesn't exist
-if (-not (Test-Path $env:PYTORCH_FINAL_PACKAGE_DIR)) {
-    New-Item -Path $env:PYTORCH_FINAL_PACKAGE_DIR -ItemType Directory -Force | Out-Null
-}
-
-# Set script helpers directory
-$env:SCRIPT_HELPERS_DIR = Join-Path $ScriptParentDir "win-test-helpers\arm64"
-
-# Run the main build script
-& "$env:SCRIPT_HELPERS_DIR\build_pytorch.ps1"
-
-Write-Host "BUILD PASSED"

.ci/pytorch/win-arm64-test.sh

@@ -1,24 +0,0 @@
-#!/bin/bash
-set -ex -o pipefail
-
-SCRIPT_PARENT_DIR=$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )
-# shellcheck source=./common.sh
-source "$SCRIPT_PARENT_DIR/common.sh"
-
-run_tests() {
-    echo Running smoke_test.py...
-    python ./.ci/pytorch/smoke_test/smoke_test.py --package torchonly
-
-    echo Running test_autograd.oy, test_nn.py, test_torch.py...
-    cd test
-
-    CORE_TEST_LIST=("test_autograd.py" "test_nn.py" "test_modules.py")
-
-    for t in "${CORE_TEST_LIST[@]}"; do
-        echo "Running test: $t"
-        python "$t" --verbose --save-xml --use-pytest -vvvv -rfEsxXP -p no:xdist
-    done
-}
-
-run_tests
-echo "TEST PASSED"

.ci/pytorch/win-test-helpers/arm64/build_pytorch.ps1

@@ -1,98 +0,0 @@
-# TODO: we may can use existing build_pytorch.bat for arm64
-
-if ($env:DEBUG -eq "1") {
-    $env:BUILD_TYPE = "debug"
-} else {
-    $env:BUILD_TYPE = "release"
-}
-
-# This inflates our log size slightly, but it is REALLY useful to be
-# able to see what our cl.exe commands are. (since you can actually
-# just copy-paste them into a local Windows setup to just rebuild a
-# single file.)
-# log sizes are too long, but leaving this here in case someone wants to use it locally
-# $env:CMAKE_VERBOSE_MAKEFILE = "1"
-
-$env:INSTALLER_DIR = Join-Path $env:SCRIPT_HELPERS_DIR "installation-helpers"
-
-cd ..
-
-# Environment variables
-$env:SCCACHE_IDLE_TIMEOUT = "0"
-$env:SCCACHE_IGNORE_SERVER_IO_ERROR = "1"
-$env:CMAKE_BUILD_TYPE = $env:BUILD_TYPE
-$env:CMAKE_C_COMPILER_LAUNCHER = "sccache"
-$env:CMAKE_CXX_COMPILER_LAUNCHER = "sccache"
-$env:libuv_ROOT = Join-Path $env:DEPENDENCIES_DIR "libuv\install"
-$env:MSSdk = "1"
-
-if ($env:PYTORCH_BUILD_VERSION) {
-    $env:PYTORCH_BUILD_VERSION = $env:PYTORCH_BUILD_VERSION
-    $env:PYTORCH_BUILD_NUMBER = "1"
-}
-
-$env:CMAKE_POLICY_VERSION_MINIMUM = "3.5"
-
-# Set BLAS type
-if ($env:ENABLE_APL -eq "1") {
-    $env:BLAS = "APL"
-    $env:USE_LAPACK = "1"
-} elseif ($env:ENABLE_OPENBLAS -eq "1") {
-    $env:BLAS = "OpenBLAS"
-    $env:OpenBLAS_HOME = Join-Path $env:DEPENDENCIES_DIR "OpenBLAS\install"
-}
-
-# Change to source directory
-Set-Location $env:PYTORCH_ROOT
-
-# Copy libuv.dll
-Copy-Item -Path (Join-Path $env:libuv_ROOT "lib\Release\uv.dll") -Destination "torch\lib\uv.dll" -Force
-
-# Create virtual environment
-python -m venv .venv
-.\.venv\Scripts\Activate.ps1
-where.exe python
-
-# Python install dependencies
-python -m pip install --upgrade pip
-pip install setuptools pyyaml
-pip install -r requirements.txt
-
-# Set after installing psutil
-$env:DISTUTILS_USE_SDK = "1"
-
-# Print all environment variables
-Get-ChildItem Env:
-
-# Start and inspect sccache
-sccache --start-server
-sccache --zero-stats
-sccache --show-stats
-
-# Build the wheel
-python setup.py bdist_wheel
-if ($LASTEXITCODE -ne 0) { exit 1 }
-
-# Install the wheel locally
-$whl = Get-ChildItem -Path "dist\*.whl" | Select-Object -First 1
-if ($whl) {
-    python -mpip install --no-index --no-deps $whl.FullName
-}
-
-# Copy final wheel
-robocopy "dist" "$env:PYTORCH_FINAL_PACKAGE_DIR" *.whl
-
-# Export test times
-python tools/stats/export_test_times.py
-
-# Copy additional CI files
-robocopy ".additional_ci_files" "$env:PYTORCH_FINAL_PACKAGE_DIR\.additional_ci_files" /E
-
-# Save ninja log
-Copy-Item -Path "build\.ninja_log" -Destination $env:PYTORCH_FINAL_PACKAGE_DIR -Force
-
-# Final sccache stats and stop
-sccache --show-stats
-sccache --stop-server
-
-exit 0

.ci/pytorch/win-test-helpers/build_pytorch.bat

@@ -42,7 +42,7 @@ call choco upgrade -y cmake --no-progress --installargs 'ADD_CMAKE_TO_PATH=Syste
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail
 
-call pip install mkl==2024.2.0 mkl-static==2024.2.0 mkl-include==2024.2.0
+call pip install mkl-include==2021.4.0 mkl-devel==2021.4.0
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail
 

.ci/pytorch/win-test.sh

@@ -41,7 +41,7 @@ fi
 python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
 
 # Install Z3 optional dependency for Windows builds.
-python -m pip install z3-solver==4.15.1.0
+python -m pip install z3-solver==4.12.2.0
 
 # Install tlparse for test\dynamo\test_structured_trace.py UTs.
 python -m pip install tlparse==0.3.30
@@ -52,9 +52,6 @@ python -m pip install parameterized==0.8.1
 # Install pulp for testing ilps under torch\distributed\_tools
 python -m pip install pulp==2.9.0
 
-# Install expecttest to merge https://github.com/pytorch/pytorch/pull/155308
-python -m pip install expecttest==0.3.0
-
 run_tests() {
     # Run nvidia-smi if available
     for path in '/c/Program Files/NVIDIA Corporation/NVSMI/nvidia-smi.exe' /c/Windows/System32/nvidia-smi.exe; do

.ci/pytorch/windows/cuda129.bat

@@ -29,7 +29,7 @@ IF "%NVTOOLSEXT_PATH%"=="" (
 
 IF "%CUDA_PATH_V129%"=="" (
     IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\bin\nvcc.exe" (
-        set "CUDA_PATH_V129=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9"
+        set "CUDA_PATH_V128=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9"
     ) ELSE (
         echo CUDA 12.9 not found, failing
         exit /b 1
@@ -37,10 +37,10 @@ IF "%CUDA_PATH_V129%"=="" (
 )
 
 IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=7.0;7.5;8.0;8.6;9.0;10.0;12.0
+    set TORCH_CUDA_ARCH_LIST=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_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 NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -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_V129%"

