[Dynamo] added warning message for tracing lru_cache wrapped functions #153744

.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

@@ -3,8 +3,10 @@ set -eux -o pipefail
 
 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
-if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="9.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="9.0;10.0;12.0"
 fi
 
 SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
@@ -25,7 +27,6 @@ if [ "$DESIRED_CUDA" = "cpu" ]; then
     USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
-    export USE_SYSTEM_NCCL=1
     #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
     USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -31,47 +31,33 @@ def build_ArmComputeLibrary() -> None:
         "build=native",
     ]
     acl_install_dir = "/acl"
-    acl_checkout_dir = os.getenv("ACL_SOURCE_DIR", "ComputeLibrary")
-    if os.path.isdir(acl_install_dir):
-        shutil.rmtree(acl_install_dir)
-    if not os.path.isdir(acl_checkout_dir) or not len(os.listdir(acl_checkout_dir)):
-        check_call(
-            [
-                "git",
-                "clone",
-                "https://github.com/ARM-software/ComputeLibrary.git",
-                "-b",
-                "v25.02",
-                "--depth",
-                "1",
-                "--shallow-submodules",
-            ]
-        )
+    acl_checkout_dir = "ComputeLibrary"
+    os.makedirs(acl_install_dir)
+    check_call(
+        [
+            "git",
+            "clone",
+            "https://github.com/ARM-software/ComputeLibrary.git",
+            "-b",
+            "v25.02",
+            "--depth",
+            "1",
+            "--shallow-submodules",
+        ]
+    )
 
     check_call(
-        ["scons", "Werror=1", f"-j{os.cpu_count()}"] + acl_build_flags,
+        ["scons", "Werror=1", "-j8", f"build_dir=/{acl_install_dir}/build"]
+        + acl_build_flags,
         cwd=acl_checkout_dir,
     )
-    for d in ["arm_compute", "include", "utils", "support", "src", "build"]:
+    for d in ["arm_compute", "include", "utils", "support", "src"]:
         shutil.copytree(f"{acl_checkout_dir}/{d}", f"{acl_install_dir}/{d}")
 
 
-def replace_tag(filename) -> None:
-    with open(filename) as f:
-        lines = f.readlines()
-    for i, line in enumerate(lines):
-        if line.startswith("Tag:"):
-            lines[i] = line.replace("-linux_", "-manylinux_2_28_")
-            print(f"Updated tag from {line} to {lines[i]}")
-            break
-
-    with open(filename, "w") as f:
-        f.writelines(lines)
-
-
-def package_cuda_wheel(wheel_path, desired_cuda) -> None:
+def update_wheel(wheel_path, desired_cuda) -> None:
     """
-    Package the cuda wheel libraries
+    Update the cuda wheel libraries
     """
     folder = os.path.dirname(wheel_path)
     wheelname = os.path.basename(wheel_path)
@@ -79,7 +65,6 @@ 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 +74,7 @@ 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/libnvToolsExt.so.1",
         "/usr/local/cuda/lib64/libnvJitLink.so.12",
         "/usr/local/cuda/lib64/libnvrtc.so.12",
         "/usr/local/cuda/lib64/libcudnn_adv.so.9",
@@ -103,19 +88,30 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
         "/usr/lib64/libgfortran.so.5",
         "/acl/build/libarm_compute.so",
         "/acl/build/libarm_compute_graph.so",
-        "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_lapack_core.so.0",
-        "/usr/local/lib/libnvpl_blas_core.so.0",
     ]
-
-    if "129" in desired_cuda:
+    if enable_cuda:
         libs_to_copy += [
-            "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.9",
-            "/usr/local/cuda/lib64/libcufile.so.0",
-            "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_lapack_core.so.0",
+            "/usr/local/lib/libnvpl_blas_core.so.0",
+        ]
+        if "126" in desired_cuda:
+            libs_to_copy += [
+                "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.6",
+                "/usr/local/cuda/lib64/libcufile.so.0",
+                "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            ]
+        elif "128" in desired_cuda:
+            libs_to_copy += [
+                "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.8",
+                "/usr/local/cuda/lib64/libcufile.so.0",
+                "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            ]
+    else:
+        libs_to_copy += [
+            "/opt/OpenBLAS/lib/libopenblas.so.0",
         ]
-
     # Copy libraries to unzipped_folder/a/lib
     for lib_path in libs_to_copy:
         lib_name = os.path.basename(lib_path)
@@ -124,13 +120,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
             f"cd {folder}/tmp/torch/lib/; "
             f"patchelf --set-rpath '$ORIGIN' --force-rpath {folder}/tmp/torch/lib/{lib_name}"
         )
-
-    # Make sure the wheel is tagged with manylinux_2_28
-    for f in os.scandir(f"{folder}/tmp/"):
-        if f.is_dir() and f.name.endswith(".dist-info"):
-            replace_tag(f"{f.path}/WHEEL")
-            break
-
     os.mkdir(f"{folder}/cuda_wheel")
     os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
     shutil.move(
@@ -205,10 +194,8 @@ if __name__ == "__main__":
     ).decode()
 
     print("Building PyTorch wheel")
-    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
-    # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
-    if enable_cuda:
-        build_vars = "MAX_JOBS=5 " + build_vars
+    build_vars = "MAX_JOBS=5 CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
+    os.system("cd /pytorch; python setup.py clean")
 
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")
@@ -255,6 +242,6 @@ if __name__ == "__main__":
         print("Updating Cuda Dependency")
         filename = os.listdir("/pytorch/dist/")
         wheel_path = f"/pytorch/dist/{filename[0]}"
-        package_cuda_wheel(wheel_path, desired_cuda)
+        update_wheel(wheel_path, desired_cuda)
     pytorch_wheel_name = complete_wheel("/pytorch/")
     print(f"Build Complete. Created {pytorch_wheel_name}..")

.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/almalinux/Dockerfile

@@ -1,7 +1,7 @@
-ARG CUDA_VERSION=12.6
+ARG CUDA_VERSION=12.4
 ARG BASE_TARGET=cuda${CUDA_VERSION}
 ARG ROCM_IMAGE=rocm/dev-almalinux-8:6.3-complete
-FROM amd64/almalinux:8.10-20250519 as base
+FROM amd64/almalinux:8 as base
 
 ENV LC_ALL en_US.UTF-8
 ENV LANG en_US.UTF-8
@@ -11,8 +11,6 @@ ARG DEVTOOLSET_VERSION=11
 
 RUN yum -y update
 RUN yum -y install epel-release
-# install glibc-langpack-en make sure en_US.UTF-8 locale is available
-RUN yum -y install glibc-langpack-en
 RUN yum install -y sudo wget curl perl util-linux xz bzip2 git patch which perl zlib-devel openssl-devel yum-utils autoconf automake make gcc-toolset-${DEVTOOLSET_VERSION}-toolchain
 # Just add everything as a safe.directory for git since these will be used in multiple places with git
 RUN git config --global --add safe.directory '*'
@@ -52,6 +50,10 @@ ENV CUDA_VERSION=${CUDA_VERSION}
 # Make things in our path by default
 ENV PATH=/usr/local/cuda-${CUDA_VERSION}/bin:$PATH
 
+FROM cuda as cuda11.8
+RUN bash ./install_cuda.sh 11.8
+ENV DESIRED_CUDA=11.8
+
 FROM cuda as cuda12.6
 RUN bash ./install_cuda.sh 12.6
 ENV DESIRED_CUDA=12.6
@@ -60,10 +62,6 @@ FROM cuda as cuda12.8
 RUN bash ./install_cuda.sh 12.8
 ENV DESIRED_CUDA=12.8
 
-FROM cuda as cuda12.9
-RUN bash ./install_cuda.sh 12.9
-ENV DESIRED_CUDA=12.9
-
 FROM ${ROCM_IMAGE} as rocm
 ENV PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
 ADD ./common/install_mkl.sh install_mkl.sh
@@ -78,8 +76,7 @@ RUN bash ./install_mnist.sh
 FROM base as all_cuda
 COPY --from=cuda11.8  /usr/local/cuda-11.8 /usr/local/cuda-11.8
 COPY --from=cuda12.6  /usr/local/cuda-12.6 /usr/local/cuda-12.6
-COPY --from=cuda12.8  /usr/local/cuda-12.8 /usr/local/cuda-12.8
-COPY --from=cuda12.9  /usr/local/cuda-12.9 /usr/local/cuda-12.9
+COPY --from=cuda12.4  /usr/local/cuda-12.8 /usr/local/cuda-12.8
 
 # Final step
 FROM ${BASE_TARGET} as final

.ci/docker/build.sh

@@ -50,23 +50,30 @@ if [[ "$image" == *xla* ]]; then
   exit 0
 fi
 
-if [[ "$image" == *-jammy* ]]; then
+if [[ "$image" == *-focal* ]]; then
+  UBUNTU_VERSION=20.04
+elif [[ "$image" == *-jammy* ]]; then
   UBUNTU_VERSION=22.04
-elif [[ "$image" == *-noble* ]]; then
-  UBUNTU_VERSION=24.04
 elif [[ "$image" == *ubuntu* ]]; then
   extract_version_from_image_name ubuntu UBUNTU_VERSION
+elif [[ "$image" == *centos* ]]; then
+  extract_version_from_image_name centos CENTOS_VERSION
 fi
 
 if [ -n "${UBUNTU_VERSION}" ]; then
   OS="ubuntu"
+elif [ -n "${CENTOS_VERSION}" ]; then
+  OS="centos"
 else
   echo "Unable to derive operating system base..."
   exit 1
 fi
 
 DOCKERFILE="${OS}/Dockerfile"
-if [[ "$image" == *rocm* ]]; then
+# When using ubuntu - 22.04, start from Ubuntu docker image, instead of nvidia/cuda docker image.
+if [[ "$image" == *cuda* && "$UBUNTU_VERSION" != "22.04" ]]; then
+  DOCKERFILE="${OS}-cuda/Dockerfile"
+elif [[ "$image" == *rocm* ]]; then
   DOCKERFILE="${OS}-rocm/Dockerfile"
 elif [[ "$image" == *xpu* ]]; then
   DOCKERFILE="${OS}-xpu/Dockerfile"
@@ -91,72 +98,10 @@ tag=$(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
-    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-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=9
-    VISION=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    TRITON=yes
-    INDUCTOR_BENCHMARKS=yes
-    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
-    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-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
-    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-cuda12.6-cudnn9-py3-gcc9)
+  pytorch-linux-focal-cuda12.6-cudnn9-py3-gcc11)
     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-cuda12.8-cudnn9-py3.12-gcc11-vllm)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -164,8 +109,9 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.6
+  pytorch-linux-focal-cuda12.4-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.4.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -175,8 +121,9 @@ case "$tag" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.6
+  pytorch-linux-focal-cuda12.4-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.4.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=9
     VISION=yes
@@ -186,8 +133,9 @@ case "$tag" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.6
+  pytorch-linux-focal-cuda12.4-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.4.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.13
     GCC_VERSION=9
     VISION=yes
@@ -197,8 +145,9 @@ case "$tag" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9)
-    CUDA_VERSION=12.8.1
+  pytorch-linux-focal-cuda12.6-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -207,36 +156,91 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3-clang12-onnx)
+  pytorch-linux-focal-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-focal-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-focal-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.13
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-focal-cuda11.8-cudnn9-py3-gcc9)
+    CUDA_VERSION=11.8.0
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    ;;
+  pytorch-linux-focal-py3-clang10-onnx)
     ANACONDA_PYTHON_VERSION=3.9
-    CLANG_VERSION=12
+    CLANG_VERSION=10
     VISION=yes
     ONNX=yes
     ;;
-  pytorch-linux-jammy-py3.9-clang12)
+  pytorch-linux-focal-py3.9-clang10)
     ANACONDA_PYTHON_VERSION=3.9
-    CLANG_VERSION=12
+    CLANG_VERSION=10
     VISION=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3.11-clang12)
+  pytorch-linux-focal-py3.11-clang10)
     ANACONDA_PYTHON_VERSION=3.11
-    CLANG_VERSION=12
+    CLANG_VERSION=10
     VISION=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3.9-gcc9)
+  pytorch-linux-focal-py3.9-gcc9)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=9
     VISION=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-noble-rocm-n-py3)
-    if [[ $tag =~ "jammy" ]]; then
-      ANACONDA_PYTHON_VERSION=3.10
-    else
-      ANACONDA_PYTHON_VERSION=3.12
-    fi
+  pytorch-linux-jammy-rocm-n-1-py3)
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=11
+    VISION=yes
+    ROCM_VERSION=6.3
+    NINJA_VERSION=1.9.0
+    TRITON=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-rocm-n-py3)
+    ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     ROCM_VERSION=6.4
@@ -247,19 +251,6 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-noble-rocm-alpha-py3)
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=11
-    VISION=yes
-    ROCM_VERSION=7.0
-    NINJA_VERSION=1.9.0
-    TRITON=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    INDUCTOR_BENCHMARKS=yes
-    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
-    ;;
   pytorch-linux-jammy-xpu-2025.0-py3)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
@@ -276,7 +267,7 @@ case "$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
@@ -285,13 +276,25 @@ case "$tag" in
     DOCS=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
+  pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-clang12)
     ANACONDA_PYTHON_VERSION=3.9
-    CUDA_VERSION=12.8.1
+    CUDA_VERSION=11.8
+    CUDNN_VERSION=9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
     ;;
+  pytorch-linux-jammy-py3-clang12-asan)
+    ANACONDA_PYTHON_VERSION=3.9
+    CLANG_VERSION=12
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3-clang15-asan)
+    ANACONDA_PYTHON_VERSION=3.10
+    CLANG_VERSION=15
+    VISION=yes
+    ;;
   pytorch-linux-jammy-py3-clang18-asan)
     ANACONDA_PYTHON_VERSION=3.10
     CLANG_VERSION=18
@@ -324,23 +327,21 @@ case "$tag" in
     GCC_VERSION=11
     TRITON_CPU=yes
     ;;
-  pytorch-linux-jammy-linter)
+  pytorch-linux-focal-linter)
     # TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
     # We will need to update mypy version eventually, but that's for another day. The task
     # would be to upgrade mypy to 1.0.0 with Python 3.11
     PYTHON_VERSION=3.9
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
+  pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-linter)
     PYTHON_VERSION=3.9
-    CUDA_VERSION=12.8.1
+    CUDA_VERSION=11.8
     ;;
   pytorch-linux-jammy-aarch64-py3.10-gcc11)
     ANACONDA_PYTHON_VERSION=3.10
     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
@@ -350,8 +351,6 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
-    CONDA_CMAKE=yes
-    OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific
     SKIP_LLVM_SRC_BUILD_INSTALL=yes
@@ -366,6 +365,7 @@ case "$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
@@ -394,6 +394,14 @@ esac
 
 tmp_tag=$(basename "$(mktemp -u)" | tr '[:upper:]' '[:lower:]')
 
