force kwarg=copy on .to in proxy_call in export

.ci/aarch64_linux/aarch64_ci_setup.sh

@@ -12,7 +12,7 @@ fi
 SCRIPTPATH="$( cd "$(dirname "$0")" ; pwd -P )"
 source $SCRIPTPATH/../manywheel/set_desired_python.sh
 
-pip install -q numpy==${NUMPY_VERSION} pyyaml==6.0.2 scons==4.7.0 ninja==1.11.1 patchelf==0.17.2
+pip install -q numpy==${NUMPY_VERSION} pyyaml==6.0.2 scons==4.7.0 ninja==1.11.1.4 patchelf==0.17.2
 
 for tool in python python3 pip pip3 ninja scons patchelf; do
     ln -sf ${DESIRED_PYTHON_BIN_DIR}/${tool} /usr/local/bin;

.ci/docker/README.md

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

.ci/docker/build.sh

@@ -93,6 +93,7 @@ tag=$(echo $image | awk -F':' '{print $2}')
 case "$tag" in
   pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11)
     CUDA_VERSION=12.4
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
@@ -103,6 +104,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
@@ -113,6 +115,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -124,6 +127,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=9
     VISION=yes
@@ -135,6 +139,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.13
     GCC_VERSION=9
     VISION=yes
@@ -146,6 +151,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9)
     CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -154,18 +160,9 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=11
-    VISION=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    TRITON=yes
-    ;;
   pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.6
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -177,6 +174,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.6
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=9
     VISION=yes
@@ -188,6 +186,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.6
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.13
     GCC_VERSION=9
     VISION=yes
@@ -199,6 +198,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -276,7 +276,7 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
+    pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
@@ -288,6 +288,7 @@ case "$tag" in
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
     ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
@@ -366,6 +367,7 @@ 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 [[ "$image" == *rocm* ]]; then
       extract_version_from_image_name rocm ROCM_VERSION
@@ -417,6 +419,9 @@ docker build \
        --build-arg "PYTHON_VERSION=${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:-}" \

.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} == "12.9" || ${CUDA_VERSION:0:4} == "12.8" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:4} == "12.6" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:2} == "11" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.1.0.70_cuda11-archive"
+    else
+        print "Unsupported CUDA version ${CUDA_VERSION}"
+        exit 1
+    fi
+    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/${CUDNN_NAME}.tar.xz
+    tar xf ${CUDNN_NAME}.tar.xz
+    cp -a ${CUDNN_NAME}/include/* /usr/local/cuda/include/
+    cp -a ${CUDNN_NAME}/lib/* /usr/local/cuda/lib64/
+    popd
+    rm -rf tmp_cudnn
+    ldconfig
+fi

.ci/docker/common/install_rocm.sh

@@ -30,7 +30,7 @@ EOF
 
     # we want the patch version of 6.4 instead
     if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-        ROCM_VERSION="${ROCM_VERSION}.2"
+        ROCM_VERSION="${ROCM_VERSION}.1"
     fi
 
     # Default url values
@@ -85,19 +85,16 @@ EOF
     # 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 6.4.1) ]]; then
+            HIP_BRANCH=release/rocm-rel-6.4
+            CLR_HASH=ca18eb3f77fa09292fcda62bc60c3e565d752ada  # branch release/rocm-rel-6.4.1-statco-hotfix
         elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-            HIP_TAG=rocm-6.4.0
+            HIP_BRANCH=release/rocm-rel-6.4
             CLR_HASH=600f5b0d2baed94d5121e2174a9de0851b040b0c  # branch release/rocm-rel-6.4-statco-hotfix
         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
+        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

.ci/docker/common/install_triton.sh

@@ -103,5 +103,5 @@ fi
 # 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
+  pip_install helion --no-deps
 fi

.ci/docker/libtorch/build.sh

@@ -41,7 +41,7 @@ case ${DOCKER_TAG_PREFIX} in
     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"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.1"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

.ci/docker/manywheel/Dockerfile_2_28

@@ -128,7 +128,7 @@ 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:$LD_LIBRARY_PATH
 # Install setuptools and wheel for python 3.12/3.13
 RUN for cpython_version in "cp312-cp312" "cp313-cp313" "cp313-cp313t"; do \
-    /opt/python/${cpython_version}/bin/python -m pip install setuptools wheel; \
+    /opt/python/${cpython_version}/bin/python -m pip install "setuptools>=77.0.0" "packaging>=24.2" wheel; \
     done;
 
 

.ci/docker/manywheel/Dockerfile_s390x

@@ -124,10 +124,9 @@ RUN python3 -mpip install cmake==3.28.0
 # 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' && \
+RUN pip3 install 'setuptools>=77.0' 'packaging>=24.2' && \
+  pip3 install flatbuffers cython 'pkgconfig>=1.5.5' 'numpy<2.3.0' && \
   pip3 install --no-build-isolation 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 && \

.ci/docker/manywheel/build.sh

@@ -77,7 +77,7 @@ case ${image} in
     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"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.1"
         fi
         TARGET=rocm_final
         MANY_LINUX_VERSION="2_28"

.ci/docker/requirements-ci.txt

@@ -104,10 +104,10 @@ networkx==2.8.8
 #Pinned versions: 2.8.8
 #test that import: functorch
 