.ci/pytorch/windows/internal/copy.bat

@@ -8,7 +8,6 @@ copy "%CUDA_PATH%\bin\cusolver*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\bin\cudnn*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\bin\nvrtc*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\cupti64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\extras\CUPTI\lib64\nvperf_host*.dll*" pytorch\torch\lib
 
 copy "C:\Program Files\NVIDIA Corporation\NvToolsExt\bin\x64\nvToolsExt64_1.dll*" pytorch\torch\lib
 copy "%PYTHON_LIB_PATH%\libiomp*5md.dll" pytorch\torch\lib

.ci/pytorch/windows/internal/install_python.bat

@@ -18,5 +18,3 @@ start /wait "" python-amd64.exe /quiet InstallAllUsers=1 PrependPath=0 Include_t
 if errorlevel 1 exit /b 1
 
 set "PATH=%CD%\Python\Scripts;%CD%\Python;%PATH%"
-%PYTHON_EXEC% -m pip install --upgrade pip setuptools packaging wheel
-if errorlevel 1 exit /b 1

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

@@ -148,7 +148,14 @@ if "%NVIDIA_GPU_EXISTS%" == "0" (
     goto end
 )
 
-cl %PYTORCH_ROOT%\.ci\pytorch\test_example_code\check-torch-cuda.cpp torch_cpu.lib c10.lib torch_cuda.lib /EHsc /std:c++17 /link /INCLUDE:?warp_size@cuda@at@@YAHXZ
+set BUILD_SPLIT_CUDA=
+if exist "%install_root%\lib\torch_cuda_cu.lib" if exist "%install_root%\lib\torch_cuda_cpp.lib" set BUILD_SPLIT_CUDA=ON
+
+if "%BUILD_SPLIT_CUDA%" == "ON" (
+    cl %PYTORCH_ROOT%\.ci\pytorch\test_example_code\check-torch-cuda.cpp torch_cpu.lib c10.lib torch_cuda_cu.lib torch_cuda_cpp.lib /EHsc /std:c++17 /link /INCLUDE:?warp_size@cuda@at@@YAHXZ /INCLUDE:?_torch_cuda_cu_linker_symbol_op_cuda@native@at@@YA?AVTensor@2@AEBV32@@Z
+) else (
+    cl %PYTORCH_ROOT%\.ci\pytorch\test_example_code\check-torch-cuda.cpp torch_cpu.lib c10.lib torch_cuda.lib /EHsc /std:c++17 /link /INCLUDE:?warp_size@cuda@at@@YAHXZ
+)
 .\check-torch-cuda.exe
 if ERRORLEVEL 1 exit /b 1
 

.ci/wheel/build_wheel.sh

@@ -127,7 +127,7 @@ export INSTALL_TEST=0 # dont install test binaries into site-packages
 export MACOSX_DEPLOYMENT_TARGET=10.15
 export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
-SETUPTOOLS_PINNED_VERSION="==70.1.0"
+SETUPTOOLS_PINNED_VERSION="=46.0.0"
 PYYAML_PINNED_VERSION="=5.3"
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
@@ -135,7 +135,7 @@ RENAME_WHEEL=true
 case $desired_python in
     3.13t)
         echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=68.0.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
         NUMPY_PINNED_VERSION="=2.1.0"
         CONDA_ENV_CREATE_FLAGS="python-freethreading"
@@ -145,31 +145,31 @@ case $desired_python in
         ;;
     3.13)
         echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=68.0.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
         NUMPY_PINNED_VERSION="=2.1.0"
         ;;
     3.12)
         echo "Using 3.12 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=68.0.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.11)
         echo "Using 3.11 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=46.0.0"
         PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.10)
         echo "Using 3.10 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=46.0.0"
         PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.9)
         echo "Using 3.9 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=46.0.0"
         PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
@@ -184,14 +184,16 @@ tmp_env_name="wheel_py$python_nodot"
 conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
 source activate "$tmp_env_name"
 
-retry pip install -r "${pytorch_rootdir}/requirements-build.txt"
-pip install "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing-extensions
+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
 
 # For USE_DISTRIBUTED=1 on macOS, need libuv, which is build as part of tensorpipe submodule
 export USE_DISTRIBUTED=1
 
+if [[ -n "$CROSS_COMPILE_ARM64" ]]; then
+    export CMAKE_OSX_ARCHITECTURES=arm64
+fi
 export USE_MKLDNN=OFF
 export USE_QNNPACK=OFF
 export BUILD_TEST=OFF
@@ -199,7 +201,16 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-python setup.py bdist_wheel -d "$whl_tmp_dir"
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 CMAKE_FRESH=1 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    python setup.py bdist_wheel -d "$whl_tmp_dir"
+fi
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.circleci/scripts/binary_linux_test.sh

@@ -65,8 +65,16 @@ fi
 
 if [[ "$PACKAGE_TYPE" != libtorch ]]; then
   if [[ "\$BUILD_ENVIRONMENT" != *s390x* ]]; then
-    pip install "\$pkg" --index-url "https://download.pytorch.org/whl/\${CHANNEL}/${DESIRED_CUDA}"
-    retry pip install -q numpy protobuf typing-extensions
+    if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+      pkg_no_python="$(ls -1 /final_pkgs/torch_no_python* | sort |tail -1)"
+      pkg_torch="$(ls -1 /final_pkgs/torch-* | sort |tail -1)"
+      # todo: after folder is populated use the pypi_pkg channel instead
+      pip install "\$pkg_no_python" "\$pkg_torch" --index-url "https://download.pytorch.org/whl/\${CHANNEL}/${DESIRED_CUDA}_pypi_pkg"
+      retry pip install -q numpy protobuf typing-extensions
+    else
+      pip install "\$pkg" --index-url "https://download.pytorch.org/whl/\${CHANNEL}/${DESIRED_CUDA}"
+      retry pip install -q numpy protobuf typing-extensions
+    fi
   else
     pip install "\$pkg"
     retry pip install -q numpy protobuf typing-extensions

.circleci/scripts/binary_populate_env.sh

@@ -75,8 +75,8 @@ TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_version.txt)
 # Here PYTORCH_EXTRA_INSTALL_REQUIREMENTS is already set for the all the wheel builds hence append TRITON_CONSTRAINT
 TRITON_CONSTRAINT="platform_system == 'Linux' and platform_machine == 'x86_64'"
 
-# CUDA 12.9 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]]; then
+# CUDA 12.8 builds have triton for Linux and Linux aarch64 binaries.
+if [[ "$DESIRED_CUDA" == cu128 ]]; then
   TRITON_CONSTRAINT="platform_system == 'Linux'"
 fi
 
@@ -134,6 +134,7 @@ export DESIRED_PYTHON="${DESIRED_PYTHON:-}"
 export DESIRED_CUDA="$DESIRED_CUDA"
 export LIBTORCH_VARIANT="${LIBTORCH_VARIANT:-}"
 export BUILD_PYTHONLESS="${BUILD_PYTHONLESS:-}"
+export USE_SPLIT_BUILD="${USE_SPLIT_BUILD:-}"
 if [[ "${OSTYPE}" == "msys" ]]; then
   export LIBTORCH_CONFIG="${LIBTORCH_CONFIG:-}"
   if [[ "${LIBTORCH_CONFIG:-}" == 'debug' ]]; then

.circleci/scripts/binary_upload.sh

@@ -23,6 +23,10 @@ if [[ "${DRY_RUN}" = "disabled" ]]; then
   AWS_S3_CP="aws s3 cp"
 fi
 