+#when using cudnn version 8 install it separately from cuda
+if [[ "$image" == *cuda*  && ${OS} == "ubuntu" ]]; then
+  IMAGE_NAME="nvidia/cuda:${CUDA_VERSION}-cudnn${CUDNN_VERSION}-devel-ubuntu${UBUNTU_VERSION}"
+  if [[ ${CUDNN_VERSION} == 9 ]]; then
+    IMAGE_NAME="nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}"
+  fi
+fi
+
 no_cache_flag=""
 progress_flag=""
 # Do not use cache and progress=plain when in CI
@@ -410,6 +418,7 @@ docker build \
        --build-arg "LLVMDEV=${LLVMDEV:-}" \
        --build-arg "VISION=${VISION:-}" \
        --build-arg "UBUNTU_VERSION=${UBUNTU_VERSION}" \
+       --build-arg "CENTOS_VERSION=${CENTOS_VERSION}" \
        --build-arg "DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}" \
        --build-arg "GLIBC_VERSION=${GLIBC_VERSION}" \
        --build-arg "CLANG_VERSION=${CLANG_VERSION}" \
@@ -417,6 +426,9 @@ 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=${ROCM_VERSION:-}" \
@@ -434,7 +446,6 @@ docker build \
        --build-arg "XPU_VERSION=${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/centos-rocm/Dockerfile

@@ -39,7 +39,6 @@ RUN bash ./install_user.sh && rm install_user.sh
 
 # Install conda and other packages (e.g., numpy, pytest)
 ARG ANACONDA_PYTHON_VERSION
-ARG BUILD_ENVIRONMENT
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
 COPY requirements-ci.txt /opt/conda/requirements-ci.txt

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-56392aa978594cc155fa8af48cd949f5b5f1823a
+b173722085b3f555d6ba4533d6bbaddfd7c71144

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

@@ -1 +1 @@
-v2.27.5-1
+v2.26.5-1

.ci/docker/ci_commit_pins/triton-xpu.txt

@@ -1 +1 @@
-ae324eeac8e102a2b40370e341460f3791353398
+0bcc8265e677e5321606a3311bf71470f14456a8

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-11ec6354315768a85da41032535e3b7b99c5f706
+96316ce50fade7e209553aba4898cd9b82aab83b

.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"
@@ -33,6 +30,18 @@ install_ubuntu() {
     maybe_libomp_dev=""
   fi
 
+  # HACK: UCC testing relies on libnccl library from NVIDIA repo, and version 2.16 crashes
+  # See https://github.com/pytorch/pytorch/pull/105260#issuecomment-1673399729
+  # TODO: Eliminate this hack, we should not relay on apt-get installation
+  # See https://github.com/pytorch/pytorch/issues/144768
+  if [[ "$UBUNTU_VERSION" == "20.04"* && "$CUDA_VERSION" == "11.8"* ]]; then
+    maybe_libnccl_dev="libnccl2=2.15.5-1+cuda11.8 libnccl-dev=2.15.5-1+cuda11.8 --allow-downgrades --allow-change-held-packages"
+  elif [[ "$UBUNTU_VERSION" == "20.04"* && "$CUDA_VERSION" == "12.4"* ]]; then
+    maybe_libnccl_dev="libnccl2=2.26.2-1+cuda12.4 libnccl-dev=2.26.2-1+cuda12.4 --allow-downgrades --allow-change-held-packages"
+  else
+    maybe_libnccl_dev=""
+  fi
+
   # Install common dependencies
   apt-get update
   # TODO: Some of these may not be necessary
@@ -61,6 +70,7 @@ install_ubuntu() {
     libasound2-dev \
     libsndfile-dev \
     ${maybe_libomp_dev} \
+    ${maybe_libnccl_dev} \
     software-properties-common \
     wget \
     sudo \

.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* ]]; 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
 
@@ -59,16 +64,11 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
   # which is provided in libstdcxx 12 and up.
   conda_install libstdcxx-ng=12.3.0 --update-deps -c conda-forge
 
-  # Miniforge installer doesn't install sqlite by default
-  if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
-    conda_install sqlite
-  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 +82,6 @@ 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
@@ -66,7 +68,7 @@ function do_cpython_build {
         ln -s pip3 ${prefix}/bin/pip
     fi
     # install setuptools since python 3.12 is required to use distutils
-    ${prefix}/bin/pip install wheel==0.45.1 setuptools==80.9.0
+    ${prefix}/bin/pip install wheel==0.34.2 setuptools==68.2.2
     local abi_tag=$(${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}
 }
@@ -74,20 +76,24 @@ function do_cpython_build {
 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
@@ -42,40 +40,18 @@ 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"
+function install_118 {
+    CUDNN_VERSION=9.1.0.70
+    echo "Installing CUDA 11.8 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.4.0"
+    install_cuda 11.8.0 cuda_11.8.0_520.61.05_linux
 
-  case "${arch_path}" in
-    sbsa)
-      dl_arch="aarch64"
-      ;;
-    x86_64)
-      dl_arch="x64"
-      ;;
-    *)
-      dl_arch="${arch}"
-      ;;
-  esac
+    install_cudnn 11 $CUDNN_VERSION
 
-  tmpdir="tmp_nvshmem"
-  mkdir -p "${tmpdir}" && cd "${tmpdir}"
+    CUDA_VERSION=11.8 bash install_nccl.sh
 
-  # 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"
+    CUDA_VERSION=11.8 bash install_cusparselt.sh
 
-  # download, unpack, install
-  wget -q "${url}"
-  tar xf "${filename}.tar.gz"
-  cp -a "libnvshmem/include/"* /usr/local/include/
-  cp -a "libnvshmem/lib/"*     /usr/local/lib/
-
-  # cleanup
-  cd ..
-  rm -rf "${tmpdir}"
-
-  echo "nvSHMEM ${nvshmem_version} for CUDA ${cuda_major_version} (${arch_path}) installed."
+    ldconfig
 }
 
 function install_124 {
@@ -93,14 +69,12 @@ function install_124 {
 }
 
 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"
+  CUDNN_VERSION=9.5.1.17
+  echo "Installing CUDA 12.6.3 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.6.3"
   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
@@ -108,22 +82,35 @@ function install_126 {
   ldconfig
 }
 
-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"
-  # install CUDA 12.9.1 in the same container
-  install_cuda 12.9.1 cuda_12.9.1_575.57.08_linux
+function prune_118 {
+    echo "Pruning CUDA 11.8 and cuDNN"
+    #####################################################################################
+    # CUDA 11.8 prune static libs
+    #####################################################################################
+    export NVPRUNE="/usr/local/cuda-11.8/bin/nvprune"
+    export CUDA_LIB_DIR="/usr/local/cuda-11.8/lib64"
 
-  # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
-  install_cudnn 12 $CUDNN_VERSION
+    export GENCODE="-gencode arch=compute_35,code=sm_35 -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_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -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"
 
-  install_nvshmem 12 $NVSHMEM_VERSION
+    if [[ -n "$OVERRIDE_GENCODE" ]]; then
+        export GENCODE=$OVERRIDE_GENCODE
+    fi
 
-  CUDA_VERSION=12.9 bash install_nccl.sh
+    # all CUDA libs except CuDNN and CuBLAS (cudnn and cublas need arch 3.7 included)
+    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/{}"
 
-  CUDA_VERSION=12.9 bash install_cusparselt.sh
+    # 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
 
-  ldconfig
+    #####################################################################################
+    # CUDA 11.8 prune visual tools
+    #####################################################################################
+    export CUDA_BASE="/usr/local/cuda-11.8/"
+    rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2022.3.0 $CUDA_BASE/nsight-systems-2022.4.2/
 }
 
 function prune_124 {
@@ -196,15 +183,13 @@ 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"
-  # install CUDA 12.8.1 in the same container
-  install_cuda 12.8.1 cuda_12.8.1_570.124.06_linux
+  echo "Installing CUDA 12.8.0 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.6.3"
+  # install CUDA 12.8.0 in the same container
+  install_cuda 12.8.0 cuda_12.8.0_570.86.10_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
@@ -216,13 +201,13 @@ function install_128 {
 while test $# -gt 0
 do
     case "$1" in
+    11.8) install_118; prune_118
+        ;;
     12.4) install_124; prune_124
         ;;
-    12.6|12.6.*) install_126; prune_126
+    12.6) install_126; prune_126
         ;;
-    12.8|12.8.*) install_128;
-        ;;
-    12.9|12.9.*) install_129;
+    12.8) install_128;
         ;;
     *) echo "bad argument $1"; exit 1
         ;;

.ci/docker/common/install_cudnn.sh

@@ -0,0 +1,26 @@
+#!/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} == "12.8" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.8.0.87_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:4} == "12.6" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.5.1.17_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:2} == "12" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.1.0.70_cuda12-archive"
+    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

@@ -5,13 +5,13 @@ set -ex
 # cuSPARSELt license: https://docs.nvidia.com/cuda/cusparselt/license.html
 mkdir tmp_cusparselt && cd tmp_cusparselt
 
-if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
+if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-8]$ ]]; then
     arch_path='sbsa'
     export TARGETARCH=${TARGETARCH:-$(uname -m)}
     if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
         arch_path='x86_64'
     fi
-    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.7.1.0-archive"
+    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.6.3.2-archive"
     curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
 elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
     arch_path='sbsa'
@@ -21,6 +21,9 @@ elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
     fi
     CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.6.2.3-archive"
     curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
+elif [[ ${CUDA_VERSION:0:4} == "11.8" ]]; then
+    CUSPARSELT_NAME="libcusparse_lt-linux-x86_64-0.4.0.7-archive"
+    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-x86_64/${CUSPARSELT_NAME}.tar.xz
 else
     echo "Not sure which libcusparselt version to install for this ${CUDA_VERSION}"
 fi

.ci/docker/common/install_onnx.sh

@@ -8,6 +8,16 @@ retry () {
     "$@" || (sleep 10 && "$@") || (sleep 20 && "$@") || (sleep 40 && "$@")
 }
 
+# A bunch of custom pip dependencies for ONNX
+pip_install \
+  beartype==0.15.0 \
+  filelock==3.9.0 \
+  flatbuffers==2.0 \
+  mock==5.0.1 \
+  ninja==1.10.2 \
+  networkx==2.5 \
+  numpy==1.24.2
+
 # ONNXRuntime should be installed before installing
 # onnx-weekly. Otherwise, onnx-weekly could be
 # overwritten by onnx.
@@ -19,8 +29,12 @@ pip_install \
   transformers==4.36.2
 
 pip_install coloredlogs packaging
+
 pip_install onnxruntime==1.18.1
-pip_install onnxscript==0.3.1
+pip_install onnx==1.17.0
+pip_install onnxscript==0.2.2 --no-deps
+# required by onnxscript
+pip_install ml_dtypes
 
 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/

.ci/docker/common/install_openblas.sh

@@ -4,9 +4,9 @@
 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 v0.3.29 --depth 1 --shallow-submodules
+
 
-OPENBLAS_CHECKOUT_DIR="OpenBLAS"
 OPENBLAS_BUILD_FLAGS="
 NUM_THREADS=128
 USE_OPENMP=1
@@ -14,8 +14,9 @@ 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
@@ -28,27 +26,13 @@ Pin: release o=repo.radeon.com
 Pin-Priority: 600
 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"
-    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 +66,25 @@ 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) -eq $(ver 6.3) ]] || [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
+        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.3) ]]; then
+            HIP_BRANCH=rocm-6.3.x
+            VER_STR=6.3
         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
         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}-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

@@ -51,12 +51,7 @@ as_jenkins git clone --recursive ${TRITON_REPO} triton
 cd triton
 as_jenkins git checkout ${TRITON_PINNED_COMMIT}
 as_jenkins git submodule update --init --recursive
-
-# Old versions of python have setup.py in ./python; newer versions have it in ./
-if [ ! -f setup.py ]; then
-  cd python
-fi
-
+cd python
 pip_install pybind11==2.13.6
 
 # TODO: remove patch setup.py once we have a proper fix for https://github.com/triton-lang/triton/issues/4527
@@ -98,10 +93,3 @@ 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==0.0.10 --no-deps
-fi

.ci/docker/common/install_xpu.sh

@@ -56,10 +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)'

.ci/docker/libtorch/Dockerfile

@@ -54,6 +54,16 @@ COPY ./ci_commit_pins/nccl-cu* /ci_commit_pins/
 COPY ./common/install_cusparselt.sh install_cusparselt.sh
 ENV CUDA_HOME /usr/local/cuda
 