-ninja==1.11.1.3
+ninja==1.11.1.4
 #Description: build system. Used in some tests. Used in build to generate build
 #time tracing information
-#Pinned versions: 1.11.1.3
+#Pinned versions: 1.11.1.4
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
 
 numba==0.49.0 ; python_version < "3.9"
@@ -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"
@@ -363,9 +363,10 @@ pwlf==2.2.1
 
 
 # To build PyTorch itself
+packaging>=24.2
 pyyaml
 pyzstd
-setuptools>=70.1.0
+setuptools>=77.0.0
 six
 
 scons==4.5.2 ; platform_machine == "aarch64"

.ci/docker/requirements-docs.txt

@@ -4,8 +4,8 @@ sphinx==5.3.0
 -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 it 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
@@ -50,8 +50,8 @@ 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
@@ -59,3 +59,4 @@ sphinx-copybutton==0.5.0
 sphinx-design==0.4.0
 sphinxcontrib-mermaid==1.0.0
 myst-parser==0.18.1
+myst-nb

.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
 
@@ -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* && "$BUILD_ENVIRONMENT" == *vulkan* ]]; then
   export USE_VULKAN=1
   # shellcheck disable=SC1091
   source /var/lib/jenkins/vulkansdk/setup-env.sh
@@ -203,7 +225,7 @@ if [[ "${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
 
@@ -247,6 +269,9 @@ if [[ "$BUILD_ENVIRONMENT" == *-bazel-* ]]; then
     tools/bazel build --config=no-tty "${BAZEL_MEM_LIMIT}" "${BAZEL_CPU_LIMIT}" //...
   fi
 else
+  # install build-system requirements before running setup.py commands
+  python -m pip install -r requirements-build.txt
+
   # check that setup.py would fail with bad arguments
   echo "The next three invocations are expected to fail with invalid command error messages."
   ( ! get_exit_code python setup.py bad_argument )

.ci/pytorch/common_utils.sh

@@ -204,32 +204,8 @@ function install_torchrec_and_fbgemm() {
     pip_build_and_install "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}" dist/torchrec
     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
@@ -247,7 +223,7 @@ function install_torchrec_and_fbgemm() {
       pushd fbgemm/fbgemm_gpu
       git checkout "${fbgemm_commit}"
       python setup.py bdist_wheel \
-        --build-variant=rocm \
+        --package_variant=rocm \
         -DHIP_ROOT_DIR="${ROCM_PATH}" \
         -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
         -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
@@ -264,7 +240,6 @@ function install_torchrec_and_fbgemm() {
     done
 
     rm -rf fbgemm
-    popd
   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

.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/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/smoke_test/smoke_test.py

@@ -35,7 +35,14 @@ MODULES = [
         "smoke_test": "./vision/test/smoke_test.py",
         "extension": "extension",
         "repo_name": "vision",
-    }
+    },
+    {
+        "name": "torchaudio",
+        "repo": "https://github.com/pytorch/audio.git",
+        "smoke_test": "./audio/test/smoke_test/smoke_test.py --no-ffmpeg",
+        "extension": "_extension",
+        "repo_name": "audio",
+    },
 ]
 
 
@@ -378,29 +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:
@@ -495,8 +479,6 @@ def main() -> None:
         options.pypi_pkg_check,
     )
 
-    smoke_test_nvshmem()
-
 
 if __name__ == "__main__":
     main()

.ci/pytorch/test.sh

@@ -201,7 +201,7 @@ fi
 
 if [[ "$BUILD_ENVIRONMENT" != *-bazel-* ]] ; then
   # JIT C++ extensions require ninja.
-  pip_install "ninja==1.10.2"
+  pip_install "ninja==1.11.1.4"
   # 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"
@@ -345,12 +345,6 @@ test_h100_symm_mem() {
   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 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-}
-
 test_lazy_tensor_meta_reference_disabled() {
   export TORCH_DISABLE_FUNCTIONALIZATION_META_REFERENCE=1
   echo "Testing lazy tensor operations without meta reference"
@@ -365,6 +359,7 @@ test_dynamo_wrapped_shard() {
     exit 1
   fi
   python tools/dynamo/verify_dynamo.py
+  python tools/dynamo/gb_id_mapping.py verify
   # PLEASE DO NOT ADD ADDITIONAL EXCLUDES HERE.
   # Instead, use @skipIfTorchDynamo on your tests.
   time python test/run_test.py --dynamo \
@@ -462,7 +457,7 @@ test_inductor_aoti() {
   # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
   /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${BUILD_COMMAND[@]}"
 
-  /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
+  /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference -dist=loadfile
 }
 
 test_inductor_cpp_wrapper_shard() {
@@ -928,6 +923,12 @@ 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 +975,6 @@ 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
 }
 
@@ -1320,13 +1319,10 @@ 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 \
@@ -1560,7 +1556,7 @@ test_executorch() {
 test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
-        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
+        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops test_cpp_extensions_open_device_registration \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1773,8 +1769,6 @@ elif [[ "${TEST_CONFIG}" == h100_distributed ]]; then
   test_h100_distributed
 elif [[ "${TEST_CONFIG}" == "h100-symm-mem" ]]; then
   test_h100_symm_mem
-elif [[ "${TEST_CONFIG}" == h100_cutlass_backend ]]; then
-  test_h100_cutlass_backend
 else
   install_torchvision
   install_monkeytype