+if [[ "${USE_SPLIT_BUILD:-false}" == "true" ]]; then
+  UPLOAD_SUBFOLDER="${UPLOAD_SUBFOLDER}_pypi_pkg"
+fi
+
 # this is special build with all dependencies packaged
 if [[ ${BUILD_NAME} == *-full* ]]; then
   UPLOAD_SUBFOLDER="${UPLOAD_SUBFOLDER}_full"

.clang-format

@@ -120,7 +120,6 @@ UseTab:          Never
 Language: ObjC
 ColumnLimit: 120
 AlignAfterOpenBracket: Align
-IndentWidth: 2
 ObjCBlockIndentWidth: 2
 ObjCSpaceAfterProperty: false
 ObjCSpaceBeforeProtocolList: false

.devcontainer/README.md

@@ -61,8 +61,8 @@ You are now all set to start developing with PyTorch in a DevContainer environme
 ## Step 8: Build PyTorch
 
 To build pytorch from source, simply run:
-   ```bash
-   python -m pip install --no-build-isolation -v -e .
+   ```
+   python setup.py develop
    ```
 
 The process involves compiling thousands of files, and would take a long time. Fortunately, the compiled objects can be useful for your next build. When you modify some files, you only need to compile the changed files the next time.

.editorconfig

@@ -1,36 +1,14 @@
 root = true
 
 [*]
-charset = utf-8
 end_of_line = lf
 insert_final_newline = true
 
 # Python
-[*.{py,pyi,py.in,pyi.in}]
+[*.py]
 indent_style = space
 indent_size = 4
 
-# C/C++/CUDA
-[*.{cpp,hpp,cxx,cc,c,h,cu,cuh}]
-indent_style = space
-indent_size = 2
-
-# Objective-C
-[*.{mm,m,M}]
-indent_style = space
-indent_size = 2
-
-# Clang tools
-[.clang-{format,tidy}]
-indent_style = space
-indent_size = 2
-
 # Make
 [Makefile]
 indent_style = tab
-
-# Batch file
-[*.bat]
-indent_style = space
-indent_size = 2
-end_of_line = crlf

.flake8

@@ -7,12 +7,12 @@ max-line-length = 120
 # C408 ignored because we like the dict keyword argument syntax
 # E501 is not flexible enough, we're using B950 instead
 ignore =
-    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,
     # shebang has extra meaning in fbcode lints, so I think it's not worth trying
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
-    B007,B008,B017,B019,B023,B028,B903,B904,B905,B906,B907,B908,B910
+    B007,B008,B017,B019,B023,B028,B903,B904,B905,B906,B907
     # these ignores are from flake8-comprehensions; please fix!
     C407,
     # these ignores are from flake8-logging-format; please fix!

.github/actionlint.yaml

@@ -53,12 +53,16 @@ self-hosted-runner:
     - linux.rocm.gpu.mi250
     - linux.rocm.gpu.2
     - linux.rocm.gpu.4
-    # gfx942 runners
-    - linux.rocm.gpu.gfx942.2
-    - linux.rocm.gpu.gfx942.4
+    # MI300 runners
+    - linux.rocm.gpu.mi300.2
+    - linux.rocm.gpu.mi300.4
     - rocm-docker
+    # Repo-specific Apple hosted  runners
+    - macos-m1-ultra
+    - macos-m2-14
     # Org wise AWS `mac2.metal` runners (2020 Mac mini hardware powered by Apple silicon M1 processors)
     - macos-m1-stable
+    - macos-m1-13
     - macos-m1-14
     # GitHub-hosted MacOS runners
     - macos-latest-xlarge

.github/actions/build-android/action.yml

@@ -0,0 +1,78 @@
+name: build android
+
+description: build android for a specific arch
+
+inputs:
+  arch:
+    description: arch to build
+    required: true
+  arch-for-build-env:
+    description: |
+      arch to pass to build environment.
+      This is currently different than the arch name we use elsewhere, which
+      should be fixed.
+    required: true
+  github-secret:
+    description: github token
+    required: true
+  build-environment:
+    required: true
+    description: Top-level label for what's being built/tested.
+  docker-image:
+    required: true
+    description: Name of the base docker image to build with.
+  branch:
+    required: true
+    description: What branch we are building on.
+outputs:
+  container_id:
+    description: Docker container identifier used to build the artifacts
+    value: ${{ steps.build.outputs.container_id }}
+
+runs:
+  using: composite
+  steps:
+    - name: Build-${{ inputs.arch }}
+      id: build
+      shell: bash
+      env:
+        BRANCH: ${{ inputs.branch }}
+        BUILD_ENVIRONMENT: pytorch-linux-xenial-py3-clang5-android-ndk-r19c-${{ inputs.arch-for-build-env }}-build"
+        AWS_DEFAULT_REGION: us-east-1
+        PR_NUMBER: ${{ github.event.pull_request.number }}
+        SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+        SCCACHE_BUCKET: ossci-compiler-cache-circleci-v2
+        SCCACHE_REGION: us-east-1
+        DOCKER_IMAGE: ${{ inputs.docker-image  }}
+        MATRIX_ARCH: ${{ inputs.arch }}
+      run: |
+        # detached container should get cleaned up by teardown_ec2_linux
+        set -exo pipefail
+        export container_name
+        container_name=$(docker run \
+          -e BUILD_ENVIRONMENT \
+          -e MAX_JOBS="$(nproc --ignore=2)" \
+          -e AWS_DEFAULT_REGION \
+          -e PR_NUMBER \
+          -e SHA1 \
+          -e BRANCH \
+          -e SCCACHE_BUCKET \
+          -e SCCACHE_REGION \
+          -e SKIP_SCCACHE_INITIALIZATION=1 \
+          --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
+          --security-opt seccomp=unconfined \
+          --cap-add=SYS_PTRACE \
+          --tty \
+          --detach \
+          --user jenkins \
+          -w /var/lib/jenkins/workspace \
+          "${DOCKER_IMAGE}"
+        )
+        git submodule sync && git submodule update -q --init --recursive --depth 1
+        docker cp "${GITHUB_WORKSPACE}/." "${container_name}:/var/lib/jenkins/workspace"
+        (echo "sudo chown -R jenkins . && .ci/pytorch/build.sh && find ${BUILD_ROOT} -type f -name "*.a" -or -name "*.o" -delete" | docker exec -u jenkins -i "${container_name}" bash) 2>&1
+
+        # Copy install binaries back
+        mkdir -p "${GITHUB_WORKSPACE}/build_android_install_${MATRIX_ARCH}"
+        docker cp "${container_name}:/var/lib/jenkins/workspace/build_android/install" "${GITHUB_WORKSPACE}/build_android_install_${MATRIX_ARCH}"
+        echo "container_id=${container_name}" >> "${GITHUB_OUTPUT}"

.github/actions/filter-test-configs/action.yml

@@ -70,7 +70,7 @@ runs:
           set -eux
           # PyYAML 6.0 doesn't work with MacOS x86 anymore
           # This must run on Python-3.7 (AmazonLinux2) so can't use request=3.32.2
-          python3 -m pip install requests==2.27.1 pyyaml==6.0.2
+          python3 -m pip install requests==2.27.1 pyyaml==6.0.1
 
     - name: Parse ref
       id: parse-ref
@@ -125,7 +125,7 @@ runs:
         TAG: ${{ steps.parse-ref.outputs.tag }}
         EVENT_NAME: ${{ github.event_name }}
         SCHEDULE: ${{ github.event.schedule }}
-        HEAD_BRANCH: ${{ steps.parse-ref.outputs.branch }}
+        HEAD_BRANCH: ${{ github.event.workflow_run.head_branch }}
       id: filter
       run: |
         echo "Workflow: ${GITHUB_WORKFLOW}"