+FROM cuda as cuda11.8
+RUN bash ./install_cuda.sh 11.8
+RUN bash ./install_magma.sh 11.8
+RUN ln -sf /usr/local/cuda-11.8 /usr/local/cuda
+
+FROM cuda as cuda12.4
+RUN bash ./install_cuda.sh 12.4
+RUN bash ./install_magma.sh 12.4
+RUN ln -sf /usr/local/cuda-12.4 /usr/local/cuda
+
 FROM cuda as cuda12.6
 RUN bash ./install_cuda.sh 12.6
 RUN bash ./install_magma.sh 12.6
@@ -64,11 +74,6 @@ RUN bash ./install_cuda.sh 12.8
 RUN bash ./install_magma.sh 12.8
 RUN ln -sf /usr/local/cuda-12.8 /usr/local/cuda
 
-FROM cuda as cuda12.9
-RUN bash ./install_cuda.sh 12.9
-RUN bash ./install_magma.sh 12.9
-RUN ln -sf /usr/local/cuda-12.9 /usr/local/cuda
-
 FROM cpu as rocm
 ARG ROCM_VERSION
 ARG PYTORCH_ROCM_ARCH

.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_2_28

@@ -26,7 +26,7 @@ ADD ./common/install_openssl.sh install_openssl.sh
 RUN bash ./install_openssl.sh && rm install_openssl.sh
 
 
-# remove unnecessary python versions
+# remove unncessary python versions
 RUN rm -rf /opt/python/cp26-cp26m /opt/_internal/cpython-2.6.9-ucs2
 RUN rm -rf /opt/python/cp26-cp26mu /opt/_internal/cpython-2.6.9-ucs4
 RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6
@@ -103,7 +103,6 @@ ENV SSL_CERT_FILE=/opt/_internal/certs.pem
 # Install LLVM version
 COPY --from=openssl            /opt/openssl                          /opt/openssl
 COPY --from=base               /opt/python                           /opt/python
-COPY --from=base               /usr/local/lib/                       /usr/local/lib/
 COPY --from=base               /opt/_internal                        /opt/_internal
 COPY --from=base               /usr/local/bin/auditwheel             /usr/local/bin/auditwheel
 COPY --from=intel              /opt/intel                            /opt/intel

.ci/docker/manywheel/Dockerfile_2_28_aarch64

@@ -2,7 +2,7 @@ FROM quay.io/pypa/manylinux_2_28_aarch64 as base
 
 ARG GCCTOOLSET_VERSION=13
 
-# Language variables
+# Language variabes
 ENV LC_ALL=en_US.UTF-8
 ENV LANG=en_US.UTF-8
 ENV LANGUAGE=en_US.UTF-8
@@ -58,13 +58,12 @@ RUN git config --global --add safe.directory "*"
 
 FROM base as openblas
 # Install openblas
-ARG OPENBLAS_VERSION
 ADD ./common/install_openblas.sh install_openblas.sh
 RUN bash ./install_openblas.sh && rm install_openblas.sh
 
 FROM base as final
 
-# remove unnecessary python versions
+# remove unncessary python versions
 RUN rm -rf /opt/python/cp26-cp26m /opt/_internal/cpython-2.6.9-ucs2
 RUN rm -rf /opt/python/cp26-cp26mu /opt/_internal/cpython-2.6.9-ucs4
 RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6

.ci/docker/manywheel/Dockerfile_cuda_aarch64

@@ -60,7 +60,7 @@ RUN bash ./install_openssl.sh && rm install_openssl.sh
 ENV SSL_CERT_FILE=/opt/_internal/certs.pem
 
 FROM openssl as final
-# remove unnecessary python versions
+# remove unncessary python versions
 RUN rm -rf /opt/python/cp26-cp26m /opt/_internal/cpython-2.6.9-ucs2
 RUN rm -rf /opt/python/cp26-cp26mu /opt/_internal/cpython-2.6.9-ucs4
 RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6

.ci/docker/manywheel/Dockerfile_s390x

@@ -5,9 +5,7 @@ ENV LC_ALL=C.UTF-8
 ENV LANG=C.UTF-8
 ENV LANGUAGE=C.UTF-8
 
-# there is a bugfix in gcc >= 14 for precompiled headers and s390x vectorization interaction.
-# with earlier gcc versions test/inductor/test_cpu_cpp_wrapper.py will fail.
-ARG DEVTOOLSET_VERSION=14
+ARG DEVTOOLSET_VERSION=13
 # Installed needed OS packages. This is to support all
 # the binary builds (torch, vision, audio, text, data)
 RUN yum -y install epel-release
@@ -60,8 +58,7 @@ RUN yum install -y \
   libxslt-devel \
   libxml2-devel \
   openssl-devel \
-  valgrind \
-  ninja-build
+  valgrind
 
 ENV PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/bin:$PATH
 ENV LD_LIBRARY_PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib64:/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib:$LD_LIBRARY_PATH
@@ -106,6 +103,9 @@ CMD ["/bin/bash"]
 # install test dependencies:
 # - grpcio requires system openssl, bundled crypto fails to build
 RUN dnf install -y \
+  protobuf-devel \
+  protobuf-c-devel \
+  protobuf-lite-devel \
   hdf5-devel \
   python3-h5py \
   git
@@ -120,22 +120,15 @@ RUN python3 -mpip install cmake==3.28.0
 # so just build it from upstream repository.
 # h5py is dependency of onnxruntime_training.
 # h5py==3.11.0 builds with hdf5-devel 1.10.5 from repository.
-# h5py 3.11.0 doesn't build with numpy >= 2.3.0.
 # install newest flatbuffers version first:
 # for some reason old version is getting pulled in otherwise.
 # packaging package is required for onnxruntime wheel build.
 RUN pip3 install flatbuffers && \
-  pip3 install cython 'pkgconfig>=1.5.5' 'setuptools>=77' 'numpy<2.3.0' && \
-  pip3 install --no-build-isolation h5py==3.11.0 && \
+  pip3 install h5py==3.11.0 && \
   pip3 install packaging && \
   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 \
-  --compile_no_warning_as_error && \
+  ./build.sh --config Release --parallel 0 --enable_pybind --build_wheel --enable_training --enable_training_apis --enable_training_ops --skip_tests --allow_running_as_root && \
   pip3 install ./build/Linux/Release/dist/onnxruntime_training-*.whl && \
   cd .. && /bin/rm -rf ./onnxruntime

.ci/docker/manywheel/build.sh

@@ -27,7 +27,6 @@ fi
 
 MANY_LINUX_VERSION=${MANY_LINUX_VERSION:-}
 DOCKERFILE_SUFFIX=${DOCKERFILE_SUFFIX:-}
-OPENBLAS_VERSION=${OPENBLAS_VERSION:-}
 
 case ${image} in
     manylinux2_28-builder:cpu)
@@ -41,7 +40,6 @@ 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"
         ;;
     manylinuxcxx11-abi-builder:cpu-cxx11-abi)
         TARGET=final
@@ -75,10 +73,6 @@ 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"
@@ -115,7 +109,6 @@ tmp_tag=$(basename "$(mktemp -u)" | tr '[:upper:]' '[:lower:]')
 DOCKER_BUILDKIT=1 docker build  \
     ${DOCKER_GPU_BUILD_ARG} \
     --build-arg "GPU_IMAGE=${GPU_IMAGE}" \
-    --build-arg "OPENBLAS_VERSION=${OPENBLAS_VERSION}" \
     --target "${TARGET}" \
     -t "${tmp_tag}" \
     $@ \

.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:
@@ -42,15 +41,18 @@ fbscribelogger==0.1.7
 #Pinned versions: 0.1.6
 #test that import:
 
-flatbuffers==24.12.23
+flatbuffers==2.0 ; platform_machine != "s390x"
 #Description: cross platform serialization library
-#Pinned versions: 24.12.23
+#Pinned versions: 2.0
 #test that import:
 
+flatbuffers ; platform_machine == "s390x"
+#Description: cross platform serialization library; Newer version is required on s390x for new python version
+
 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
@@ -64,7 +66,6 @@ lark==0.12.0
 #test that import:
 
 librosa>=0.6.2 ; python_version < "3.11"
-librosa==0.10.2 ; python_version == "3.12"
 #Description: A python package for music and audio analysis
 #Pinned versions: >=0.6.2
 #test that import: test_spectral_ops.py
@@ -92,10 +93,10 @@ librosa==0.10.2 ; python_version == "3.12"
 #Pinned versions:
 #test that import:
 
-mypy==1.16.0
+mypy==1.14.0
 # Pin MyPy version because new errors are likely to appear with each release
 #Description: linter
-#Pinned versions: 1.16.0
+#Pinned versions: 1.14.0
 #test that import: test_typing.py, test_type_hints.py
 
 networkx==2.8.8
@@ -113,7 +114,6 @@ ninja==1.11.1.3
 numba==0.49.0 ; python_version < "3.9"
 numba==0.55.2 ; python_version == "3.9"
 numba==0.55.2 ; python_version == "3.10"
-numba==0.60.0 ; python_version == "3.12"
 #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
@@ -166,10 +166,10 @@ pillow==11.0.0
 #Pinned versions: 10.3.0
 #test that import:
 
-protobuf==5.29.4
-#Description:  Google's data interchange format
-#Pinned versions: 5.29.4
-#test that import: test_tensorboard.py, test/onnx/*
+protobuf==3.20.2
+#Description:  Google’s data interchange format
+#Pinned versions: 3.20.1
+#test that import: test_tensorboard.py
 
 psutil
 #Description: information on running processes and system utilization
@@ -221,9 +221,9 @@ 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
@@ -233,7 +233,7 @@ 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"
@@ -307,7 +307,7 @@ pytest-cpp==2.3.0
 #Pinned versions: 2.3.0
 #test that import:
 
-z3-solver==4.15.1.0
+z3-solver==4.12.6.0
 #Description: The Z3 Theorem Prover Project
 #Pinned versions:
 #test that import:
@@ -337,12 +337,12 @@ sympy==1.13.3
 #Pinned versions:
 #test that import:
 
-onnx==1.18.0
-#Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
+onnx==1.17.0
+#Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
 
-onnxscript==0.3.1
+onnxscript==0.2.2
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -363,10 +363,10 @@ pwlf==2.2.1
 
 
 # To build PyTorch itself
-pyyaml
+astunparse
+PyYAML
 pyzstd
-setuptools>=70.1.0
-six
+setuptools
 
 scons==4.5.2 ; platform_machine == "aarch64"
 
@@ -382,16 +382,3 @@ dataclasses_json==0.6.7
 
 cmake==4.0.0
 #Description: required for building
-
-tlparse==0.3.30
-#Description: required for log parsing
-
-cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"
-#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
@@ -15,14 +15,9 @@ sphinxext-opengraph==0.9.1
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 0.9.1
 
-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 +45,8 @@ 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
+3.3.0