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

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

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

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

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

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

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

@@ -126,6 +126,11 @@ if "%USE_CUDA%"=="1" (
   set CMAKE_CUDA_COMPILER_LAUNCHER=%TMP_DIR%/bin/randomtemp.exe;%TMP_DIR%\bin\sccache.exe
 )
 
+:: Install build-system requirements before running setup.py commands
+python -m pip install -r requirements-build.txt
+if errorlevel 1 goto fail
+if not errorlevel 0 goto fail
+
 :: Print all existing environment variable for debugging
 set
 

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

@@ -18,5 +18,5 @@ 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
+%PYTHON_EXEC% -m pip install --upgrade pip "setuptools>=77.0.0" "packaging>=24.2" wheel
 if errorlevel 1 exit /b 1

.ci/pytorch/windows/setup_build.bat

@@ -7,6 +7,9 @@ call "internal\install_python.bat"
 
 %PYTHON_EXEC% --version
 set "PATH=%CD%\Python\Lib\site-packages\cmake\data\bin;%CD%\Python\Scripts;%CD%\Python;%PATH%"
+
+%PYTHON_EXEC% -m pip install "setuptools>=77.0.0" "packaging>=24.2"
+
 if "%DESIRED_PYTHON%" == "3.13t" %PYTHON_EXEC% -m pip install numpy==2.2.1 cmake
 if "%DESIRED_PYTHON%" == "3.13" %PYTHON_EXEC% -m pip install numpy==2.1.2 cmake
 if "%DESIRED_PYTHON%" == "3.12" %PYTHON_EXEC% -m pip install numpy==2.0.2 cmake
@@ -16,7 +19,7 @@ if "%DESIRED_PYTHON%" == "3.9" %PYTHON_EXEC% -m pip install numpy==2.0.2 cmake
 
 %PYTHON_EXEC% -m pip install pyyaml
 %PYTHON_EXEC% -m pip install mkl-include mkl-static
-%PYTHON_EXEC% -m pip install boto3 ninja typing_extensions setuptools==72.1.0
+%PYTHON_EXEC% -m pip install boto3 ninja typing-extensions
 
 where cmake.exe
 

.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="==77.0.0"
 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=">=77.0.0"
         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=">=77.0.0"
         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=">=77.0.0"
         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=">=77.0.0"
         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=">=77.0.0"
         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=">=77.0.0"
         PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="=2.0.2"
         ;;

.flake8

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

.github/actionlint.yaml

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

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

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

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

@@ -70,7 +70,7 @@ runs:
           set -eux
           # PyYAML 6.0 doesn't work with MacOS x86 anymore
           # This must run on Python-3.7 (AmazonLinux2) so can't use request=3.32.2
-          python3 -m pip install requests==2.27.1 pyyaml==6.0.2
+          python3 -m pip install requests==2.27.1 pyyaml==6.0.1
 
     - name: Parse ref
       id: parse-ref

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-bf305f538005f2e900f8850ed57146024a8bc559
+00b0c91db92c51a11356249262577b9fa26c18c5

.github/ci_commit_pins/fbgemm_rocm.txt

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

.github/ci_commit_pins/vllm.txt

@@ -1 +0,0 @@
-ca9e2be3ed6320b51f52f536595cd24e254f8bb2

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-29ae4c76c026185f417a25e841d2cd5e65f087a3
+1c00dea2c9adb2137903c86b4191e8c247f8fda9

.github/merge_rules.yaml

@@ -76,8 +76,8 @@
   - .github/ci_commit_pins/audio.txt
   - .github/ci_commit_pins/vision.txt
   - .github/ci_commit_pins/torchdynamo.txt
-  - .github/ci_commit_pins/vllm.txt
   - .ci/docker/ci_commit_pins/triton.txt
+  - .ci/docker/ci_commit_pins/vllm.txt
   approved_by:
   - pytorchbot
   mandatory_checks_name:
@@ -131,6 +131,21 @@
   - Lint
   - pull
 
+- name: Mobile
+  patterns:
+  - ios/**
+  - android/**
+  - test/mobile/**
+  approved_by:
+  - linbinyu
+  - IvanKobzarev
+  - dreiss
+  - raziel
+  mandatory_checks_name:
+  - EasyCLA
+  - Lint
+  - pull
+
 - name: PrimTorch
   patterns:
   - torch/_meta_registrations.py
@@ -477,19 +492,6 @@
   - srossross
   - chillee
   - zou3519
-  - guilhermeleobas
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
-- name: Dynamo
-  patterns:
-  - torch/_dynamo/**
-  - torch/csrc/dynamo/**
-  - test/dynamo/**
-  approved_by:
-  - guilhermeleobas
   mandatory_checks_name:
   - EasyCLA
   - Lint

.github/pytorch-probot.yml

@@ -31,9 +31,7 @@ ciflow_push_tags:
 - ciflow/pull
 - ciflow/h100
 - ciflow/h100-distributed
-- ciflow/win-arm64
 - ciflow/h100-symm-mem
-- ciflow/h100-cutlass-backend
 retryable_workflows:
 - pull
 - trunk