.github/actions/linux-test/action.yml

@@ -126,7 +126,7 @@ runs:
       shell: bash
       continue-on-error: true
       run: |
-        python3 -m pip install psutil==5.9.8 nvidia-ml-py==11.525.84
+        python3 -m pip install psutil==5.9.1 nvidia-ml-py==11.525.84
         python3 -m tools.stats.monitor > usage_log.txt 2>&1 &
         echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 

.github/actions/reuse-old-whl/reuse_old_whl.py

@@ -304,7 +304,8 @@ def unzip_artifact_and_replace_files() -> None:
 
 
 def set_output() -> None:
-    print("Setting output reuse=true")
+    # Disable for now so we can monitor first
+    # pass
     if os.getenv("GITHUB_OUTPUT"):
         with open(str(os.getenv("GITHUB_OUTPUT")), "a") as env:
             print("reuse=true", file=env)

.github/actions/test-pytorch-binary/action.yml

@@ -24,6 +24,7 @@ runs:
           -e PYTORCH_FINAL_PACKAGE_DIR \
           -e PYTORCH_ROOT \
           -e SKIP_ALL_TESTS \
+          -e USE_SPLIT_BUILD \
           --tty \
           --detach \
           -v "${GITHUB_WORKSPACE}/pytorch:/pytorch" \

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-e500f0cf88bc57ffd8b0029033da305eef24ae25
+4e94321c54617dd738a05bfedfc28bc0fa635b5c

.github/ci_commit_pins/fbgemm_rocm.txt

@@ -1 +1 @@
-7f1de94a4c2d14f59ad4ca84538c36084ea6b2c8
+5fb5024118e9bb9decf96c2b0b1a8f0010bf56be

.github/ci_commit_pins/vllm.txt

@@ -1 +0,0 @@
-35afe1b30b154114dc2ee8329e12f8cf3fe9f576

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-095faec1e7b6cc47220181e74ae9cde2605f9b00
+55a75404c9b75cd5fd62ab5d4deafc8c506b3af2

.github/label_to_label.yml

@@ -48,12 +48,3 @@
   - "module: dynamic shapes"
   then:
   - "oncall: pt2"
-- any:
-  - "release notes: distributed (c10d)"
-  - "release notes: distributed (symm_mem)"
-  - "release notes: distributed (pipeline)"
-  - "release notes: distributed (fsdp)"
-  - "release notes: distributed (dtensor)"
-  - "oncall: distributed"
-  then:
-  - "ciflow/h100-distributed"

.github/merge_rules.yaml

@@ -76,7 +76,6 @@
   - .github/ci_commit_pins/audio.txt
   - .github/ci_commit_pins/vision.txt
   - .github/ci_commit_pins/torchdynamo.txt
-  - .github/ci_commit_pins/vllm.txt
   - .ci/docker/ci_commit_pins/triton.txt
   approved_by:
   - pytorchbot
@@ -131,6 +130,21 @@
   - Lint
   - pull
 
+- name: Mobile
+  patterns:
+  - ios/**
+  - android/**
+  - test/mobile/**
+  approved_by:
+  - linbinyu
+  - IvanKobzarev
+  - dreiss
+  - raziel
+  mandatory_checks_name:
+  - EasyCLA
+  - Lint
+  - pull
+
 - name: PrimTorch
   patterns:
   - torch/_meta_registrations.py
@@ -370,7 +384,6 @@
   - leslie-fang-intel
   - jgong5
   - EikanWang
-  - CaoE
   mandatory_checks_name:
   - EasyCLA
   - Lint
@@ -422,7 +435,6 @@
   approved_by:
   - leslie-fang-intel
   - jgong5
-  - CaoE
   mandatory_checks_name:
   - EasyCLA
   - Lint
@@ -477,23 +489,6 @@
   - srossross
   - chillee
   - zou3519
-  - guilhermeleobas
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
-- name: Dynamo
-  patterns:
-  - torch/_dynamo/**
-  - torch/csrc/dynamo/**
-  - test/dynamo/**
-  - test/dynamo_expected_failures/**
-  - test/dynamo_skips/**
-  - test/inductor_expected_failures/**
-  - test/inductor_skips/**
-  approved_by:
-  - guilhermeleobas
   mandatory_checks_name:
   - EasyCLA
   - Lint

.github/pytorch-probot.yml

@@ -4,7 +4,6 @@ ciflow_push_tags:
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
-- ciflow/triton_binaries
 - ciflow/inductor
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
@@ -31,9 +30,6 @@ ciflow_push_tags:
 - ciflow/pull
 - ciflow/h100
 - ciflow/h100-distributed
-- ciflow/win-arm64
-- ciflow/h100-symm-mem
-- ciflow/h100-cutlass-backend
 retryable_workflows:
 - pull
 - trunk

.github/requirements-gha-cache.txt

@@ -1,15 +1,14 @@
 # This file is to cache other dependencies not specified elsewhere in:
-#   requirements.txt
-#   requirements-build.txt
+#   requirement.txt
 #   docs/requirements.txt
 #   docs/cpp/requirements.txt
 #   functorch/docs/requirements.txt
 #   .ci/docker/requirements-ci.txt
 boto3==1.35.42
 jinja2==3.1.6
-lintrunner==0.12.7
+lintrunner==0.10.7
 ninja==1.10.0.post1
 nvidia-ml-py==11.525.84
-pyyaml==6.0.2
+pyyaml==6.0
 requests==2.32.4
-rich==14.1.0
+rich==10.9.0

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

@@ -2,7 +2,7 @@ boto3==1.35.42
 cmake==3.27.*
 expecttest==0.3.0
 fbscribelogger==0.1.7
-filelock==3.18.0
+filelock==3.6.0
 hypothesis==6.56.4
 librosa>=0.6.2
 mpmath==1.3.0
@@ -16,7 +16,7 @@ packaging==23.1
 parameterized==0.8.1
 pillow==10.3.0
 protobuf==5.29.4
-psutil==5.9.8
+psutil==5.9.1
 pygments==2.15.0
 pytest-cpp==2.3.0
 pytest-flakefinder==1.1.0
@@ -33,4 +33,4 @@ tensorboard==2.13.0
 typing-extensions==4.12.2
 unittest-xml-reporting<=3.2.0,>=2.0.0
 xdoctest==1.1.0
-z3-solver==4.15.1.0
+z3-solver==4.12.2.0

.github/scripts/delete_old_branches.py

@@ -275,7 +275,7 @@ def delete_branches() -> None:
         delete_branch(git_repo, branch)
 
 
-def delete_old_tags() -> None:
+def delete_old_ciflow_tags() -> None:
     # Deletes ciflow tags if they are associated with a closed PR or a specific
     # commit.  Lightweight tags don't have information about the date they were
     # created, so we can't check how old they are.  The script just assumes that
@@ -288,29 +288,23 @@ def delete_old_tags() -> None:
         delete_branch(git_repo, f"refs/tags/{tag}")
 
     tags = git_repo._run_git("tag").splitlines()
+    open_pr_numbers = [x["number"] for x in get_open_prs()]
 