.ci/docker/triton_xpu_version.txt

@@ -1 +0,0 @@
-3.4.0

.ci/docker/ubuntu-cuda/Dockerfile

@@ -0,0 +1,170 @@
+ARG UBUNTU_VERSION
+ARG CUDA_VERSION
+ARG IMAGE_NAME
+
+FROM ${IMAGE_NAME} as base
+
+ARG UBUNTU_VERSION
+ARG CUDA_VERSION
+
+ENV DEBIAN_FRONTEND noninteractive
+
+# Install common dependencies (so that this step can be cached separately)
+COPY ./common/install_base.sh install_base.sh
+RUN bash ./install_base.sh && rm install_base.sh
+
+# Install user
+COPY ./common/install_user.sh install_user.sh
+RUN bash ./install_user.sh && rm install_user.sh
+
+# Install katex
+ARG KATEX
+COPY ./common/install_docs_reqs.sh install_docs_reqs.sh
+RUN bash ./install_docs_reqs.sh && rm install_docs_reqs.sh
+
+# Install conda and other packages (e.g., numpy, pytest)
+ARG ANACONDA_PYTHON_VERSION
+ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
+ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
+COPY requirements-ci.txt /opt/conda/requirements-ci.txt
+COPY ./common/install_conda.sh install_conda.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./common/install_magma_conda.sh install_magma_conda.sh
+RUN bash ./install_conda.sh && rm install_conda.sh install_magma_conda.sh common_utils.sh /opt/conda/requirements-ci.txt
+
+# Install gcc
+ARG GCC_VERSION
+COPY ./common/install_gcc.sh install_gcc.sh
+RUN bash ./install_gcc.sh && rm install_gcc.sh
+
+# Install clang
+ARG CLANG_VERSION
+COPY ./common/install_clang.sh install_clang.sh
+RUN bash ./install_clang.sh && rm install_clang.sh
+
+# (optional) Install vision packages like OpenCV
+ARG VISION
+COPY ./common/install_vision.sh ./common/cache_vision_models.sh ./common/common_utils.sh ./
+RUN if [ -n "${VISION}" ]; then bash ./install_vision.sh; fi
+RUN rm install_vision.sh cache_vision_models.sh common_utils.sh
+ENV INSTALLED_VISION ${VISION}
+
+# (optional) Install UCC
+ARG UCX_COMMIT
+ARG UCC_COMMIT
+ENV UCX_COMMIT $UCX_COMMIT
+ENV UCC_COMMIT $UCC_COMMIT
+ENV UCX_HOME /usr
+ENV UCC_HOME /usr
+ADD ./common/install_ucc.sh install_ucc.sh
+RUN if [ -n "${UCX_COMMIT}" ] && [ -n "${UCC_COMMIT}" ]; then bash ./install_ucc.sh; fi
+RUN rm install_ucc.sh
+
+COPY ./common/install_openssl.sh install_openssl.sh
+ENV OPENSSL_ROOT_DIR /opt/openssl
+RUN bash ./install_openssl.sh
+ENV OPENSSL_DIR /opt/openssl
+
+ARG INDUCTOR_BENCHMARKS
+ARG ANACONDA_PYTHON_VERSION
+ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
+COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ci_commit_pins/huggingface.txt huggingface.txt
+COPY ci_commit_pins/timm.txt timm.txt
+RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
+
+ARG TRITON
+
+FROM base as triton-builder
+# Install triton, this needs to be done before sccache because the latter will
+# try to reach out to S3, which docker build runners don't have access
+COPY ./common/install_triton.sh install_triton.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ci_commit_pins/triton.txt triton.txt
+COPY triton_version.txt triton_version.txt
+RUN bash ./install_triton.sh
+
+FROM base as final
+COPY --from=triton-builder /opt/triton /opt/triton
+RUN if [ -n "${TRITON}" ]; then pip install /opt/triton/*.whl; chown -R jenkins:jenkins /opt/conda; fi
+RUN rm -rf /opt/triton
+
+ARG HALIDE
+# Build and install halide
+COPY ./common/install_halide.sh install_halide.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ci_commit_pins/halide.txt halide.txt
+RUN if [ -n "${HALIDE}" ]; then bash ./install_halide.sh; fi
+RUN rm install_halide.sh common_utils.sh halide.txt
+
+# Install ccache/sccache (do this last, so we get priority in PATH)
+COPY ./common/install_cache.sh install_cache.sh
+ENV PATH /opt/cache/bin:$PATH
+# See https://github.com/pytorch/pytorch/issues/82174
+# TODO(sdym@fb.com):
+# check if this is needed after full off Xenial migration
+ENV CARGO_NET_GIT_FETCH_WITH_CLI true
+RUN bash ./install_cache.sh && rm install_cache.sh
+ENV CMAKE_CUDA_COMPILER_LAUNCHER=/opt/cache/bin/sccache
+
+# Add jni.h for java host build
+COPY ./common/install_jni.sh install_jni.sh
+COPY ./java/jni.h jni.h
+RUN bash ./install_jni.sh && rm install_jni.sh
+
+# Install Open MPI for CUDA
+COPY ./common/install_openmpi.sh install_openmpi.sh
+RUN if [ -n "${CUDA_VERSION}" ]; then bash install_openmpi.sh; fi
+RUN rm install_openmpi.sh
+
+# Include BUILD_ENVIRONMENT environment variable in image
+ARG BUILD_ENVIRONMENT
+ENV BUILD_ENVIRONMENT ${BUILD_ENVIRONMENT}
+
+# AWS specific CUDA build guidance
+ENV TORCH_CUDA_ARCH_LIST Maxwell
+ENV TORCH_NVCC_FLAGS "-Xfatbin -compress-all"
+ENV CUDA_PATH /usr/local/cuda
+
+# Install LLVM dev version (Defined in the pytorch/builder github repository)
+COPY --from=pytorch/llvm:9.0.1 /opt/llvm /opt/llvm
+
+# Install CUDNN
+ARG CUDNN_VERSION
+ARG CUDA_VERSION
+COPY ./common/install_cudnn.sh install_cudnn.sh
+RUN if [ -n "${CUDNN_VERSION}" ]; then bash install_cudnn.sh; fi
+RUN rm install_cudnn.sh
+
+# Install CUSPARSELT
+ARG CUDA_VERSION
+COPY ./common/install_cusparselt.sh install_cusparselt.sh
+RUN bash install_cusparselt.sh
+RUN rm install_cusparselt.sh
+
+# Install NCCL
+ARG CUDA_VERSION
+COPY ./common/install_nccl.sh install_nccl.sh
+COPY ./ci_commit_pins/nccl-cu* /ci_commit_pins/
+RUN bash install_nccl.sh
+RUN rm install_nccl.sh /ci_commit_pins/nccl-cu*
+ENV USE_SYSTEM_NCCL=1
+ENV NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+ENV NCCL_LIB_DIR="/usr/local/cuda/lib64/"
+
+# Install CUDSS
+ARG CUDA_VERSION
+COPY ./common/install_cudss.sh install_cudss.sh
+RUN bash install_cudss.sh
+RUN rm install_cudss.sh
+
+# Delete /usr/local/cuda-11.X/cuda-11.X symlinks
+RUN if [ -h /usr/local/cuda-11.6/cuda-11.6 ]; then rm /usr/local/cuda-11.6/cuda-11.6; fi
+RUN if [ -h /usr/local/cuda-11.7/cuda-11.7 ]; then rm /usr/local/cuda-11.7/cuda-11.7; fi
+RUN if [ -h /usr/local/cuda-12.1/cuda-12.1 ]; then rm /usr/local/cuda-12.1/cuda-12.1; fi
+RUN if [ -h /usr/local/cuda-12.4/cuda-12.4 ]; then rm /usr/local/cuda-12.4/cuda-12.4; fi
+
+USER jenkins
+CMD ["bash"]

.ci/docker/ubuntu-rocm/Dockerfile

@@ -25,7 +25,6 @@ RUN bash ./install_docs_reqs.sh && rm install_docs_reqs.sh
 
 # Install conda and other packages (e.g., numpy, pytest)
 ARG ANACONDA_PYTHON_VERSION
-ARG BUILD_ENVIRONMENT
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
 COPY requirements-ci.txt /opt/conda/requirements-ci.txt

.ci/docker/ubuntu-xpu/Dockerfile

@@ -72,7 +72,7 @@ ARG TRITON
 COPY ./common/install_triton.sh install_triton.sh
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/triton-xpu.txt triton-xpu.txt
-COPY triton_xpu_version.txt triton_version.txt
+COPY triton_version.txt triton_version.txt
 RUN if [ -n "${TRITON}" ]; then bash ./install_triton.sh; fi
 RUN rm install_triton.sh common_utils.sh triton-xpu.txt triton_version.txt
 

.ci/docker/ubuntu/Dockerfile

@@ -147,12 +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/magma/Makefile

@@ -1,7 +1,7 @@
 SHELL=/usr/bin/env bash
 
 DOCKER_CMD ?= docker
-DESIRED_CUDA ?= 12.8
+DESIRED_CUDA ?= 11.8
 DESIRED_CUDA_SHORT = $(subst .,,$(DESIRED_CUDA))
 PACKAGE_NAME = magma-cuda
 CUDA_ARCH_LIST ?= -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90
@@ -16,21 +16,15 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma/build_magma.sh
 
 .PHONY: all
-all: magma-cuda129
 all: magma-cuda128
 all: magma-cuda126
+all: magma-cuda118
 
 .PHONY:
 clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
-.PHONY: magma-cuda129
-magma-cuda129: DESIRED_CUDA := 12.9
-magma-cuda129: CUDA_ARCH_LIST += -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
-magma-cuda129:
-	$(DOCKER_RUN)
-
 .PHONY: magma-cuda128
 magma-cuda128: DESIRED_CUDA := 12.8
 magma-cuda128: CUDA_ARCH_LIST += -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
@@ -41,3 +35,9 @@ magma-cuda128:
 magma-cuda126: DESIRED_CUDA := 12.6
 magma-cuda126:
 	$(DOCKER_RUN)
+
+.PHONY: magma-cuda118
+magma-cuda118: DESIRED_CUDA := 11.8
+magma-cuda118: CUDA_ARCH_LIST += -gencode arch=compute_37,code=sm_37
+magma-cuda118:
+	$(DOCKER_RUN)

.ci/manywheel/build_common.sh

@@ -18,10 +18,12 @@ retry () {
     $*  || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
 }
 
-PLATFORM=""
+PLATFORM="manylinux2014_x86_64"
 # TODO move this into the Docker images
 OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
-if [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
     retry yum install -q -y zip openssl
     PLATFORM="manylinux_2_28_x86_64"
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
@@ -31,11 +33,9 @@ elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
     # Comment out nvidia repositories to prevent them from getting apt-get updated, see https://github.com/pytorch/pytorch/issues/74968
     # shellcheck disable=SC2046
     sed -i 's/.*nvidia.*/# &/' $(find /etc/apt/ -type f -name "*.list")
+
     retry apt-get update
     retry apt-get -y install zip openssl
-else
-    echo "Unknown OS: '$OS_NAME'"
-    exit 1
 fi
 
 # We use the package name to test the package by passing this to 'pip install'
@@ -79,6 +79,8 @@ if [[ -e /opt/openssl ]]; then
     export CMAKE_INCLUDE_PATH="/opt/openssl/include":$CMAKE_INCLUDE_PATH
 fi
 
+
+
 mkdir -p /tmp/$WHEELHOUSE_DIR
 
 export PATCHELF_BIN=/usr/local/bin/patchelf
@@ -97,7 +99,6 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
     exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -qUr requirements-build.txt
 python setup.py clean
 retry pip install -qr requirements.txt
 case ${DESIRED_PYTHON} in
@@ -151,7 +152,7 @@ if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
     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/$WHEELHOUSE_DIR
+    python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR --cmake
     echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
 else
     time CMAKE_ARGS=${CMAKE_ARGS[@]} \

.ci/manywheel/build_cuda.sh

@@ -15,9 +15,6 @@ export INSTALL_TEST=0 # dont install test binaries into site-packages
 export USE_CUPTI_SO=0
 export USE_CUSPARSELT=${USE_CUSPARSELT:-1} # Enable if not disabled by libtorch build
 export USE_CUFILE=${USE_CUFILE:-1}
-export USE_SYSTEM_NCCL=1
-export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
-export NCCL_LIB_DIR="/usr/local/cuda/lib64/"
 
 # Keep an array of cmake variables to add to
 if [[ -z "$CMAKE_ARGS" ]]; then
@@ -39,8 +36,10 @@ if [[ -n "$DESIRED_CUDA" ]]; then
     if [[ ${DESIRED_CUDA} =~ ^[0-9]+\.[0-9]+$ ]]; then
         CUDA_VERSION=${DESIRED_CUDA}
     else
-        # cu126, cu128 etc...
-        if [[ ${#DESIRED_CUDA} -eq 5 ]]; then
+        # cu90, cu92, cu100, cu101
+        if [[ ${#DESIRED_CUDA} -eq 4 ]]; then
+            CUDA_VERSION="${DESIRED_CUDA:2:1}.${DESIRED_CUDA:3:1}"
+        elif [[ ${#DESIRED_CUDA} -eq 5 ]]; then
             CUDA_VERSION="${DESIRED_CUDA:2:2}.${DESIRED_CUDA:4:1}"
         fi
     fi
@@ -51,23 +50,24 @@ 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 ${CUDA_VERSION} in
-    #removing sm_50-sm_60 as these architectures are deprecated in CUDA 12.8/9 and will be removed in future releases
-    #however we would like to keep sm_70 architecture see: https://github.com/pytorch/pytorch/issues/157517
     12.8)
-        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0"
-        ;;
-    12.9)
-        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0+PTX"
-        # WAR to resolve the ld error in libtorch build with CUDA 12.9
-        if [[ "$PACKAGE_TYPE" == "libtorch" ]]; then
-            TORCH_CUDA_ARCH_LIST="7.5;8.0;9.0;10.0;12.0+PTX"
-        fi
+        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 and will be removed in future releases
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         ;;
     12.6)
-        TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6;9.0"
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    12.4)
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    11.8)
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};3.7;9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         ;;
     *)
         echo "unknown cuda version $CUDA_VERSION"
@@ -91,15 +91,14 @@ fi
 mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
 
 OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
-if [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
     LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
     LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
 elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
     LIBGOMP_PATH="/usr/lib/x86_64-linux-gnu/libgomp.so.1"
-else
-    echo "Unknown OS: '$OS_NAME'"
-    exit 1
 fi
 
 DEPS_LIST=(
@@ -109,12 +108,31 @@ DEPS_SONAME=(
     "libgomp.so.1"
 )
 
+# CUDA 11.8 have to ship the libcusparseLt.so.0 with the binary
+# since nvidia-cusparselt-cu11 is not available in PYPI
+if [[ $USE_CUSPARSELT == "1" && $CUDA_VERSION == "11.8" ]]; then
+        DEPS_SONAME+=(
+            "libcusparseLt.so.0"
+        )
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcusparseLt.so.0"
+        )
+fi
 