.github/requirements-gha-cache.txt

@@ -7,9 +7,9 @@
 #   .ci/docker/requirements-ci.txt
 boto3==1.35.42
 jinja2==3.1.6
-lintrunner==0.12.7
-ninja==1.10.0.post1
+lintrunner==0.10.7
+ninja==1.11.1.4
 nvidia-ml-py==11.525.84
-pyyaml==6.0.2
+pyyaml==6.0
 requests==2.32.4
-rich==14.1.0
+rich==10.9.0

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

@@ -2,17 +2,17 @@ boto3==1.35.42
 cmake==3.27.*
 expecttest==0.3.0
 fbscribelogger==0.1.7
-filelock==3.18.0
+filelock==3.6.0
 hypothesis==6.56.4
 librosa>=0.6.2
 mpmath==1.3.0
 networkx==2.8.7
-ninja==1.10.2.4
+ninja==1.11.1.4
 numba==0.59.0
 numpy==1.26.4
 opt-einsum>=3.3
 optree==0.13.0
-packaging==23.1
+packaging==25.0
 parameterized==0.8.1
 pillow==10.3.0
 protobuf==5.29.4
@@ -26,7 +26,7 @@ pytest-xdist==3.3.1
 pytest==7.3.2
 pyyaml==6.0.2
 scipy==1.12.0
-setuptools==72.1.0
+setuptools==80.9.0
 sympy==1.13.3
 tlparse==0.3.30
 tensorboard==2.13.0

.github/scripts/lintrunner.sh

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

.github/scripts/trymerge.py

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

.github/scripts/windows/build_triton.bat

@@ -10,7 +10,7 @@ if "%PY_VERS%" == "3.13t" (
     call conda create -n %PYTHON_PREFIX% -y -c=conda-forge python=%PY_VERS%
 )
 :: Fix cmake version for issue https://github.com/pytorch/pytorch/issues/150480
-call conda run -n %PYTHON_PREFIX% pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==72.1.0 ninja
+call conda run -n %PYTHON_PREFIX% pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==78.1.1 ninja
 
 dir "%VC_INSTALL_PATH%"
 

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

@@ -27,7 +27,7 @@ jobs:
             PR_NUMBER="${{ github.event.number }}"
 
             # Use gh CLI to get changed files in the PR with explicit repo
-            CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
+            CHANGED_FILES=$(gh pr view "$PR_NUMBER" --repo "${{ github.repository }}" --json files --jq '.files[].path' | tr '\n' ' ' | sed 's/ $//')
 
             if [ -z "$CHANGED_FILES" ]; then
               echo "No changed files found, setting to '*'"
@@ -40,4 +40,4 @@ jobs:
           else
             echo "Not in PR context, setting changed files to '*'"
             echo "changed-files=*" >> "$GITHUB_OUTPUT"
-          fi
+          fi

.github/workflows/_mac-test.yml

@@ -80,6 +80,11 @@ jobs:
         run: |
           sysctl machdep.cpu.brand_string kern.osproductversion
 
+      - name: Install build toolchain
+        run: |
+          brew update --quiet
+          brew install --formula cmake ninja
+
       - name: Clean up leftover processes on MacOS pet runner
         continue-on-error: true
         run: |

.github/workflows/_rocm-test.yml

@@ -269,8 +269,8 @@ jobs:
           # copy test results back to the mounted workspace, needed sudo, resulting permissions were correct
           docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "cd ../pytorch && sudo cp -R test/test-reports ../workspace/test"
 
-      - name: Change permissions (only needed for kubernetes runners for now)
-        if: ${{ always() && steps.test.conclusion && (contains(matrix.runner, 'gfx942') || contains(matrix.runner, 'mi355')) }}
+      - name: Change permissions (only needed for MI300 runners for now)
+        if: ${{ always() && steps.test.conclusion && contains(matrix.runner, 'mi300') }}
         run: |
           docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "sudo chown -R 1001:1001 test"
 

.github/workflows/build-triton-wheel.yml

@@ -50,7 +50,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
+        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t" ]
         device: ["cuda", "rocm", "xpu", "aarch64"]
         docker-image: ["pytorch/manylinux2_28-builder:cpu"]
         include:
@@ -126,12 +126,6 @@ jobs:
           3.13t)
             PYTHON_EXECUTABLE=/opt/python/cp313-cp313t/bin/python
             ;;
-          3.14)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314/bin/python
-            ;;
-          3.14t)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314t/bin/python
-            ;;
           *)
             echo "Unsupported python version ${PY_VERS}"
             exit 1

.github/workflows/check_mergeability_ghstack.yml

@@ -56,7 +56,7 @@ jobs:
           cache: pip
           architecture: x64
 
-      - run: pip install pyyaml==6.0.2
+      - run: pip install pyyaml==6.0
         shell: bash
 
       - name: Verify mergeability

.github/workflows/cherry-pick.yml

@@ -26,7 +26,7 @@ jobs:
           cache: pip
 
       # Not the direct dependencies but the script uses trymerge
-      - run: pip install pyyaml==6.0.2
+      - run: pip install pyyaml==6.0
 
       - name: Setup committer id
         run: |

.github/workflows/docker-builds.yml

@@ -50,7 +50,6 @@ jobs:
         runner: [linux.12xlarge]
         docker-image-name: [
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm,
           pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks,