-    CIFLOW_TAG_REGEX = re.compile(r"^ciflow\/.*\/(\d{5,6}|[0-9a-f]{40})$")
-    AUTO_REVERT_TAG_REGEX = re.compile(r"^trunk\/[0-9a-f]{40}$")
     for tag in tags:
         try:
             if ESTIMATED_TOKENS[0] > 400:
                 print("Estimated tokens exceeded, exiting")
                 break
-
-            if not CIFLOW_TAG_REGEX.match(tag) and not AUTO_REVERT_TAG_REGEX.match(tag):
+            if not tag.startswith("ciflow/"):
                 continue
-
-            # This checks the date of the commit associated with the tag instead
-            # of the tag itself since lightweight tags don't have this
-            # information.  I think it should be ok since this only runs once a
-            # day
-            tag_info = git_repo._run_git("show", "-s", "--format=%ct", tag)
-            tag_timestamp = int(tag_info.strip())
-            # Maybe some timezone issues, but a few hours shouldn't matter
-            tag_age_days = (datetime.now().timestamp() - tag_timestamp) / SEC_IN_DAY
-
-            if tag_age_days > 7:
-                print(f"[{tag}] Tag is older than 7 days, deleting")
+            re_match_pr = re.match(r"^ciflow\/.*\/(\d{5,6})$", tag)
+            re_match_sha = re.match(r"^ciflow\/.*\/([0-9a-f]{40})$", tag)
+            if re_match_pr:
+                pr_number = int(re_match_pr.group(1))
+                if pr_number in open_pr_numbers:
+                    continue
+                delete_tag(tag)
+            elif re_match_sha:
                 delete_tag(tag)
         except Exception as e:
             print(f"Failed to check tag {tag}: {e}")
@@ -318,4 +312,4 @@ def delete_old_tags() -> None:
 
 if __name__ == "__main__":
     delete_branches()
-    delete_old_tags()
+    delete_old_ciflow_tags()

.github/scripts/filter_test_configs.py

@@ -18,7 +18,6 @@ import yaml
 
 
 REENABLE_TEST_REGEX = "(?i)(Close(d|s)?|Resolve(d|s)?|Fix(ed|es)?) (#|https://github.com/pytorch/pytorch/issues/)([0-9]+)"
-MAIN_BRANCH = "main"
 
 PREFIX = "test-config/"
 
@@ -98,7 +97,7 @@ def parse_args() -> Any:
     parser.add_argument(
         "--branch",
         type=str,
-        default=MAIN_BRANCH,
+        default="main",
         help="the branch name",
     )
     return parser.parse_args()
@@ -457,7 +456,6 @@ def download_json(url: str, headers: dict[str, str], num_retries: int = 3) -> An
 
 
 def set_output(name: str, val: Any) -> None:
-    print(f"Setting output {name}={val}")
     if os.getenv("GITHUB_OUTPUT"):
         with open(str(os.getenv("GITHUB_OUTPUT")), "a") as env:
             print(f"{name}={val}", file=env)
@@ -497,20 +495,13 @@ def check_for_setting(labels: set[str], body: str, setting: str) -> bool:
 
 
 def perform_misc_tasks(
-    labels: set[str],
-    test_matrix: dict[str, list[Any]],
-    job_name: str,
-    pr_body: str,
-    branch: Optional[str] = None,
+    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
     misc tasks to set keep-going and is-unstable variables
     """
-    set_output(
-        "keep-going",
-        branch == MAIN_BRANCH or check_for_setting(labels, pr_body, "keep-going"),
-    )
+    set_output("keep-going", check_for_setting(labels, pr_body, "keep-going"))
     set_output(
         "ci-verbose-test-logs",
         check_for_setting(labels, pr_body, "ci-verbose-test-logs"),
@@ -633,7 +624,6 @@ def main() -> None:
         test_matrix=filtered_test_matrix,
         job_name=args.job_name,
         pr_body=pr_body if pr_body else "",
-        branch=args.branch,
     )
 
     # Set the filtered test matrix as the output

.github/scripts/generate_binary_build_matrix.py

@@ -17,7 +17,7 @@ from typing import Optional
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
 CUDA_ARCHES = ["12.6", "12.8", "12.9"]
-CUDA_STABLE = "12.8"
+CUDA_STABLE = "12.6"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
@@ -53,8 +53,8 @@ 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==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.3; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.2.5; 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'"
@@ -70,8 +70,8 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.3; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.2.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'"
@@ -87,8 +87,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.3; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'"
@@ -193,7 +192,7 @@ LIBTORCH_CONTAINER_IMAGES: dict[str, str] = {
     "cpu": "libtorch-cxx11-builder:cpu",
 }
 
-FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
+FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t"]
 
 
 def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
@@ -225,6 +224,8 @@ def generate_libtorch_matrix(
             arches += ROCM_ARCHES
         elif os == "windows":
             arches += CUDA_ARCHES
+            if "12.9" in arches:
+                arches.remove("12.9")
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -273,6 +274,7 @@ def generate_wheels_matrix(
     os: str,
     arches: Optional[list[str]] = None,
     python_versions: Optional[list[str]] = None,
+    use_split_build: bool = False,
 ) -> list[dict[str, str]]:
     package_type = "wheel"
     if os == "linux" or os == "linux-aarch64" or os == "linux-s390x":
@@ -289,6 +291,9 @@ def generate_wheels_matrix(
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
             arches += CUDA_ARCHES + XPU_ARCHES
+            # skip CUDA 12.9 builds on Windows
+            if "12.9" in arches:
+                arches.remove("12.9")
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -314,11 +319,15 @@ def generate_wheels_matrix(
             # TODO: Enable python 3.13t on cpu-s390x
             if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
                 continue
-            # TODO: Enable python 3.14 on non linux OSes
-            if os != "linux" and (
-                python_version == "3.14" or python_version == "3.14t"
+
+            if use_split_build and (
+                arch_version not in ["12.6", "12.8", "12.9", "cpu"] or os != "linux"
             ):
-                continue
+                raise RuntimeError(
+                    "Split build is only supported on linux with cuda 12* 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."
+                )
 
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
@@ -334,6 +343,7 @@ def generate_wheels_matrix(
                         "gpu_arch_type": gpu_arch_type,
                         "gpu_arch_version": gpu_arch_version,
                         "desired_cuda": desired_cuda,
+                        "use_split_build": "True" if use_split_build else "False",
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],
@@ -366,6 +376,7 @@ def generate_wheels_matrix(
                             "desired_cuda": translate_desired_cuda(
                                 gpu_arch_type, gpu_arch_version
                             ),
+                            "use_split_build": "True" if use_split_build else "False",
                             "container_image": WHEEL_CONTAINER_IMAGES[
                                 arch_version
                             ].split(":")[0],
@@ -388,6 +399,7 @@ def generate_wheels_matrix(
                         "desired_cuda": translate_desired_cuda(
                             gpu_arch_type, gpu_arch_version
                         ),
+                        "use_split_build": "True" if use_split_build else "False",
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],

.github/scripts/generate_ci_workflows.py

@@ -22,7 +22,6 @@ LABEL_CIFLOW_BINARIES = "ciflow/binaries"
 LABEL_CIFLOW_PERIODIC = "ciflow/periodic"
 LABEL_CIFLOW_BINARIES_LIBTORCH = "ciflow/binaries_libtorch"
 LABEL_CIFLOW_BINARIES_WHEEL = "ciflow/binaries_wheel"
-LABEL_CIFLOW_ROCM = "ciflow/rocm"
 
 
 @dataclass
@@ -59,7 +58,9 @@ class BinaryBuildWorkflow:
     is_scheduled: str = ""
     branches: str = "nightly"
     # Mainly for macos
+    cross_compile_arm64: bool = False
     macos_runner: str = "macos-14-xlarge"
+    use_split_build: bool = False
     # Mainly used for libtorch builds
     build_variant: str = ""
 