-# CUDA_VERSION 12.6, 12.8, 12.9
+
+# Turn USE_CUFILE off for CUDA 11.8, 12.4 since nvidia-cufile-cu11 and 1.9.0.20 are
+# not available in PYPI
+if [[ $CUDA_VERSION == "11.8" || $CUDA_VERSION == "12.4" ]]; then
+    export USE_CUFILE=0
+fi
+
+
+# CUDA_VERSION 12.4, 12.6, 12.8
 if [[ $CUDA_VERSION == 12* ]]; then
     export USE_STATIC_CUDNN=0
     # Try parallelizing nvcc as well
     export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
+
     if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
         echo "Bundling with cudnn and cublas."
         DEPS_LIST+=(
@@ -130,12 +148,9 @@ if [[ $CUDA_VERSION == 12* ]]; then
             "/usr/local/cuda/lib64/libcublasLt.so.12"
             "/usr/local/cuda/lib64/libcusparseLt.so.0"
             "/usr/local/cuda/lib64/libcudart.so.12"
+            "/usr/local/cuda/lib64/libnvToolsExt.so.1"
             "/usr/local/cuda/lib64/libnvrtc.so.12"
             "/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"
@@ -150,17 +165,19 @@ if [[ $CUDA_VERSION == 12* ]]; then
             "libcublasLt.so.12"
             "libcusparseLt.so.0"
             "libcudart.so.12"
+            "libnvToolsExt.so.1"
             "libnvrtc.so.12"
             "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
-            DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
-            DEPS_SONAME+=("libnvToolsExt.so.1")
+        if [[ $USE_CUFILE == 1 ]]; then
+            DEPS_LIST+=(
+                "/usr/local/cuda/lib64/libcufile.so.0"
+                "/usr/local/cuda/lib64/libcufile_rdma.so.1"
+            )
+            DEPS_SONAME+=(
+                "libcufile.so.0"
+                "libcufile_rdma.so.1"
+            )
         fi
     else
         echo "Using nvidia libs from pypi."
@@ -174,21 +191,94 @@ if [[ $CUDA_VERSION == 12* ]]; then
             '$ORIGIN/../../nvidia/curand/lib'
             '$ORIGIN/../../nvidia/cusolver/lib'
             '$ORIGIN/../../nvidia/cusparse/lib'
-            '$ORIGIN/../../nvidia/cusparselt/lib'
             '$ORIGIN/../../cusparselt/lib'
             '$ORIGIN/../../nvidia/nccl/lib'
-            '$ORIGIN/../../nvidia/nvshmem/lib'
             '$ORIGIN/../../nvidia/nvtx/lib'
-            '$ORIGIN/../../nvidia/cufile/lib'
+        )
+        if [[ $USE_CUFILE == 1 ]]; then
+            CUDA_RPATHS+=(
+                '$ORIGIN/../../nvidia/cufile/lib'
+            )
+        fi
+        CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
+        export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
+        export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'
+        export FORCE_RPATH="--force-rpath"
+        export USE_STATIC_NCCL=0
+        export USE_SYSTEM_NCCL=1
+        export ATEN_STATIC_CUDA=0
+        export USE_CUDA_STATIC_LINK=0
+        export USE_CUPTI_SO=1
+        export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+        export NCCL_LIB_DIR="/usr/local/cuda/lib64/"
+    fi
+elif [[ $CUDA_VERSION == "11.8" ]]; then
+    export USE_STATIC_CUDNN=0
+    # Try parallelizing nvcc as well
+    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
+    # Bundle ptxas into the wheel, see https://github.com/pytorch/pytorch/pull/119750
+    export BUILD_BUNDLE_PTXAS=1
+
+    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
+        echo "Bundling with cudnn and cublas."
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcudnn_adv.so.9"
+            "/usr/local/cuda/lib64/libcudnn_cnn.so.9"
+            "/usr/local/cuda/lib64/libcudnn_graph.so.9"
+            "/usr/local/cuda/lib64/libcudnn_ops.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9"
+            "/usr/local/cuda/lib64/libcudnn.so.9"
+            "/usr/local/cuda/lib64/libcublas.so.11"
+            "/usr/local/cuda/lib64/libcublasLt.so.11"
+            "/usr/local/cuda/lib64/libcudart.so.11.0"
+            "/usr/local/cuda/lib64/libnvToolsExt.so.1"
+            "/usr/local/cuda/lib64/libnvrtc.so.11.2"    # this is not a mistake, it links to more specific cuda version
+            "/usr/local/cuda/lib64/libnvrtc-builtins.so.11.8"
+        )
+        DEPS_SONAME+=(
+            "libcudnn_adv.so.9"
+            "libcudnn_cnn.so.9"
+            "libcudnn_graph.so.9"
+            "libcudnn_ops.so.9"
+            "libcudnn_engines_runtime_compiled.so.9"
+            "libcudnn_engines_precompiled.so.9"
+            "libcudnn_heuristic.so.9"
+            "libcudnn.so.9"
+            "libcublas.so.11"
+            "libcublasLt.so.11"
+            "libcudart.so.11.0"
+            "libnvToolsExt.so.1"
+            "libnvrtc.so.11.2"
+            "libnvrtc-builtins.so.11.8"
+        )
+    else
+        echo "Using nvidia libs from pypi."
+        CUDA_RPATHS=(
+            '$ORIGIN/../../nvidia/cublas/lib'
+            '$ORIGIN/../../nvidia/cuda_cupti/lib'
+            '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
+            '$ORIGIN/../../nvidia/cuda_runtime/lib'
+            '$ORIGIN/../../nvidia/cudnn/lib'
+            '$ORIGIN/../../nvidia/cufft/lib'
+            '$ORIGIN/../../nvidia/curand/lib'
+            '$ORIGIN/../../nvidia/cusolver/lib'
+            '$ORIGIN/../../nvidia/cusparse/lib'
+            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/nvtx/lib'
         )
         CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
         export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
         export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'
         export FORCE_RPATH="--force-rpath"
         export USE_STATIC_NCCL=0
+        export USE_SYSTEM_NCCL=1
         export ATEN_STATIC_CUDA=0
         export USE_CUDA_STATIC_LINK=0
         export USE_CUPTI_SO=1
+        export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+        export NCCL_LIB_DIR="/usr/local/cuda/lib64/"
     fi
 else
     echo "Unknown cuda version $CUDA_VERSION"

.ci/manywheel/build_libtorch.sh

@@ -22,7 +22,9 @@ retry () {
 
 # TODO move this into the Docker images
 OS_NAME=`awk -F= '/^NAME/{print $2}' /etc/os-release`
-if [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
     retry yum install -q -y zip openssl
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
     retry dnf install -q -y zip openssl
@@ -33,9 +35,6 @@ elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
     sed -i 's/.*nvidia.*/# &/' $(find /etc/apt/ -type f -name "*.list")
     retry apt-get update
     retry apt-get -y install zip openssl
-else
-    echo "Unknown OS: '$OS_NAME'"
-    exit 1
 fi
 
 # Version: setup.py uses $PYTORCH_BUILD_VERSION.post$PYTORCH_BUILD_NUMBER if
@@ -92,7 +91,6 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
     exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -qUr requirements-build.txt
 python setup.py clean
 retry pip install -qr requirements.txt
 retry pip install -q numpy==2.0.1
@@ -104,7 +102,7 @@ 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 +118,7 @@ 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

@@ -95,7 +95,6 @@ ROCM_SO_FILES=(
     "libroctracer64.so"
     "libroctx64.so"
     "libhipblaslt.so"
-    "libhipsparselt.so"
     "libhiprtc.so"
 )
 
@@ -187,28 +186,20 @@ do
     OS_SO_FILES[${#OS_SO_FILES[@]}]=$file_name # Append lib to array
 done
 
-ARCH=$(echo $PYTORCH_ROCM_ARCH | sed 's/;/|/g') # Replace ; separated arch list to bar for grep
-
 # rocBLAS library files
 ROCBLAS_LIB_SRC=$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 $OTHER_FILES)
+ARCH=$(echo $PYTORCH_ROCM_ARCH | sed 's/;/|/g') # Replace ; seperated arch list to bar for grep
+ARCH_SPECIFIC_FILES=$(ls $ROCBLAS_LIB_SRC | grep -E $ARCH)
+OTHER_FILES=$(ls $ROCBLAS_LIB_SRC | grep -v gfx)
+ROCBLAS_LIB_FILES=($ARCH_SPECIFIC_FILES $OTHER_FILES)
 
 # hipblaslt library files
 HIPBLASLT_LIB_SRC=$ROCM_HOME/lib/hipblaslt/library
 HIPBLASLT_LIB_DST=lib/hipblaslt/library
-HIPBLASLT_ARCH_SPECIFIC_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -E $ARCH)
-HIPBLASLT_OTHER_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -v gfx)
-HIPBLASLT_LIB_FILES=($HIPBLASLT_ARCH_SPECIFIC_FILES $HIPBLASLT_OTHER_FILES)
-
-# hipsparselt library files
-HIPSPARSELT_LIB_SRC=$ROCM_HOME/lib/hipsparselt/library
-HIPSPARSELT_LIB_DST=lib/hipsparselt/library
-HIPSPARSELT_ARCH_SPECIFIC_FILES=$(ls $HIPSPARSELT_LIB_SRC | grep -E $ARCH)
-#HIPSPARSELT_OTHER_FILES=$(ls $HIPSPARSELT_LIB_SRC | grep -v gfx)
-HIPSPARSELT_LIB_FILES=($HIPSPARSELT_ARCH_SPECIFIC_FILES $HIPSPARSELT_OTHER_FILES)
+ARCH_SPECIFIC_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -E $ARCH)
+OTHER_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -v gfx)
+HIPBLASLT_LIB_FILES=($ARCH_SPECIFIC_FILES $OTHER_FILES)
 
 # ROCm library files
 ROCM_SO_PATHS=()
@@ -243,14 +234,12 @@ DEPS_SONAME=(
 DEPS_AUX_SRCLIST=(
     "${ROCBLAS_LIB_FILES[@]/#/$ROCBLAS_LIB_SRC/}"
     "${HIPBLASLT_LIB_FILES[@]/#/$HIPBLASLT_LIB_SRC/}"
-    "${HIPSPARSELT_LIB_FILES[@]/#/$HIPSPARSELT_LIB_SRC/}"
     "/opt/amdgpu/share/libdrm/amdgpu.ids"
 )
 
 DEPS_AUX_DSTLIST=(
     "${ROCBLAS_LIB_FILES[@]/#/$ROCBLAS_LIB_DST/}"
     "${HIPBLASLT_LIB_FILES[@]/#/$HIPBLASLT_LIB_DST/}"
-    "${HIPSPARSELT_LIB_FILES[@]/#/$HIPSPARSELT_LIB_DST/}"
     "share/libdrm/amdgpu.ids"
 )
 

.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
 
@@ -23,12 +27,6 @@ cmake --version
 echo "Environment variables:"
 env
 
-# The sccache wrapped version of nvcc gets put in /opt/cache/lib in docker since
-# there are some issues if it is always wrapped, so we need to add it to PATH
-# during CI builds.
-# https://github.com/pytorch/pytorch/blob/0b6c0898e6c352c8ea93daec854e704b41485375/.ci/docker/common/install_cache.sh#L97
-export PATH="/opt/cache/lib:$PATH"
-
 if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
   # Use jemalloc during compilation to mitigate https://github.com/pytorch/pytorch/issues/116289
   export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so.2
@@ -54,6 +52,12 @@ fi
 export USE_LLVM=/opt/llvm
 export LLVM_DIR=/opt/llvm/lib/cmake/llvm
 
+if [[ "$BUILD_ENVIRONMENT" == *executorch* ]]; then
+  # To build test_edge_op_registration
+  export BUILD_EXECUTORCH=ON
+  export USE_CUDA=0
+fi
+
 if ! which conda; then
   # In ROCm CIs, we are doing cross compilation on build machines with
   # intel cpu and later run tests on machines with amd cpu.
@@ -120,8 +124,26 @@ if [[ "$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
@@ -176,8 +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* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]]; 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 [[ "${BUILD_ENVIRONMENT}" == *clang* ]]; then
@@ -203,7 +227,7 @@ if [[ "${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
 
@@ -233,7 +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
@@ -284,22 +307,6 @@ else
     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/
@@ -387,8 +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/check_binary.sh

@@ -313,7 +313,7 @@ if [[ "$(uname)" == 'Linux' &&  "$PACKAGE_TYPE" == 'manywheel' ]]; then
   # Please see issue for reference: https://github.com/pytorch/pytorch/issues/152426
   if [[ "$(uname -m)" == "s390x" ]]; then
     cxx_abi="19"
-  elif [[ "$DESIRED_CUDA" != 'xpu' && "$DESIRED_CUDA" != 'rocm'* ]]; then
+  elif [[ "$DESIRED_CUDA" != 'cu118' && "$DESIRED_CUDA" != 'xpu' && "$DESIRED_CUDA" != 'rocm'* ]]; then
     cxx_abi="18"
   else
     cxx_abi="16"

.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.sh

@@ -15,6 +15,6 @@ if [[ "${BUILD_ENVIRONMENT}" == *rocm* ]]; then
   export PYTORCH_TEST_WITH_ROCM=1
 fi
 
-# TODO: Reenable libtorch testing for MacOS, see https://github.com/pytorch/pytorch/issues/62598
+# TODO: Renable libtorch testing for MacOS, see https://github.com/pytorch/pytorch/issues/62598
 # shellcheck disable=SC2034
 BUILD_TEST_LIBTORCH=0

.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() {
@@ -174,19 +153,17 @@ 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
 }
 
+function install_tlparse() {
+  pip_install --user "tlparse==0.3.30"
+  PATH="$(python -m site --user-base)/bin:$PATH"
+}
+
 function install_torchrec_and_fbgemm() {
   local torchrec_commit
   torchrec_commit=$(get_pinned_commit torchrec)
@@ -201,73 +178,25 @@ 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.$MINOR_VERSION"
-
     pip_install tabulate  # needed for newer fbgemm
     pip_install patchelf  # needed for rocm fbgemm
-    pushd /tmp
-
-    local wheel_dir=dist/fbgemm_gpu
-    local found_whl=0
-    for file in "${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 "${fbgemm_commit}"
-      python setup.py bdist_wheel \
-        --build-variant=rocm \
-        -DHIP_ROOT_DIR="${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
-
-    rm -rf fbgemm
+    git clone --recursive https://github.com/pytorch/fbgemm
+    pushd fbgemm/fbgemm_gpu
+    git checkout "${fbgemm_commit}"
+    python setup.py install \
+      --package_variant=rocm \
+      -DHIP_ROOT_DIR="${ROCM_PATH}" \
+      -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
+      -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
     popd
+    rm -rf fbgemm
   else
-    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
 }
 
@@ -310,7 +239,7 @@ function checkout_install_torchbench() {
 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-build.sh

@@ -40,7 +40,7 @@ if [[ ${BUILD_ENVIRONMENT} == *"distributed"* ]]; then
 else
   # Explicitly set USE_DISTRIBUTED=0 to align with the default build config on mac. This also serves as the sole CI config that tests
   # that building with USE_DISTRIBUTED=0 works at all. See https://github.com/pytorch/pytorch/issues/86448
-  USE_DISTRIBUTED=0 USE_OPENMP=1 MACOSX_DEPLOYMENT_TARGET=11.0 WERROR=1 BUILD_TEST=OFF USE_PYTORCH_METAL=1 python setup.py bdist_wheel --plat-name macosx_11_0_arm64
+  USE_DISTRIBUTED=0 USE_OPENMP=1 MACOSX_DEPLOYMENT_TARGET=11.0 WERROR=1 BUILD_TEST=OFF USE_PYTORCH_METAL=1 python setup.py bdist_wheel
 fi
 if which sccache > /dev/null; then
   print_sccache_stats