.github/workflows/docker-release.yml

@@ -144,7 +144,7 @@ jobs:
         run: |
           make -f docker.Makefile "${BUILD_IMAGE_TYPE}-image"
       - name: Push nightly tags
-        if: ${{ github.event.ref == 'refs/heads/nightly' && matrix.image_type == 'runtime' && matrix.platform == 'linux/amd4' }}
+        if: ${{ github.event.ref == 'refs/heads/nightly' && matrix.image_type == 'runtime' && matrix.build_platforms == 'linux/amd4' }}
         run: |
           PYTORCH_DOCKER_TAG="${PYTORCH_VERSION}-cuda${CUDA_VERSION_SHORT}-cudnn${CUDNN_VERSION}-runtime"
           CUDA_SUFFIX="-cu${CUDA_VERSION}"

.github/workflows/h100-cutlass-backend.yml

@@ -1,58 +0,0 @@
-name: Limited CI for CUTLASS backend on H100
-
-on:
-  pull_request:
-    paths:
-      - .github/workflows/h100-cutlass-backend.yml
-  workflow_dispatch:
-  schedule:
-    - cron: 22 9 * * *  # every 24 hours about 2:22am PDT
-  push:
-    tags:
-      - ciflow/h100-cutlass-backend/*
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-
-  get-label-type:
-    if: github.repository_owner == 'pytorch'
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-cuda12_8-py3_10-gcc11-sm90-build-cutlass-backend:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm90-cutlass-backend
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90-cutlass-backend
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '9.0'
-      test-matrix: |
-        { include: [
-          { config: "h100_cutlass_backend", shard: 1, num_shards: 1, runner: "linux.aws.h100", owners: ["oncall:pt2"] },
-        ]}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc11-sm90-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm90-cutlass-backend
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-sm90-build-cutlass-backend
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90-cutlass-backend
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm90-build-cutlass-backend.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm90-build-cutlass-backend.outputs.test-matrix }}
-    secrets: inherit

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

@@ -2,7 +2,7 @@ name: inductor-perf-nightly-h100
 
 on:
   schedule:
-    - cron: 15 0,12 * * 1-6
+    - cron: 15 0,4,8,12,16,20 * * 1-6
     - cron: 0 7 * * 0
   # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
   # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
@@ -126,7 +126,7 @@ jobs:
     name: cuda12.8-py3.10-gcc9-sm90
     uses: ./.github/workflows/_linux-test.yml
     needs: build
-    if: github.event.schedule == '15 0,12 * * 1-6'
+    if: github.event.schedule == '15 0,4,8,12,16,20 * * 1-6'
     with:
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true

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

@@ -88,23 +88,23 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.gfx942.2" },
+          { config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
         ]}
     secrets: inherit
 

.github/workflows/inductor-rocm-mi300.yml

@@ -47,8 +47,8 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       test-matrix: |
         { include: [
-          { config: "inductor", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "inductor", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
+          { config: "inductor", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
         ]}
     secrets: inherit
 

.github/workflows/mac-mps.yml

@@ -28,6 +28,7 @@ jobs:
       # than our AWS macos-m1-14 runners
       test-matrix: |
         { include: [
+          { config: "test_mps", shard: 1, num_shards: 1, runner: "macos-m1-13" },
           { config: "test_mps", shard: 1, num_shards: 1, runner: "macos-m1-14" },
           { config: "test_mps", shard: 1, num_shards: 1, runner: "macos-m2-15" },
         ]}

.github/workflows/nightly.yml

@@ -86,7 +86,7 @@ jobs:
           - repo-name: vllm
             repo-owner: vllm-project
             branch: main
-            pin-folder: .github/ci_commit_pins
+            pin-folder: .ci/docker/ci_commit_pins
     # Allow this to be triggered on either a schedule or on workflow_dispatch to allow for easier testing
     if: github.repository_owner == 'pytorch' && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
     steps:

.github/workflows/periodic-rocm-mi300.yml

@@ -59,9 +59,9 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       test-matrix: |
         { include: [
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi300.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi300.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi300.4", owners: ["module:rocm", "oncall:distributed"] },
         ]}
     secrets: inherit
 

.github/workflows/pull.yml

@@ -315,6 +315,21 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.test-matrix }}
     secrets: inherit
 
+  linux-jammy-py3-clang18-mobile-build:
+    name: linux-jammy-py3-clang18-mobile-build
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-py3-clang12-mobile-build
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
+      build-generates-artifacts: false
+      test-matrix: |
+        { include: [
+          { config: "default", shard: 1, num_shards: 1 },
+        ]}
+    secrets: inherit
+
   linux-jammy-cuda12_8-cudnn9-py3_9-clang12-build:
     name: linux-jammy-cuda12.8-cudnn9-py3.9-clang12
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/revert.yml

@@ -26,7 +26,7 @@ jobs:
           architecture: x64
           check-latest: false
           cache: pip
-      - run: pip install pyyaml==6.0.2
+      - run: pip install pyyaml==6.0
 
       - name: Setup committer id
         run: |

.github/workflows/rocm-mi300.yml

@@ -48,12 +48,12 @@ jobs:
       sync-tag: rocm-build
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.gfx942.2" },
+          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi300.2" },
         ]}
     secrets: inherit
 