@@ -70,6 +71,9 @@ class BinaryBuildWorkflow:
                 for item in [self.os, "binary", self.package_type, self.build_variant]
                 if item != ""
             )
+        if self.use_split_build:
+            # added to distinguish concurrency groups
+            self.build_environment += "-split"
 
     def generate_workflow_file(self, workflow_template: jinja2.Template) -> None:
         output_file_path = (
@@ -112,6 +116,21 @@ LINUX_BINARY_BUILD_WORFKLOWS = [
             isolated_workflow=True,
         ),
     ),
+    # See https://github.com/pytorch/pytorch/issues/138750
+    #   BinaryBuildWorkflow(
+    #     os=OperatingSystem.LINUX,
+    #     package_type="manywheel",
+    #     build_configs=generate_binary_build_matrix.generate_wheels_matrix(
+    #         OperatingSystem.LINUX,
+    #         use_split_build=True,
+    #         arches=["11.8", "12.1", "12.4", "cpu"],
+    #     ),
+    #     ciflow_config=CIFlowConfig(
+    #         labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},
+    #         isolated_workflow=True,
+    #     ),
+    #     use_split_build=True,
+    # ),
     BinaryBuildWorkflow(
         os=OperatingSystem.LINUX,
         package_type="libtorch",
@@ -127,39 +146,33 @@ LINUX_BINARY_BUILD_WORFKLOWS = [
     ),
 ]
 
-ROCM_SMOKE_WORKFLOWS = [
-    BinaryBuildWorkflow(
-        os=OperatingSystem.LINUX,
-        package_type="manywheel",
-        build_variant="rocm",
-        build_configs=generate_binary_build_matrix.generate_wheels_matrix(
-            OperatingSystem.LINUX,
-            arches=["6.4"],
-            python_versions=["3.9"],
-        ),
-        ciflow_config=CIFlowConfig(
-            labels={
-                LABEL_CIFLOW_BINARIES,
-                LABEL_CIFLOW_BINARIES_WHEEL,
-                LABEL_CIFLOW_ROCM,
-            },
-            isolated_workflow=True,
-        ),
-        branches="main",
-    ),
-]
-
 LINUX_BINARY_SMOKE_WORKFLOWS = [
     BinaryBuildWorkflow(
         os=OperatingSystem.LINUX,
         package_type="manywheel",
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
-            arches=["12.8"],
-            python_versions=["3.12"],
+            arches=["12.6", "12.8", "12.9", "6.4"],
+            python_versions=["3.9"],
         ),
         branches="main",
     ),
+    # See https://github.com/pytorch/pytorch/issues/138750
+    # BinaryBuildWorkflow(
+    #     os=OperatingSystem.LINUX,
+    #     package_type="manywheel",
+    #     build_configs=generate_binary_build_matrix.generate_wheels_matrix(
+    #         OperatingSystem.LINUX,
+    #         arches=["11.8", "12.1", "12.4"],
+    #         python_versions=["3.9"],
+    #         use_split_build=True,
+    #     ),
+    #     ciflow_config=CIFlowConfig(
+    #         labels={LABEL_CIFLOW_PERIODIC},
+    #     ),
+    #     branches="main",
+    #     use_split_build=True,
+    # ),
     BinaryBuildWorkflow(
         os=OperatingSystem.LINUX,
         package_type="libtorch",
@@ -302,6 +315,7 @@ MACOS_BINARY_BUILD_WORKFLOWS = [
             generate_binary_build_matrix.RELEASE,
             libtorch_variants=["shared-with-deps"],
         ),
+        cross_compile_arm64=False,
         macos_runner="macos-14-xlarge",
         ciflow_config=CIFlowConfig(
             labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_LIBTORCH},
@@ -314,6 +328,7 @@ MACOS_BINARY_BUILD_WORKFLOWS = [
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.MACOS_ARM64
         ),
+        cross_compile_arm64=False,
         macos_runner="macos-14-xlarge",
         ciflow_config=CIFlowConfig(
             labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},
@@ -372,11 +387,6 @@ def main() -> None:
             jinja_env.get_template("linux_binary_build_workflow.yml.j2"),
             S390X_BINARY_BUILD_WORKFLOWS,
         ),
-        (
-            # Give rocm it's own workflow file
-            jinja_env.get_template("linux_binary_build_workflow.yml.j2"),
-            ROCM_SMOKE_WORKFLOWS,
-        ),
         (
             jinja_env.get_template("linux_binary_build_workflow.yml.j2"),
             LINUX_BINARY_SMOKE_WORKFLOWS,

.github/scripts/get_workflow_job_id.py

@@ -136,10 +136,10 @@ def find_job_id_name(args: Any) -> tuple[str, str]:
 
 
 def set_output(name: str, val: Any) -> None:
-    print(f"Setting output {name}={val}")
     if os.getenv("GITHUB_OUTPUT"):
         with open(str(os.getenv("GITHUB_OUTPUT")), "a") as env:
             print(f"{name}={val}", file=env)
+        print(f"setting {name}={val}")
     else:
         print(f"::set-output name={name}::{val}")
 

.github/scripts/lintrunner.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Use uv to speed up lintrunner init
-python3 -m pip install -U uv==0.8.* setuptools
+python3 -m pip install uv==0.1.45 setuptools
 
 CACHE_DIRECTORY="/tmp/.lintbin"
 # Try to recover the cached binaries

.github/scripts/parse_ref.py

@@ -5,7 +5,6 @@ import re
 
 
 def set_output(name: str, val: str) -> None:
-    print(f"Setting output {name}={val}")
     if os.getenv("GITHUB_OUTPUT"):
         with open(str(os.getenv("GITHUB_OUTPUT")), "a") as env:
             print(f"{name}={val}", file=env)

.github/scripts/runner_determinator.py

@@ -262,12 +262,7 @@ def is_exception_branch(branch: str) -> bool:
     """
     Branches that get opted out of experiments by default, until they're explicitly enabled.
     """
-    return branch.split("/", maxsplit=1)[0] in {
-        "main",
-        "nightly",
-        "release",
-        "landchecks",
-    }
+    return branch.split("/")[0] in {"main", "nightly", "release", "landchecks"}
 
 
 def load_yaml(yaml_text: str) -> Any:

.github/scripts/tag_docker_images_for_release.py

@@ -0,0 +1,64 @@
+import argparse
+import subprocess
+
+import generate_binary_build_matrix
+
+
+def tag_image(
+    image: str,
+    default_tag: str,
+    release_version: str,
+    dry_run: str,
+    tagged_images: dict[str, bool],
+) -> None:
+    if image in tagged_images:
+        return
+    release_image = image.replace(f"-{default_tag}", f"-{release_version}")
+    print(f"Tagging {image} to {release_image} , dry_run: {dry_run}")
+
+    if dry_run == "disabled":
+        subprocess.check_call(["docker", "pull", image])
+        subprocess.check_call(["docker", "tag", image, release_image])
+        subprocess.check_call(["docker", "push", release_image])
+    tagged_images[image] = True
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--version",
+        help="Version to tag",
+        type=str,
+        default="2.2",
+    )
+    parser.add_argument(
+        "--dry-run",
+        help="No Runtime Error check",
+        type=str,
+        choices=["enabled", "disabled"],
+        default="enabled",
+    )
+
+    options = parser.parse_args()
+    tagged_images: dict[str, bool] = {}
+    platform_images = [
+        generate_binary_build_matrix.WHEEL_CONTAINER_IMAGES,
+        generate_binary_build_matrix.LIBTORCH_CONTAINER_IMAGES,
+    ]
+    default_tag = generate_binary_build_matrix.DEFAULT_TAG
+
+    for platform_image in platform_images:  # type: ignore[attr-defined]
+        for arch in platform_image.keys():  # type: ignore[attr-defined]
+            if arch == "cpu-s390x":
+                continue
+            tag_image(
+                platform_image[arch],  # type: ignore[index]
+                default_tag,
+                options.version,
+                options.dry_run,
+                tagged_images,
+            )
+
+
+if __name__ == "__main__":
+    main()