.ci/pytorch/macos-common.sh

@@ -20,4 +20,14 @@ print_cmake_info() {
   CONDA_INSTALLATION_DIR=$(dirname "$CMAKE_EXEC")
   # Print all libraries under cmake rpath for debugging
   ls -la "$CONDA_INSTALLATION_DIR/../lib"
+
+  export CMAKE_EXEC
+  # Explicitly add conda env lib folder to cmake rpath to address the flaky issue
+  # where cmake dependencies couldn't be found. This seems to point to how conda
+  # links $CMAKE_EXEC to its package cache when cloning a new environment
+  install_name_tool -add_rpath @executable_path/../lib "${CMAKE_EXEC}" || true
+  # Adding the rpath will invalidate cmake signature, so signing it again here
+  # to trust the executable. EXC_BAD_ACCESS (SIGKILL (Code Signature Invalid))
+  # with an exit code 137 otherwise
+  codesign -f -s - "${CMAKE_EXEC}" || true
 }

.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
@@ -160,7 +165,6 @@ test_jit_hooks() {
 torchbench_setup_macos() {
   git clone --recursive https://github.com/pytorch/vision torchvision
   git clone --recursive https://github.com/pytorch/audio torchaudio
-  brew install jpeg-turbo libpng
 
   pushd torchvision
   git fetch
@@ -175,8 +179,7 @@ torchbench_setup_macos() {
   git checkout "$(cat ../.github/ci_commit_pins/audio.txt)"
   git submodule update --init --recursive
   python setup.py clean
-  #TODO: Remove me, when figure out how to make TorchAudio find brew installed openmp
-  USE_OPENMP=0 python setup.py develop
+  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
@@ -184,8 +187,9 @@ torchbench_setup_macos() {
   checkout_install_torchbench
 }
 
-pip_benchmark_deps() {
-  python -mpip install --no-input requests cython scikit-learn six
+conda_benchmark_deps() {
+  conda install -y astunparse numpy scipy ninja pyyaml setuptools cmake typing-extensions requests protobuf numba cython scikit-learn
+  conda install -y -c conda-forge librosa
 }
 
 
@@ -193,7 +197,7 @@ test_torchbench_perf() {
   print_cmake_info
 
   echo "Launching torchbench setup"
-  pip_benchmark_deps
+  conda_benchmark_deps
   torchbench_setup_macos
 
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
@@ -220,7 +224,7 @@ test_torchbench_smoketest() {
   print_cmake_info
 
   echo "Launching torchbench setup"
-  pip_benchmark_deps
+  conda_benchmark_deps
   # shellcheck disable=SC2119,SC2120
   torchbench_setup_macos
 
@@ -228,52 +232,53 @@ test_torchbench_smoketest() {
   mkdir -p "$TEST_REPORTS_DIR"
 
   local device=mps
-  local dtypes=(undefined float16 bfloat16 notset)
-  local dtype=${dtypes[$1]}
-  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor timm_resnet timm_vovnet vgg16)
+  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo)
+  local hf_models=(GoogleFnet YituTechConvBert)
 
   for backend in eager inductor; do
 
-    echo "Launching torchbench inference performance run for backend ${backend} and dtype ${dtype}"
-    local dtype_arg="--${dtype}"
-    if [ "$dtype" == notset ]; then
-        dtype_arg="--float32"
-    fi
-    touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv"
-    for model in "${models[@]}"; do
-      PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-        --performance --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-        --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv" || true
-      if [ "$backend" == "inductor" ]; then
-        PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-          --accuracy --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-          --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_accuracy.csv" || true
-      fi
-    done
-    if [ "$backend" == "inductor" ]; then
-      PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-        --performance --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-        --output "$TEST_REPORTS_DIR/inductor_${backend}_huggingface_${dtype}_inference_${device}_performance.csv" || true
-      PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-        --accuracy --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-        --output "$TEST_REPORTS_DIR/inductor_${backend}_huggingface_${dtype}_inference_${device}_accuracy.csv" || true
-    fi
-
-    if [ "$dtype" == notset ]; then
-      for dtype_ in notset amp; do
-        echo "Launching torchbench training performance run for backend ${backend} and dtype ${dtype_}"
-        touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype_}_training_${device}_performance.csv"
-        local dtype_arg="--${dtype_}"
-        if [ "$dtype_" == notset ]; then
+    for dtype in notset float16 bfloat16; do
+      echo "Launching torchbench inference performance run for backend ${backend} and dtype ${dtype}"
+      local dtype_arg="--${dtype}"
+      if [ "$dtype" == notset ]; then
           dtype_arg="--float32"
-        fi
-        for model in "${models[@]}"; do
+      fi
+      touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv"
+      for model in "${models[@]}"; do
+        PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
+          --performance --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
+          --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv" || true
+        if [ "$backend" == "inductor" ]; then
           PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-            --performance --only "$model" --backend "$backend" --training --devices "$device" "$dtype_arg" \
-            --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype_}_training_${device}_performance.csv" || true
-        done
+            --accuracy --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
+            --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_accuracy.csv" || true
+        fi
       done
-    fi
+      for model in "${hf_models[@]}"; do
+        if [ "$backend" == "inductor" ]; then
+          PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
+            --performance --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
+            --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv" || true
+          PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
+            --accuracy --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
+            --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_accuracy.csv" || true
+        fi
+      done
+    done
+
+    for dtype in notset amp; do
+      echo "Launching torchbench training performance run for backend ${backend} and dtype ${dtype}"
+      touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_training_${device}_performance.csv"
+      local dtype_arg="--${dtype}"
+      if [ "$dtype" == notset ]; then
+          dtype_arg="--float32"
+      fi
+      for model in "${models[@]}"; do
+        PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
+          --performance --only "$model" --backend "$backend" --training --devices "$device" "$dtype_arg" \
+          --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_training_${device}_performance.csv" || true
+      done
+    done
 
   done
 
@@ -284,7 +289,7 @@ test_hf_perf() {
   print_cmake_info
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
-  pip_benchmark_deps
+  conda_benchmark_deps
   torchbench_setup_macos
 
   echo "Launching HuggingFace training perf run"
@@ -300,7 +305,7 @@ test_timm_perf() {
   print_cmake_info
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
-  pip_benchmark_deps
+  conda_benchmark_deps
   torchbench_setup_macos
 
   echo "Launching timm training perf run"
@@ -312,6 +317,8 @@ test_timm_perf() {
   echo "timm benchmark on mps device completed"
 }
 
+install_tlparse
+
 if [[ $TEST_CONFIG == *"perf_all"* ]]; then
   test_torchbench_perf
   test_hf_perf
@@ -323,7 +330,7 @@ elif [[ $TEST_CONFIG == *"perf_hf"* ]]; then
 elif [[ $TEST_CONFIG == *"perf_timm"* ]]; then
   test_timm_perf
 elif [[ $TEST_CONFIG == *"perf_smoketest"* ]]; then
-  test_torchbench_smoketest "${SHARD_NUMBER}"
+  test_torchbench_smoketest
 elif [[ $TEST_CONFIG == *"mps"* ]]; then
   test_python_mps
 elif [[ $NUM_TEST_SHARDS -gt 1 ]]; then

.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/check_binary_symbols.py

@@ -93,7 +93,7 @@ def check_lib_symbols_for_abi_correctness(lib: str) -> None:
             f"Found pre-cxx11 symbols, but there shouldn't be any, see: {pre_cxx11_symbols[:100]}"
         )
     if num_cxx11_symbols < 100:
-        raise RuntimeError("Didn't find enough cxx11 symbols")
+        raise RuntimeError("Didn't find enought cxx11 symbols")
 
 
 def main() -> None:

.ci/pytorch/smoke_test/check_gomp.py

@@ -46,9 +46,6 @@ def get_gomp_thread():
 
     # use the default gomp path of AlmaLinux OS
     libgomp_path = "/usr/lib64/libgomp.so.1"
-    # if it does not exist, try Ubuntu path
-    if not os.path.exists(libgomp_path):
-        libgomp_path = f"/usr/lib/{os.uname().machine}-linux-gnu/libgomp.so.1"
 
     os.environ["GOMP_CPU_AFFINITY"] = "0-3"
 

.ci/pytorch/smoke_test/smoke_test.py

@@ -276,7 +276,7 @@ def smoke_test_cuda(
             torch_nccl_version = ".".join(str(v) for v in torch.cuda.nccl.version())
             print(f"Torch nccl; version: {torch_nccl_version}")
 
-        # Pypi dependencies are installed on linux only and nccl is available only on Linux.
+        # Pypi dependencies are installed on linux ony and nccl is availbale only on Linux.
         if pypi_pkg_check == "enabled" and sys.platform in ["linux", "linux2"]:
             compare_pypi_to_torch_versions(
                 "cudnn", find_pypi_package_version("nvidia-cudnn"), torch_cudnn_version
@@ -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 [[ "$TEST_CONFIG" == *crossref* ]]; then
@@ -196,12 +196,12 @@ if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   # shellcheck disable=SC1091
   source /opt/intel/oneapi/mpi/latest/env/vars.sh
   # Check XPU status before testing
-  timeout 30 xpu-smi discovery || true
+  xpu-smi discovery
 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:$PATH"
@@ -212,6 +212,8 @@ if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
   export VALGRIND=OFF
 fi
 
+install_tlparse
+
 # DANGER WILL ROBINSON.  The LD_PRELOAD here could cause you problems
 # if you're not careful.  Check this if you made some changes and the
 # ASAN test is not working
@@ -224,7 +226,7 @@ if [[ "$BUILD_ENVIRONMENT" == *asan* ]]; then
     export PYTORCH_TEST_WITH_ASAN=1
     export PYTORCH_TEST_WITH_UBSAN=1
     # TODO: Figure out how to avoid hard-coding these paths
-    export ASAN_SYMBOLIZER_PATH=/usr/lib/llvm-18/bin/llvm-symbolizer
+    export ASAN_SYMBOLIZER_PATH=/usr/lib/llvm-15/bin/llvm-symbolizer
     export TORCH_USE_RTLD_GLOBAL=1
     # NB: We load libtorch.so with RTLD_GLOBAL for UBSAN, unlike our
     # default behavior.
@@ -289,12 +291,6 @@ 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
@@ -328,29 +324,6 @@ test_python_smoke() {
   assert_git_not_dirty
 }
 
-test_h100_distributed() {
-  # Distributed tests at H100
-  time python test/run_test.py --include distributed/_composable/test_composability/test_pp_composability.py  $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  # This test requires multicast support
-  time python test/run_test.py --include distributed/_composable/fsdp/test_fully_shard_comm.py -k TestFullyShardAllocFromPG $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  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 +352,12 @@ 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 +409,14 @@ 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 +429,47 @@ 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 +482,7 @@ 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"
@@ -615,9 +590,7 @@ test_perf_for_dashboard() {
 
   local device=cuda
   if [[ "${TEST_CONFIG}" == *cpu* ]]; then
-    if [[ "${TEST_CONFIG}" == *cpu_x86_zen* ]]; then
-      device=cpu_x86_zen
-    elif [[ "${TEST_CONFIG}" == *cpu_x86* ]]; then
+    if [[ "${TEST_CONFIG}" == *cpu_x86* ]]; then
       device=cpu_x86
     elif [[ "${TEST_CONFIG}" == *cpu_aarch64* ]]; then
       device=cpu_aarch64
@@ -633,11 +606,7 @@ test_perf_for_dashboard() {
 
   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
@@ -649,10 +618,6 @@ 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 "$@" \
@@ -855,7 +820,16 @@ test_inductor_torchbench_smoketest_perf() {
   done
 }
 
+test_inductor_get_core_number() {
+  if [[ "${TEST_CONFIG}" == *aarch64* ]]; then
+    echo "$(($(lscpu | grep 'Cluster(s):' | awk '{print $2}') * $(lscpu | grep 'Core(s) per cluster:' | awk '{print $4}')))"
+  else
+    echo "$(($(lscpu | grep 'Socket(s):' | awk '{print $2}') * $(lscpu | grep 'Core(s) per socket:' | awk '{print $4}')))"
+  fi
+}
+
 test_inductor_set_cpu_affinity(){
+  #set jemalloc
   JEMALLOC_LIB="$(find /usr/lib -name libjemalloc.so.2)"
   export LD_PRELOAD="$JEMALLOC_LIB":"$LD_PRELOAD"
   export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1"
@@ -867,23 +841,14 @@ test_inductor_set_cpu_affinity(){
     export KMP_AFFINITY=granularity=fine,compact,1,0
     export KMP_BLOCKTIME=1
   fi
-
-  # Use nproc here instead of lscpu because it takes into account cgroups slice
-  cpus=$(nproc)
-  thread_per_core=$(lscpu | grep 'Thread(s) per core:' | awk '{print $4}')
-  cores=$((cpus / thread_per_core))
-
-  # Set number of cores to 16 on aarch64 for performance runs
+  cores=$(test_inductor_get_core_number)
+  # Set number of cores to 16 on Aarch64 for performance runs.
   if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
     cores=16
   fi
   export OMP_NUM_THREADS=$cores
-
-  # Handle cgroups slice start and end CPU
-  start_cpu=$(python -c 'import os; print(min(os.sched_getaffinity(0)))')
-  # Leaving one physical CPU for other tasks
-  end_cpu=$(($(python -c 'import os; print(max(os.sched_getaffinity(0)))') - thread_per_core))
-  export TASKSET="taskset -c $start_cpu-$end_cpu"
+  end_core=$((cores-1))
+  export TASKSET="taskset -c 0-$end_core"
 }
 
 test_inductor_torchbench_cpu_smoketest_perf(){
@@ -928,6 +893,12 @@ 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:
@@ -974,8 +945,6 @@ 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
 }
 