.github/workflows/rocm-mi355.yml

@@ -1,68 +0,0 @@
-name: rocm-mi355
-
-on:
-  workflow_dispatch:
-  schedule:
-    - cron: 30 11,1 * * *  # about 4:30am PDT and 6:30pm PDT
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  target-determination:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-noble-rocm-py3_12-build:
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    name: linux-noble-rocm-py3.12-mi355
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-noble-rocm-py3.12-mi355
-      docker-image-name: ci-image:pytorch-linux-noble-rocm-alpha-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-        ]}
-    secrets: inherit
-
-  linux-noble-rocm-py3_12-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-noble-rocm-py3.12-mi355
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-noble-rocm-py3_12-build
-      - target-determination
-    with:
-      build-environment: linux-noble-rocm-py3.12-mi355
-      docker-image: ${{ needs.linux-noble-rocm-py3_12-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-noble-rocm-py3_12-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
-    secrets: inherit

.github/workflows/trunk.yml

@@ -94,6 +94,7 @@ jobs:
           { config: "default", shard: 1, num_shards: 3, runner: "macos-m1-stable" },
           { config: "default", shard: 2, num_shards: 3, runner: "macos-m1-stable" },
           { config: "default", shard: 3, num_shards: 3, runner: "macos-m1-stable" },
+          { config: "mps", shard: 1, num_shards: 1, runner: "macos-m1-13" },
           { config: "mps", shard: 1, num_shards: 1, runner: "macos-m1-14" },
           { config: "mps", shard: 1, num_shards: 1, runner: "macos-m2-15" },
         ]}

.github/workflows/trymerge.yml

@@ -28,7 +28,7 @@ jobs:
           check-latest: false
           cache: pip
           architecture: x64
-      - run: pip install pyyaml==6.0.2
+      - run: pip install pyyaml==6.0
 
       - name: Setup committer id
         run: |

.github/workflows/tryrebase.yml

@@ -25,7 +25,7 @@ jobs:
           architecture: x64
           check-latest: false
           cache: pip
-      - run: pip install pyyaml==6.0.2
+      - run: pip install pyyaml==6.0
 
       - name: Setup committer id
         run: |

.github/workflows/upload-test-stats.yml

@@ -14,7 +14,6 @@ on:
       - inductor-periodic
       - rocm
       - rocm-mi300
-      - rocm-mi355
       - inductor-micro-benchmark
       - inductor-micro-benchmark-x86
       - inductor-cu124