.github/scripts/td_llm_indexer.sh

@@ -6,7 +6,7 @@ set -euxo pipefail
 cd llm-target-determinator
 pip install -q -r requirements.txt
 cd ../codellama
-pip install --no-build-isolation -v -e .
+pip install -e .
 pip install numpy==1.26.0
 
 # Run indexer

.github/scripts/test_delete_old_branches.py

@@ -1,56 +0,0 @@
-import os
-import unittest
-from datetime import datetime
-from unittest.mock import MagicMock, patch
-
-
-os.environ["GITHUB_TOKEN"] = "test_token"
-
-from delete_old_branches import delete_old_tags
-
-
-@patch("delete_old_branches.delete_branch")
-@patch("gitutils.GitRepo._run_git")
-class TestDeleteTag(unittest.TestCase):
-    def test_delete_tag(
-        self, mock_run_git: "MagicMock", mock_delete_tag: "MagicMock"
-    ) -> None:
-        for tag in [
-            "ciflow/branch/12345",
-            "ciflow/commitsha/1234567890abcdef1234567890abcdef12345678",
-            "trunk/1234567890abcdef1234567890abcdef12345678",
-        ]:
-            mock_run_git.side_effect = [
-                tag,
-                str(int(datetime.now().timestamp() - 8 * 24 * 60 * 60)),  # 8 days ago
-            ]
-            delete_old_tags()
-            mock_delete_tag.assert_called_once()
-            mock_delete_tag.reset_mock()
-
-            # Don't delete if the tag is not old enough
-            mock_run_git.side_effect = [
-                tag,
-                str(int(datetime.now().timestamp() - 6 * 24 * 60 * 60)),  # 6 days ago
-            ]
-            delete_old_tags()
-            mock_delete_tag.assert_not_called()
-
-    def test_do_not_delete_tag(
-        self, mock_run_git: "MagicMock", mock_delete_tag: "MagicMock"
-    ) -> None:
-        for tag in [
-            "ciflow/doesntseemtomatch",
-            "trunk/doesntseemtomatch",
-            "doesntseemtomatch",
-        ]:
-            mock_run_git.side_effect = [
-                tag,
-                str(int(datetime.now().timestamp() - 8 * 24 * 60 * 60)),  # 8 days ago
-            ]
-            delete_old_tags()
-            mock_delete_tag.assert_not_called()
-
-
-if __name__ == "__main__":
-    unittest.main()

.github/scripts/trymerge.py

@@ -1891,9 +1891,7 @@ def validate_revert(
         else pr.get_comment_by_id(comment_id)
     )
     if comment.editor_login is not None:
-        raise PostCommentError(
-            "Halting the revert as the revert comment has been edited."
-        )
+        raise PostCommentError("Don't want to revert based on edited command")
     author_association = comment.author_association
     author_login = comment.author_login
     allowed_reverters = ["COLLABORATOR", "MEMBER", "OWNER"]

.github/templates/macos_binary_build_workflow.yml.j2

@@ -47,6 +47,9 @@ env:
   GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PR_NUMBER: ${{ github.event.pull_request.number }}
   SKIP_ALL_TESTS: 0
+{%- if cross_compile_arm64 %}
+  CROSS_COMPILE_ARM64: 1
+{% endif %}
 !{{ common.concurrency(build_environment) }}
 
 jobs:

.github/templates/upload.yml.j2

@@ -25,6 +25,11 @@
       DOCKER_IMAGE: !{{ config["container_image"] }}
       DOCKER_IMAGE_TAG_PREFIX: !{{ config["container_image_tag_prefix"] }}
 {%- endif %}
+{%- if config["package_type"] == "manywheel" %}
+  {%- if config.use_split_build is defined %}
+      use_split_build: !{{ config["use_split_build"] }}
+  {%- endif %}
+{%- endif %}
 {%- if config["package_type"] == "libtorch" %}
   {%- if config["libtorch_config"] %}
       LIBTORCH_CONFIG: !{{ config["libtorch_config"] }}

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

@@ -26,6 +26,13 @@ on:
         default: 240
         type: number
         description: timeout for the job
+      use_split_build:
+        description: |
+          [Experimental] Build a libtorch only wheel and build pytorch such that
+          are built from the libtorch wheel.
+        required: false
+        type: boolean
+        default: false
       ALPINE_IMAGE:
         required: false
         type: string
@@ -110,6 +117,7 @@ jobs:
       PR_NUMBER: ${{ github.event.pull_request.number }}
       PYTORCH_FINAL_PACKAGE_DIR: /artifacts
       SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+      USE_SPLIT_BUILD: ${{ inputs.use_split_build }}
     steps:
       - name: Make the env permanent during this workflow (but not the secrets)
         shell: bash
@@ -134,6 +142,7 @@ jobs:
             echo "PR_NUMBER=${{ env.PR_NUMBER }}"
             echo "PYTORCH_FINAL_PACKAGE_DIR=${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
             echo "SHA1=${{ env.SHA1 }}"
+            echo "USE_SPLIT_BUILD=${{ env.use_split_build }}"
           } >> "${GITHUB_ENV} }}"
 
       - name: List the env
@@ -252,6 +261,7 @@ jobs:
             -e PYTORCH_ROOT \
             -e SKIP_ALL_TESTS \
             -e PYTORCH_EXTRA_INSTALL_REQUIREMENTS \
+            -e USE_SPLIT_BUILD \
             --tty \
             --detach \
             -v "${GITHUB_WORKSPACE}/pytorch:/pytorch" \

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

@@ -64,6 +64,13 @@ on:
         required: true
         type: string
         description: Hardware to run this job on. Valid values are linux.4xlarge, linux.4xlarge.nvidia.gpu, linux.arm64.2xlarge, and linux.rocm.gpu
+      use_split_build:
+        description: |
+          [Experimental] Build a libtorch only wheel and build pytorch such that
+          are built from the libtorch wheel.
+        required: false
+        type: boolean
+        default: false
     secrets:
       github-token:
         required: true
@@ -97,6 +104,7 @@ jobs:
       PR_NUMBER: ${{ github.event.pull_request.number }}
       PYTORCH_FINAL_PACKAGE_DIR: /artifacts
       SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+      USE_SPLIT_BUILD: ${{ inputs.use_split_build }}
     steps:
       - name: Make the env permanent during this workflow (but not the secrets)
         shell: bash
@@ -121,6 +129,7 @@ jobs:
             echo "PR_NUMBER=${{ env.PR_NUMBER }}"
             echo "PYTORCH_FINAL_PACKAGE_DIR=${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
             echo "SHA1=${{ env.SHA1 }}"
+            echo "USE_SPLIT_BUILD=${{ env.USE_SPLIT_BUILD }}"
           } >> "${GITHUB_ENV} }}"
 
       - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"