@@ -1162,12 +1131,6 @@ test_custom_backend() {
 
 test_custom_script_ops() {
   echo "Testing custom script operators"
-
-  if [[ "$BUILD_ENVIRONMENT" == *s390x* ]]; then
-    echo "Skipping custom script operators until it's fixed"
-    return 0
-  fi
-
   CUSTOM_OP_BUILD="${CUSTOM_TEST_ARTIFACT_BUILD_DIR}/custom-op-build"
   pushd test/custom_operator
   cp -a "$CUSTOM_OP_BUILD" build
@@ -1320,13 +1283,10 @@ 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 \
@@ -1481,8 +1441,8 @@ 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
@@ -1590,12 +1550,11 @@ 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 "$2" \
-      --output-csv "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
-      --output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.json" \
+      --output-dir "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv"
 
   pip_install pandas
   python check_perf_csv.py \
@@ -1610,13 +1569,7 @@ 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
@@ -1670,19 +1623,23 @@ 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
   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
   checkout_install_torchbench nanogpt
   PYTHONPATH=$(pwd)/torchbench test_verify_cachebench
 elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
-  install_torchaudio
+  if [[ "${TEST_CONFIG}" == *cpu* ]]; then
+    install_torchaudio cpu
+  else
+    install_torchaudio cuda
+  fi
   install_torchvision
-  install_torchao
+  TORCH_CUDA_ARCH_LIST="8.0;8.6" pip_install git+https://github.com/pytorch/ao.git
   id=$((SHARD_NUMBER-1))
   # https://github.com/opencv/opencv-python/issues/885
   pip_install opencv-python==4.8.0.74
@@ -1707,11 +1664,11 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
     PYTHONPATH=$(pwd)/torchbench test_dynamo_benchmark torchbench "$id"
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
+  install_torchaudio cuda
   install_torchvision
+  checkout_install_torchbench hf_T5 llama moco
   PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
-  if [[ "$SHARD_NUMBER" -eq "1" ]]; then
-    test_inductor_aoti
-  fi
+  test_inductor_aoti
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   test_inductor_shard "${SHARD_NUMBER}"
@@ -1720,8 +1677,6 @@ elif [[ "${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 "${SHARD_NUMBER}"
@@ -1769,12 +1724,6 @@ elif [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
   test_xpu_bin
 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-build.sh

@@ -31,7 +31,7 @@ PYLONG_API_CHECK=$?
 if [[ $PYLONG_API_CHECK == 0 ]]; then
   echo "Usage of PyLong_{From,As}{Unsigned}Long API may lead to overflow errors on Windows"
   echo "because \`sizeof(long) == 4\` and \`sizeof(unsigned long) == 4\`."
-  echo "Please include \"torch/csrc/utils/python_numbers.h\" and use the corresponding APIs instead."
+  echo "Please include \"torch/csrc/utils/python_numbers.h\" and use the correspoding APIs instead."
   echo "PyLong_FromLong -> THPUtils_packInt32 / THPUtils_packInt64"
   echo "PyLong_AsLong -> THPUtils_unpackInt (32-bit) / THPUtils_unpackLong (64-bit)"
   echo "PyLong_FromUnsignedLong -> THPUtils_packUInt32 / THPUtils_packUInt64"

.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

@@ -10,7 +10,7 @@ set PATH=C:\Program Files\CMake\bin;C:\Program Files\7-Zip;C:\ProgramData\chocol
 :: 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
+:: log sizes are too long, but leaving this here incase someone wants to use it locally
 :: set CMAKE_VERBOSE_MAKEFILE=1
 
 
@@ -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-helpers/run_python_nn_smoketests.py

@@ -52,7 +52,7 @@ if __name__ == "__main__":
             if os.path.exists(debugger):
                 command_args = [debugger, "-o", "-c", "~*g; q"] + command_args
                 command_string = " ".join(command_args)
-                print("Rerunning with traceback enabled")
+                print("Reruning with traceback enabled")
                 print("Command:", command_string)
                 subprocess.run(command_args, check=False)
             sys.exit(e.returncode)

.ci/pytorch/win-test.sh

@@ -38,10 +38,10 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
 fi
 
 # TODO: Move both of them to Windows AMI
-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
+python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 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/arm64/bootstrap_libuv.bat

@@ -7,7 +7,7 @@ if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
 if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
 
 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe
 
 cd %DEPENDENCIES_DIR%

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

@@ -7,7 +7,7 @@ if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
 if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%
 
 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe
 
 :: Clone OpenBLAS

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

@@ -2,7 +2,7 @@
 cd %PYTORCH_ROOT%
 
 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe
 
 :: create virtual environment

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

@@ -21,7 +21,7 @@ if %ENABLE_APL% == 1 (
 )
 
 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe
 
 :: change to source directory

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

@@ -21,7 +21,7 @@ if %ENABLE_APL% == 1 (
 )
 
 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe
 
 :: change to source directory

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

@@ -33,7 +33,7 @@ pushd tmp
 set VC_VERSION_LOWER=14
 set VC_VERSION_UPPER=36
 
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 
 set install_root=%CD%
 set INCLUDE=%INCLUDE%;%install_root%\include;%install_root%\include\torch\csrc\api\include

.ci/pytorch/windows/cuda118.bat

@@ -0,0 +1,59 @@
+@echo off
+
+set MODULE_NAME=pytorch
+
+IF NOT EXIST "setup.py" IF NOT EXIST "%MODULE_NAME%" (
+    call internal\clone.bat
+    cd %~dp0
+) ELSE (
+    call internal\clean.bat
+)
+IF ERRORLEVEL 1 goto :eof
+
+call internal\check_deps.bat
+IF ERRORLEVEL 1 goto :eof
+
+REM Check for optional components
+
+set USE_CUDA=
+set CMAKE_GENERATOR=Visual Studio 15 2017 Win64
+
+IF "%NVTOOLSEXT_PATH%"=="" (
+    IF EXIST "C:\Program Files\NVIDIA Corporation\NvToolsExt\lib\x64\nvToolsExt64_1.lib"  (
+        set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
+    ) ELSE (
+        echo NVTX ^(Visual Studio Extension ^for CUDA^) ^not installed, failing
+        exit /b 1
+    )
+)
+
+IF "%CUDA_PATH_V118%"=="" (
+    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin\nvcc.exe" (
+        set "CUDA_PATH_V118=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8"
+    ) ELSE (
+        echo CUDA 11.8 not found, failing
+        exit /b 1
+    )
+)
+
+IF "%BUILD_VISION%" == "" (
+    set TORCH_CUDA_ARCH_LIST=3.7+PTX;5.0;6.0;6.1;7.0;7.5;8.0;8.6;9.0
+    set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
+) ELSE (
+    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90
+)
+
+set "CUDA_PATH=%CUDA_PATH_V118%"
+set "PATH=%CUDA_PATH_V118%\bin;%PATH%"
+
+:optcheck
+
+call internal\check_opts.bat
+IF ERRORLEVEL 1 goto :eof
+
+if exist "%NIGHTLIES_PYTORCH_ROOT%" cd %NIGHTLIES_PYTORCH_ROOT%\..
+call  %~dp0\internal\copy.bat
+IF ERRORLEVEL 1 goto :eof
+
+call  %~dp0\internal\setup.bat
+IF ERRORLEVEL 1 goto :eof

.ci/pytorch/windows/cuda124.bat

@@ -27,24 +27,24 @@ 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"
+IF "%CUDA_PATH_V124%"=="" (
+    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\nvcc.exe" (
+        set "CUDA_PATH_V124=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
     ) ELSE (
-        echo CUDA 12.9 not found, failing
+        echo CUDA 12.4 not found, failing
         exit /b 1
     )
 )
 
 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=6.1;7.0;7.5;8.0;8.6;9.0
     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_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90
 )
 
-set "CUDA_PATH=%CUDA_PATH_V129%"
-set "PATH=%CUDA_PATH_V129%\bin;%PATH%"
+set "CUDA_PATH=%CUDA_PATH_V124%"
+set "PATH=%CUDA_PATH_V124%\bin;%PATH%"
 
 :optcheck
 

.ci/pytorch/windows/internal/check_deps.bat

@@ -65,7 +65,7 @@ for /F "usebackq delims=" %%i in (`python -c "import sys; print('{0[0]}{0[1]}'.f
 if  %PYVER% LSS 35 (
     echo Warning: PyTorch for Python 2 under Windows is experimental.
     echo Python x64 3.5 or up is recommended to compile PyTorch on Windows
-    echo Maybe you can create a virtual environment if you have conda installed:
+    echo Maybe you can create a virual environment if you have conda installed:
     echo ^> conda create -n test python=3.6 pyyaml numpy
     echo ^> activate test
 )

.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/cuda_install.bat

@@ -23,13 +23,66 @@ set CUDNN_LIB_FOLDER="lib\x64"
 :: Skip all of this if we already have cuda installed
 if exist "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v%CUDA_VERSION_STR%\bin\nvcc.exe" goto set_cuda_env_vars
 
+if %CUDA_VER% EQU 118 goto cuda118
+if %CUDA_VER% EQU 124 goto cuda124
 if %CUDA_VER% EQU 126 goto cuda126
 if %CUDA_VER% EQU 128 goto cuda128
-if %CUDA_VER% EQU 129 goto cuda129
 
 echo CUDA %CUDA_VERSION_STR% is not supported
 exit /b 1
 
+:cuda118
+
+set CUDA_INSTALL_EXE=cuda_11.8.0_522.06_windows.exe
+if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
+    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" & REM @lint-ignore
+    if errorlevel 1 exit /b 1
+    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
+    set "ARGS=cuda_profiler_api_11.8 thrust_11.8 nvcc_11.8 cuobjdump_11.8 nvprune_11.8 nvprof_11.8 cupti_11.8 cublas_11.8 cublas_dev_11.8 cudart_11.8 cufft_11.8 cufft_dev_11.8 curand_11.8 curand_dev_11.8 cusolver_11.8 cusolver_dev_11.8 cusparse_11.8 cusparse_dev_11.8 npp_11.8 npp_dev_11.8 nvrtc_11.8 nvrtc_dev_11.8 nvml_dev_11.8 nvtx_11.8"
+)
+
+set CUDNN_FOLDER=cudnn-windows-x86_64-9.5.0.50_cuda11-archive
+set CUDNN_LIB_FOLDER="lib"
+set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
+if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
+    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" & REM @lint-ignore
+    if errorlevel 1 exit /b 1
+    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
+)
+
+@REM cuDNN 8.3+ required zlib to be installed on the path
+echo Installing ZLIB dlls
+curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
+7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
+xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
+
+goto cuda_common
+
+:cuda124
+
+set CUDA_INSTALL_EXE=cuda_12.4.0_551.61_windows.exe
+if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
+    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" & REM @lint-ignore
+    if errorlevel 1 exit /b 1
+    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
+    set "ARGS=cuda_profiler_api_12.4 thrust_12.4 nvcc_12.4 cuobjdump_12.4 nvprune_12.4 nvprof_12.4 cupti_12.4 cublas_12.4 cublas_dev_12.4 cudart_12.4 cufft_12.4 cufft_dev_12.4 curand_12.4 curand_dev_12.4 cusolver_12.4 cusolver_dev_12.4 cusparse_12.4 cusparse_dev_12.4 npp_12.4 npp_dev_12.4 nvrtc_12.4 nvrtc_dev_12.4 nvml_dev_12.4 nvjitlink_12.4 nvtx_12.4"
+)
+
+set CUDNN_FOLDER=cudnn-windows-x86_64-9.5.0.50_cuda12-archive
+set CUDNN_LIB_FOLDER="lib"
+set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
+if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
+    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" & REM @lint-ignore
+    if errorlevel 1 exit /b 1
+    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
+)
+
+@REM cuDNN 8.3+ required zlib to be installed on the path
+echo Installing ZLIB dlls
+curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
+7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
+xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
+
 goto cuda_common
 
 :cuda126
@@ -86,33 +139,6 @@ xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
 
 goto cuda_common
 
-:cuda129
-
-set CUDA_INSTALL_EXE=cuda_12.9.1_576.57_windows.exe
-if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
-    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" & REM @lint-ignore
-    if errorlevel 1 exit /b 1
-    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
-    set "ARGS=cuda_profiler_api_12.9 thrust_12.9 nvcc_12.9 cuobjdump_12.9 nvprune_12.9 nvprof_12.9 cupti_12.9 cublas_12.9 cublas_dev_12.9 cudart_12.9 cufft_12.9 cufft_dev_12.9 curand_12.9 curand_dev_12.9 cusolver_12.9 cusolver_dev_12.9 cusparse_12.9 cusparse_dev_12.9 npp_12.9 npp_dev_12.9 nvrtc_12.9 nvrtc_dev_12.9 nvml_dev_12.9 nvjitlink_12.9 nvtx_12.9"
-)
-
-set CUDNN_FOLDER=cudnn-windows-x86_64-9.10.2.21_cuda12-archive
-set CUDNN_LIB_FOLDER="lib"
-set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
-if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
-    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" & REM @lint-ignore
-    if errorlevel 1 exit /b 1
-    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
-)
-
-@REM cuDNN 8.3+ required zlib to be installed on the path
-echo Installing ZLIB dlls
-curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
-7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
-xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
-
-goto cuda_common
-
 :cuda_common
 :: NOTE: We only install CUDA if we don't have it installed already.
 :: With GHA runners these should be pre-installed as part of our AMI process

.ci/pytorch/windows/internal/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

@@ -99,6 +99,7 @@ goto end
 :libtorch
 echo "install and test libtorch"
 
+if "%VC_YEAR%" == "2019" powershell internal\vs2019_install.ps1
 if "%VC_YEAR%" == "2022" powershell internal\vs2022_install.ps1
 
 if ERRORLEVEL 1 exit /b 1
@@ -110,6 +111,10 @@ pushd tmp\libtorch
 
 set VC_VERSION_LOWER=17
 set VC_VERSION_UPPER=18
+IF "%VC_YEAR%" == "2019" (
+    set VC_VERSION_LOWER=16
+    set VC_VERSION_UPPER=17
+)
 
 for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe" -legacy -products * -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
     if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (
@@ -148,7 +153,14 @@ 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/pytorch/windows/internal/vc_install_helper.bat