.github/workflows/win-arm64-build-test.yml

@@ -1,187 +0,0 @@
-name: windows-arm64-build-test
-
-on:
-  push:
-    tags:
-      - ciflow/win-arm64/*
-
-env:
-  GIT_DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
-  PYTHON_VERSION: "3.12"
-  PYTORCH_ROOT: ${{ github.workspace }}/pytorch
-  DOWNLOADS_DIR: c:\temp\downloads
-  DEPENDENCIES_DIR: c:\temp\dependencies
-  ENABLE_APL: 1
-  ENABLE_OPENBLAS: 0
-  BUILD_TYPE: release
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  build:
-    # Don't run on forked repos.
-    if: github.repository_owner == 'pytorch'
-    runs-on: "windows-11-arm64-preview"
-    timeout-minutes: 240
-    steps:
-      - name: configure aws credentials
-        id: aws_creds
-        uses: aws-actions/configure-aws-credentials@v4
-        with:
-          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_sscache
-          aws-region: us-east-1
-          role-duration-seconds: 18000
-
-      - name: Enable long paths
-        shell: cmd
-        run: |
-          git config --system --get core.longpaths || echo "core.longpaths is not set, setting it now"
-          git config --system core.longpaths true
-
-      - name: Git checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          path: pytorch
-          submodules: recursive
-
-      - name: Bootstrap Python
-        shell: cmd
-        run: |
-          "pytorch/.ci/pytorch/windows/arm64/bootstrap_python.bat"
-
-      - name: Parse ref
-        id: parse-ref
-        shell: bash
-        run: python pytorch/.github/scripts/parse_ref.py
-
-      - name: Get workflow job id
-        shell: bash
-        id: get-job-id
-        run: |
-          set -eux
-          python pytorch/.github/scripts/get_workflow_job_id.py "${GITHUB_RUN_ID}" "${RUNNER_NAME}"
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Bootstrap APL
-        shell: cmd
-        run: |
-          "pytorch/.ci/pytorch/windows/arm64/bootstrap_apl.bat"
-
-      - name: Bootstrap Rust
-        shell: cmd
-        run: |
-          "pytorch/.ci/pytorch/windows/arm64/bootstrap_rust.bat"
-
-      - name: Bootstrap sccache
-        shell: cmd
-        run: |
-          "pytorch/.ci/pytorch/windows/arm64/bootstrap_sccache.bat"
-
-      - name: Bootstrap Libuv
-        shell: cmd
-        run: |
-          "pytorch/.ci/pytorch/windows/arm64/bootstrap_libuv.bat"
-
-      - name: Build
-        id: build
-        shell: cmd
-        env:
-          PYTORCH_FINAL_PACKAGE_DIR: C:/${{ github.run_id }}/build-results/
-          BRANCH: ${{ steps.parse-ref.outputs.branch }}
-          BUILD_WHEEL: 1
-          MAX_JOBS: 8
-          PYTHON_VERSION: "3.12"
-          SCCACHE_BUCKET: "ossci-compiler-cache"
-          SCCACHE_S3_KEY_PREFIX: ${{ github.workflow }}
-          SCCACHE_REGION: us-east-1
-          VC_PRODUCT: "BuildTools"
-          VC_VERSION: ""
-          ALPINE_IMAGE: "308535385114.dkr.ecr.us-east-1.amazonaws.com/tool/alpine"
-          AWS_DEFAULT_REGION: us-east-1
-          USE_CUDA: '0'
-          USE_XPU: '0'
-          OUR_GITHUB_JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
-        run: |
-          cd pytorch
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
-          powershell -ExecutionPolicy Bypass -File ".ci/pytorch/win-arm64-build.ps1"
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4.4.0
-        if: always()
-        with:
-          name: torch-wheel-win-arm64-py3-12
-          retention-days: 14
-          if-no-files-found: error
-          path: C:\${{ github.run_id }}\build-results
-
-  test:
-    if: github.repository_owner == 'pytorch'
-    strategy:
-      fail-fast: false
-    runs-on: "windows-11-arm64-preview"
-    needs: build
-    steps:
-      - name: Enable long paths
-        shell: cmd
-        run: |
-          git config --system --get core.longpaths || echo "core.longpaths is not set, setting it now"
-          git config --system core.longpaths true
-
-      - name: Git checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          path: pytorch
-          submodules: recursive
-
-      - name: Bootstrap Python
-        shell: cmd
-        run: |
-          "pytorch/.ci/pytorch/windows/arm64/bootstrap_python.bat"
-
-      - name: Bootstrap Rust
-        shell: cmd
-        run: |
-          "pytorch/.ci/pytorch/windows/arm64/bootstrap_rust.bat"
-
-      - name: Get workflow job id
-        shell: bash
-        id: get-job-id
-        run: |
-          set -eux
-          python pytorch/.github/scripts/get_workflow_job_id.py "${GITHUB_RUN_ID}" "${RUNNER_NAME}"
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Download Build Artifacts
-        uses: actions/download-artifact@v4.1.7
-        with:
-          name: torch-wheel-win-arm64-py3-12
-          path: C:\${{ github.run_id }}\build-results
-
-      - name: Test
-        id: test
-        shell: cmd
-        env:
-          USE_CUDA: '0'
-          INSTALL_WINDOWS_SDK: 1
-          PYTHON_VERSION: "3.12"
-          VC_PRODUCT: "BuildTools"
-          AWS_DEFAULT_REGION: us-east-1
-          GITHUB_REPOSITORY: ${{ github.repository }}
-          GITHUB_WORKFLOW: ${{ github.workflow }}
-          GITHUB_JOB: ${{ github.job }}
-          GITHUB_RUN_ID: ${{ github.run_id }}
-          GITHUB_RUN_NUMBER: ${{ github.run_number }}
-          GITHUB_RUN_ATTEMPT: ${{ github.run_attempt }}
-          JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
-          JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
-          PYTORCH_FINAL_PACKAGE_DIR: C:/${{ github.run_id }}/build-results/
-        run: |
-          mkdir "%PYTORCH_FINAL_PACKAGE_DIR%"
-          call pytorch/.ci/pytorch/windows/arm64/bootstrap_tests.bat
-          set GIT_BASH=C:\Program Files\Git\usr\bin\bash.exe
-          "%GIT_BASH%" -c "bash --noprofile --norc .ci/pytorch/win-arm64-test.sh"

.lintrunner.toml

@@ -39,16 +39,16 @@ init_command = [
     'python3',
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
-    'flake8==7.3.0',
-    'flake8-bugbear==24.12.12',
-    'flake8-comprehensions==3.16.0',
+    'flake8==6.1.0',
+    'flake8-bugbear==23.3.23',
+    'flake8-comprehensions==3.15.0',
     'flake8-executable==2.1.3',
-    'flake8-logging-format==2024.24.12',
-    'flake8-pyi==25.5.0',
-    'flake8-simplify==0.22.0',
+    'flake8-logging-format==0.9.0',
+    'flake8-pyi==23.3.1',
+    'flake8-simplify==0.19.3',
     'mccabe==0.7.0',
-    'pycodestyle==2.14.0',
-    'pyflakes==3.4.0',
+    'pycodestyle==2.11.1',
+    'pyflakes==3.1.0',
     'torchfix==0.4.0 ; python_version >= "3.9" and python_version < "3.13"',
 ]
 
@@ -158,16 +158,16 @@ init_command = [
     'mypy==1.16.0',
     'sympy==1.13.3',
     'types-requests==2.27.25',
-    'types-pyyaml==6.0.2',
+    'types-pyyaml==6.0.1',
     'types-tabulate==0.8.8',
     'types-protobuf==5.29.1.20250403',
     'types-setuptools==79.0.0.20250422',
     'types-jinja2==2.11.9',
     'types-colorama==0.4.6',
-    'filelock==3.18.0',
+    'filelock==3.13.1',
     'junitparser==2.1.1',
-    'rich==14.1.0',
-    'pyyaml==6.0.2',
+    'rich==10.9.0',
+    'pyyaml==6.0.1',
     'optree==0.13.0',
     'dataclasses-json==0.6.7',
     'pandas==2.2.3',
@@ -1111,7 +1111,7 @@ init_command = [
     'python3',
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
-    'pyyaml==6.0.2',
+    'PyYAML==6.0.1',
 ]
 