.github/workflows/_binary-upload.yml

@@ -51,6 +51,13 @@ on:
         required: false
         type: string
         description: Desired python version
+      use_split_build:
+        description: |
+          [Experimental] Build a libtorch only wheel and build pytorch such that
+          are built from the libtorch wheel.
+        required: false
+        type: boolean
+        default: false
     secrets:
       github-token:
         required: true
@@ -79,6 +86,7 @@ jobs:
       PR_NUMBER: ${{ github.event.pull_request.number }}
       PYTORCH_FINAL_PACKAGE_DIR: /artifacts
       SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+      USE_SPLIT_BUILD: ${{ inputs.use_split_build }}
     steps:
       - name: Checkout PyTorch
         uses: pytorch/pytorch/.github/actions/checkout-pytorch@main

.github/workflows/_docs.yml

@@ -70,7 +70,7 @@ jobs:
             runner: ${{ inputs.runner_prefix }}linux.12xlarge
             # TODO: Nightly cpp docs take longer and longer to finish (more than 3h now)
             # Let's try to figure out how this can be improved
-            timeout-minutes: 360
+            timeout-minutes: 240
           - docs_type: python
             runner: ${{ inputs.runner_prefix }}linux.2xlarge
             # It takes less than 30m to finish python docs unless there are issues

.github/workflows/_get-changed-files.yml

@@ -1,43 +0,0 @@
-name: Get Changed Files
-
-on:
-  workflow_call:
-    outputs:
-      changed-files:
-        description: "List of changed files (space-separated) or '*' if not in a PR"
-        value: ${{ jobs.get-changed-files.outputs.changed-files }}
-
-jobs:
-  get-changed-files:
-    runs-on: ubuntu-latest
-    outputs:
-      changed-files: ${{ steps.get-files.outputs.changed-files }}
-
-    steps:
-      - name: Get changed files
-        id: get-files
-        env:
-          GH_TOKEN: ${{ github.token }}
-        run: |
-          # Check if we're in a pull request context
-          if [ "${{ github.event_name }}" = "pull_request" ] || [ "${{ github.event_name }}" = "pull_request_target" ]; then
-            echo "Running in PR context"
-
-            # Get the PR number from the github context
-            PR_NUMBER="${{ github.event.number }}"
-
-            # Use gh CLI to get changed files in the PR with explicit repo
-            CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
-
-            if [ -z "$CHANGED_FILES" ]; then
-              echo "No changed files found, setting to '*'"
-              CHANGED_FILES="*"
-            fi
-
-            echo "Changed files: $CHANGED_FILES"
-            echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
-
-          else
-            echo "Not in PR context, setting changed files to '*'"
-            echo "changed-files=*" >> "$GITHUB_OUTPUT"
-          fi

.github/workflows/_linux-build.yml

@@ -16,6 +16,11 @@ on:
         type: boolean
         default: true
         description: If set, upload generated build artifacts.
+      build-with-debug:
+        required: false
+        type: boolean
+        default: false
+        description: If set, build in debug mode.
       sync-tag:
         required: false
         type: string
@@ -64,6 +69,11 @@ on:
         required: false
         type: string
         default: ""
+      max-jobs:
+        description: |
+          Overwrite the number of jobs to use for the build
+        required: false
+        type: string
       disable-monitor:
         description: |
           Disable utilization monitoring for build job
@@ -82,6 +92,7 @@ on:
         required: false
         type: number
         default: 1
+
       allow-reuse-old-whl:
         description: |
           If set, the build try to pull an old wheel from s3 that was built on a
@@ -89,13 +100,6 @@ on:
         required: false
         type: boolean
         default: true
-      build-additional-packages:
-        description: |
-          If set, the build job will also builds these packages and saves their
-          wheels as artifacts
-        required: false
-        type: string
-        default: ""
 
     secrets:
       HUGGING_FACE_HUB_TOKEN:
@@ -107,6 +111,7 @@ on:
         description: |
           FB app token to write to scribe endpoint
 
+
     outputs:
       docker-image:
         value: ${{ jobs.build.outputs.docker-image }}
@@ -131,9 +136,6 @@ jobs:
         if: inputs.build-environment != 'linux-s390x-binary-manywheel'
         with:
           github-secret: ${{ secrets.GITHUB_TOKEN }}
-          instructions: |
-            Build is done inside the container, to start an interactive session run:
-              docker exec -it $(docker container ps --format '{{.ID}}') bash
 
       # [pytorch repo ref]
       # Use a pytorch/pytorch reference instead of a reference to the local
@@ -225,7 +227,7 @@ jobs:
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
         run: |
           mkdir -p ../../usage_logs
-          python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7
+          python3 -m pip install psutil==5.9.1 dataclasses_json==0.6.7
           python3 -m tools.stats.monitor \
           --log-interval "$MONITOR_LOG_INTERVAL" \
           --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" \
@@ -247,6 +249,8 @@ jobs:
         env:
           BUILD_ENVIRONMENT: ${{ inputs.build-environment }}
           BRANCH: ${{ steps.parse-ref.outputs.branch }}
+          # TODO duplicated
+          AWS_DEFAULT_REGION: us-east-1
           PR_NUMBER: ${{ github.event.pull_request.number }}
           SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
           # Do not set SCCACHE_S3_KEY_PREFIX to share the cache between all build jobs
@@ -258,10 +262,11 @@ jobs:
           DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
           DOCKER_IMAGE_S390X: ${{ inputs.docker-image-name }}
           XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
+          DEBUG: ${{ inputs.build-with-debug && '1' || '0' }}
           OUR_GITHUB_JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
           HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
           SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.SCRIBE_GRAPHQL_ACCESS_TOKEN }}
-          BUILD_ADDITIONAL_PACKAGES: ${{ inputs.build-additional-packages }}
+          MAX_JOBS_OVERRIDE: ${{ inputs.max-jobs }}
         run: |
           START_TIME=$(date +%s)
           if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
@@ -281,6 +286,12 @@ jobs:
             DOCKER_SHELL_CMD=
           fi
 
+          if [[ ${MAX_JOBS_OVERRIDE} == "" ]]; then
+            MAX_JOBS="$(nproc --ignore=2)"
+          else
+            MAX_JOBS="${MAX_JOBS_OVERRIDE}"
+          fi
+
           # Leaving 1GB for the runner and other things
           TOTAL_AVAILABLE_MEMORY_IN_GB=$(awk '/MemTotal/ { printf "%.3f \n", $2/1024/1024 - 1 }' /proc/meminfo)
           # https://docs.docker.com/engine/containers/resource_constraints/#--memory-swap-details, the 3GB swap
@@ -292,7 +303,9 @@ jobs:
           # shellcheck disable=SC2086
           container_name=$(docker run \
             -e BUILD_ENVIRONMENT \
-            -e MAX_JOBS="$(nproc --ignore=2)" \
+            -e MAX_JOBS=${MAX_JOBS} \
+            -e MAX_JOBS_OVERRIDE \
+            -e AWS_DEFAULT_REGION \
             -e PR_NUMBER \
             -e SHA1 \
             -e BRANCH \
@@ -306,7 +319,7 @@ jobs:
             -e OUR_GITHUB_JOB_ID \
             -e HUGGING_FACE_HUB_TOKEN \
             -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
-            -e BUILD_ADDITIONAL_PACKAGES \
+            -e USE_SPLIT_BUILD \
             --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
             --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
@@ -320,11 +333,6 @@ jobs:
             "${USED_IMAGE}" \
             ${DOCKER_SHELL_CMD}
           )
-
-          if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
-            docker exec -t "${container_name}" sh -c "python3 -m pip install -r requirements.txt"
-          fi
-
           docker exec -t "${container_name}" sh -c '.ci/pytorch/build.sh'
 
           END_TIME=$(date +%s)