@@ -1,7 +1,12 @@
+if "%VC_YEAR%" == "2019" powershell windows/internal/vs2019_install.ps1
 if "%VC_YEAR%" == "2022" powershell windows/internal/vs2022_install.ps1
 
 set VC_VERSION_LOWER=17
 set VC_VERSION_UPPER=18
+if "%VC_YEAR%" == "2019" (
+    set VC_VERSION_LOWER=16
+    set VC_VERSION_UPPER=17
+)
 
 for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"  -products Microsoft.VisualStudio.Product.BuildTools -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
     if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (

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

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

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

@@ -25,8 +25,8 @@ set XPU_EXTRA_INSTALLED=0
 set XPU_EXTRA_UNINSTALL=0
 
 if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.1] (
-    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
-    set XPU_BUNDLE_VERSION=2025.1.3+5
+    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/1a9fff3d-04c2-4d77-8861-3d86c774b66f/intel-deep-learning-essentials-2025.1.1.26_offline.exe
+    set XPU_BUNDLE_VERSION=2025.1.1+23
 )
 
 :: Check if XPU bundle is target version or already installed

.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 $desired_python in
         ;;
     3.13)
         echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=68.0.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
         NUMPY_PINNED_VERSION="=2.1.0"
         ;;
     3.12)
         echo "Using 3.12 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=68.0.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.11)
         echo "Using 3.11 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=46.0.0"
         PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.10)
         echo "Using 3.10 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=46.0.0"
         PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.9)
         echo "Using 3.9 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=46.0.0"
         PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;
@@ -184,8 +184,7 @@ tmp_env_name="wheel_py$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
 
@@ -207,7 +206,7 @@ if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
     BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 python setup.py bdist_wheel -d "$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 "$whl_tmp_dir"
+    BUILD_PYTHON_ONLY=1 BUILD_LIBTORCH_WHL=0 python setup.py bdist_wheel -d "$whl_tmp_dir" --cmake
     echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
 else
     python setup.py bdist_wheel -d "$whl_tmp_dir"

.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
 
@@ -105,7 +105,6 @@ fi
 
 # Set triton via PYTORCH_EXTRA_INSTALL_REQUIREMENTS for triton xpu package
 if [[ "$PACKAGE_TYPE" =~ .*wheel.* && -n "$PYTORCH_BUILD_VERSION" && "$PYTORCH_BUILD_VERSION" =~ .*xpu.* ]]; then
-    TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_xpu_version.txt)
     TRITON_REQUIREMENT="pytorch-triton-xpu==${TRITON_VERSION}"
     if [[ -n "$PYTORCH_BUILD_VERSION" && "$PYTORCH_BUILD_VERSION" =~ .*dev.* ]]; then
         TRITON_SHORTHASH=$(cut -c1-8 $PYTORCH_ROOT/.ci/docker/ci_commit_pins/triton-xpu.txt)

.circleci/scripts/binary_windows_build.sh

@@ -9,7 +9,7 @@ if [[ "$OS" != "windows-arm64" ]]; then
     export USE_SCCACHE=1
     export SCCACHE_BUCKET=ossci-compiler-cache
     export SCCACHE_IGNORE_SERVER_IO_ERROR=1
-    export VC_YEAR=2022
+    export VC_YEAR=2019
 fi
 
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then

.circleci/scripts/binary_windows_test.sh

@@ -4,7 +4,7 @@ set -eux -o pipefail
 source "${BINARY_ENV_FILE:-/c/w/env}"
 
 export CUDA_VERSION="${DESIRED_CUDA/cu/}"
-export VC_YEAR=2022
+export VC_YEAR=2019
 
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
     export VC_YEAR=2022

.circleci/scripts/trigger_azure_pipeline.py

@@ -0,0 +1,157 @@
+# Documentation: https://docs.microsoft.com/en-us/rest/api/azure/devops/build/?view=azure-devops-rest-6.0
+
+import json
+import os
+import re
+import sys
+import time
+
+import requests
+
+
+AZURE_PIPELINE_BASE_URL = "https://aiinfra.visualstudio.com/PyTorch/"
+AZURE_DEVOPS_PAT_BASE64 = os.environ.get("AZURE_DEVOPS_PAT_BASE64_SECRET", "")
+PIPELINE_ID = "911"
+PROJECT_ID = "0628bce4-2d33-499e-bac5-530e12db160f"
+TARGET_BRANCH = os.environ.get("CIRCLE_BRANCH", "main")
+TARGET_COMMIT = os.environ.get("CIRCLE_SHA1", "")
+
+build_base_url = AZURE_PIPELINE_BASE_URL + "_apis/build/builds?api-version=6.0"
+
+s = requests.Session()
+s.headers.update({"Authorization": "Basic " + AZURE_DEVOPS_PAT_BASE64})
+
+
+def submit_build(pipeline_id, project_id, source_branch, source_version):
+    print("Submitting build for branch: " + source_branch)
+    print("Commit SHA1: ", source_version)
+
+    run_build_raw = s.post(
+        build_base_url,
+        json={
+            "definition": {"id": pipeline_id},
+            "project": {"id": project_id},
+            "sourceBranch": source_branch,
+            "sourceVersion": source_version,
+        },
+    )
+
+    try:
+        run_build_json = run_build_raw.json()
+    except json.decoder.JSONDecodeError as e:
+        print(e)
+        print(
+            "Failed to parse the response. Check if the Azure DevOps PAT is incorrect or expired."
+        )
+        sys.exit(-1)
+
+    build_id = run_build_json["id"]
+
+    print("Submitted bulid: " + str(build_id))
+    print("Bulid URL: " + run_build_json["url"])
+    return build_id
+
+
+def get_build(_id):
+    get_build_url = (
+        AZURE_PIPELINE_BASE_URL + f"/_apis/build/builds/{_id}?api-version=6.0"
+    )
+    get_build_raw = s.get(get_build_url)
+    return get_build_raw.json()
+
+
+def get_build_logs(_id):
+    get_build_logs_url = (
+        AZURE_PIPELINE_BASE_URL + f"/_apis/build/builds/{_id}/logs?api-version=6.0"
+    )
+    get_build_logs_raw = s.get(get_build_logs_url)
+    return get_build_logs_raw.json()
+
+
+def get_log_content(url):
+    resp = s.get(url)
+    return resp.text
+
+
+def wait_for_build(_id):
+    build_detail = get_build(_id)
+    build_status = build_detail["status"]
+
+    while build_status == "notStarted":
+        print("Waiting for run to start: " + str(_id))
+        sys.stdout.flush()
+        try:
+            build_detail = get_build(_id)
+            build_status = build_detail["status"]
+        except Exception as e:
+            print("Error getting build")
+            print(e)
+
+        time.sleep(30)
+
+    print("Bulid started: ", str(_id))
+
+    handled_logs = set()
+    while build_status == "inProgress":
+        try:
+            print("Waiting for log: " + str(_id))
+            logs = get_build_logs(_id)
+        except Exception as e:
+            print("Error fetching logs")
+            print(e)
+            time.sleep(30)
+            continue
+
+        for log in logs["value"]:
+            log_id = log["id"]
+            if log_id in handled_logs:
+                continue
+            handled_logs.add(log_id)
+            print("Fetching log: \n" + log["url"])
+            try:
+                log_content = get_log_content(log["url"])
+                print(log_content)
+            except Exception as e:
+                print("Error getting log content")
+                print(e)
+            sys.stdout.flush()
+        build_detail = get_build(_id)
+        build_status = build_detail["status"]
+        time.sleep(30)
+
+    build_result = build_detail["result"]
+
+    print("Bulid status: " + build_status)
+    print("Bulid result: " + build_result)
+
+    return build_status, build_result
+
+
+if __name__ == "__main__":
+    # Convert the branch name for Azure DevOps
+    match = re.search(r"pull/(\d+)", TARGET_BRANCH)
+    if match is not None:
+        pr_num = match.group(1)
+        SOURCE_BRANCH = f"refs/pull/{pr_num}/head"
+    else:
+        SOURCE_BRANCH = f"refs/heads/{TARGET_BRANCH}"
+
+    MAX_RETRY = 2
+    retry = MAX_RETRY
+
+    while retry > 0:
+        build_id = submit_build(PIPELINE_ID, PROJECT_ID, SOURCE_BRANCH, TARGET_COMMIT)
+        build_status, build_result = wait_for_build(build_id)
+
+        if build_result != "succeeded":
+            retry = retry - 1
+            if retry > 0:
+                print("Retrying... remaining attempt: " + str(retry))
+                # Wait a bit before retrying
+                time.sleep((MAX_RETRY - retry) * 120)
+                continue
+            else:
+                print("No more chance to retry. Giving up.")
+                sys.exit(-1)
+        else:
+            break

.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/ISSUE_TEMPLATE/bug-report.yml

@@ -12,9 +12,7 @@ body:
     description: |
       Please provide a clear and concise description of what the bug is.
 
-      If relevant, add a minimal example so that we can reproduce the error by running the code. It is very important for the snippet to be as succinct (minimal) as possible, so please take time to trim down any irrelevant code to help us debug efficiently.
-      Your example should be fully self-contained and not rely on any artifact that should be downloaded.
-      For example:
+      If relevant, add a minimal example so that we can reproduce the error by running the code. It is very important for the snippet to be as succinct (minimal) as possible, so please take time to trim down any irrelevant code to help us debug efficiently. We are going to copy-paste your code and we expect to get the same result as you did: avoid any external data, and include the relevant imports, etc. For example:
 
       ```python
       # All necessary imports at the beginning
@@ -28,7 +26,6 @@ body:
       If the code is too long (hopefully, it isn't), feel free to put it in a public gist and link it in the issue: https://gist.github.com.
 
       Please also paste or describe the results you observe instead of the expected results. If you observe an error, please paste the error message including the **full** traceback of the exception. It may be relevant to wrap error messages in ```` ```triple quotes blocks``` ````.
-      If your issue is related to numerical accuracy or reproducibility, please read the [numerical accuracy](https://docs.pytorch.org/docs/stable/notes/numerical_accuracy.html) and [reproducibility](https://docs.pytorch.org/docs/stable/notes/randomness.html) notes. If the difference is not expected as described in these documents, please provide appropriate justification on why one result is wrong and the other is correct.
     placeholder: |
       A clear and concise description of what the bug is.
 

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

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

.github/ISSUE_TEMPLATE/release-feature-request.yml

@@ -1,111 +0,0 @@
-name: 🚀 Release highlight for proposed Feature
-description: Submit a Release highlight for proposed Feature
-labels: ["release-feature-request"]
-
-body:
-- type: textarea
-  attributes:
-    label: Release highlight for proposed Feature
-    description: >
-      Example: “A torch.special module, analogous to SciPy's special module.”
-- type: input
-  id: contact
-  attributes:
-    label: Point(s) of contact
-    description: How can we get in touch with you if we need more info?
-    placeholder: ex. github username
-  validations:
-    required: false
-- type: dropdown
-  attributes:
-    label: Release Mode (pytorch/pytorch features only)
-    description: |
-      If "out-of-tree", please include the GH repo name
-    options:
-      - In-tree
-      - Out-of-tree
-  validations:
-    required: true
-- type: textarea
-  attributes:
-    label: Out-Of-Tree Repo
-    description: >
-      please include the GH repo name
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: Description and value to the user
-    description: >
-      Please provide a brief description of the feature and how it will benefit the user.
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: Link to design doc, GitHub issues, past submissions, etc
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: What feedback adopters have provided
-    description: >
-      Please list users/teams that have tried the feature and provided feedback. If that feedback motivated material changes (API, doc, etc..), a quick overview of the changes and the status (planned, in progress, implemented) would be helpful as well.
-  validations:
-    required: false
-- type: dropdown
-  attributes:
-    label: Plan for documentations / tutorials
-    description: |
-      Select One of the following options
-    options:
-      - Tutorial exists
-      - Will submit a PR to pytorch/tutorials
-      - Will submit a PR to a repo
-      - Tutorial is not needed
-  validations:
-    required: true
-- type: textarea
-  attributes:
-    label: Additional context for tutorials
-    description: >
-      Please provide a link for existing tutorial or link to a repo or context for why tutorial is not needed.
-  validations:
-    required: false
-- type: dropdown
-  attributes:
-    label: Marketing/Blog Coverage
-    description: |
-      Are you requesting feature Inclusion in the release blogs?
-    options:
-      - "Yes"
-      - "No"
-  validations:
-    required: true
-- type: textarea
-  attributes:
-    label: Are you requesting other marketing assistance with this feature?
-    description: >
-      E.g. supplementary blogs, social media amplification, etc.
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: Release Version
-    description: >
-      Please include release version for marketing coverage.
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: OS / Platform / Compute Coverage
-    description: >
-      Please list the platforms supported by the proposed feature. If the feature supports all the platforms, write "all". Goal of this section is to clearly share if this feature works in all PyTorch configurations or is it limited to only certain platforms/configurations (e.g. CPU only, GPU only, Linux only, etc...)
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: Testing Support (CI, test cases, etc..)
-    description: >
-      Please provide an overview of test coverage. This includes unit testing and integration testing, but if E2E validation testing has been done to show that the feature works for a certain set of use cases or models please mention that as well.
-  validations:
-    required: false

.github/actionlint.yaml

@@ -14,7 +14,6 @@ self-hosted-runner:
     - linux.12xlarge
     - linux.24xlarge
     - linux.24xlarge.ephemeral
-    - linux.24xlarge.amd
     - linux.arm64.2xlarge
     - linux.arm64.2xlarge.ephemeral
     - linux.arm64.m7g.4xlarge
@@ -46,19 +45,21 @@ self-hosted-runner:
     - windows.g5.4xlarge.nvidia.gpu
     # Windows ARM64 runners
     - windows-11-arm64
-    - windows-11-arm64-preview
     # Organization-wide AMD-hosted runners
     # MI2xx runners
     - linux.rocm.gpu
-    - 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 elswhere, 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}"