 [[linter]]
@@ -1133,7 +1133,7 @@ init_command = [
     'python3',
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
-    'pyyaml==6.0.2',
+    'PyYAML==6.0.1',
 ]
 
 [[linter]]
@@ -1794,12 +1794,3 @@ include_patterns = [
     'torch/header_only_apis.txt',
 ]
 is_formatter = false
-
-
-[[linter]]
-code = "GB_REGISTRY"
-include_patterns = ["torch/_dynamo/**/*.py"]
-command = [
-  "python3",
-  "tools/linter/adapters/gb_registry_linter.py",
-]

BUILD.bazel

@@ -679,7 +679,6 @@ cc_library(
         [
             "torch/*.h",
             "torch/csrc/**/*.h",
-            "torch/nativert/**/*.h",
             "torch/csrc/distributed/c10d/**/*.hpp",
             "torch/lib/libshm/*.h",
         ],

CMakeLists.txt

@@ -564,7 +564,7 @@ if(MSVC)
   set(CMAKE_NINJA_CMCLDEPS_RC OFF)
   if(MSVC_Z7_OVERRIDE)
     # CMake set debug flags to use /Z7
-    set(CMAKE_MSVC_DEBUG_INFORMATION_FORMAT "$<$<CONFIG:Debug,RelWithDebInfo>:Embedded>")
+    set(CMAKE_MSVC_DEBUG_INFORMATION_FORMAT Embedded)
   endif()
   foreach(
     flag_var
@@ -872,14 +872,6 @@ cmake_dependent_option(
   "USE_CUDA OR USE_ROCM;NOT MSVC"
   OFF)
 
-cmake_dependent_option(
-  USE_FBGEMM_GENAI
-  "Whether to build FBGEMM GenAI quantized GEMM kernels.\
-  Will be disabled if not supported by the platform"
-  OFF
-  "USE_CUDA OR USE_ROCM"
-  OFF)
-
 # CAVEAT: Again, Flash Attention2 will error while building for sm52 while Mem
 # Eff Attention won't
 cmake_dependent_option(
@@ -913,10 +905,6 @@ if(USE_FBGEMM)
   string(APPEND CMAKE_CXX_FLAGS " -DUSE_FBGEMM")
 endif()
 
-if(USE_FBGEMM_GENAI)
-  string(APPEND CMAKE_CXX_FLAGS " -DUSE_FBGEMM_GENAI")
-endif()
-
 if(USE_PYTORCH_QNNPACK)
   string(APPEND CMAKE_CXX_FLAGS " -DUSE_PYTORCH_QNNPACK")
 endif()

CODEOWNERS

@@ -51,12 +51,12 @@ nn/qat/ @jerryzh168
 /torch/csrc/distributed/c10d/Ops.* @kwen2501
 
 # ONNX Export
-/torch/_dynamo/backends/onnxrt.py @titaiwangms @xadupre @justinchuby
-/torch/csrc/jit/passes/onnx.h @titaiwangms @xadupre
-/torch/csrc/jit/passes/onnx.cpp @titaiwangms @xadupre
-/torch/csrc/jit/passes/onnx/ @titaiwangms @xadupre
-/torch/onnx/ @titaiwangms @xadupre @justinchuby
-/test/onnx/  @titaiwangms @xadupre @justinchuby
+/torch/_dynamo/backends/onnxrt.py @wschin
+/torch/csrc/jit/passes/onnx.h @titaiwangms @shubhambhokare1
+/torch/csrc/jit/passes/onnx.cpp @titaiwangms @shubhambhokare1
+/torch/csrc/jit/passes/onnx/ @titaiwangms @shubhambhokare1
+/torch/onnx/ @titaiwangms @shubhambhokare1 @justinchuby @wschin
+/test/onnx/  @titaiwangms @shubhambhokare1 @justinchuby @wschin
 
 # CI
 /.ci  @pytorch/pytorch-dev-infra

Dockerfile

@@ -47,6 +47,18 @@ WORKDIR /opt/pytorch
 COPY . .
 RUN git submodule update --init --recursive
 
+FROM conda as build
+ARG CMAKE_VARS
+WORKDIR /opt/pytorch
+COPY --from=conda /opt/conda /opt/conda
+COPY --from=submodule-update /opt/pytorch /opt/pytorch
+RUN make triton
+RUN --mount=type=cache,target=/opt/ccache \
+    export eval ${CMAKE_VARS} && \
+    TORCH_CUDA_ARCH_LIST="7.0 7.2 7.5 8.0 8.6 8.7 8.9 9.0 9.0a" TORCH_NVCC_FLAGS="-Xfatbin -compress-all" \
+    CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" \
+    python -m pip install --no-build-isolation -v .
+
 FROM conda as conda-installs
 ARG PYTHON_VERSION=3.11
 ARG CUDA_PATH=cu121
@@ -97,5 +109,4 @@ WORKDIR /workspace
 
 FROM official as dev
 # Should override the already installed version from the official-image stage
-COPY --from=conda /opt/conda /opt/conda
-COPY --from=submodule-update /opt/pytorch /opt/pytorch
+COPY --from=build /opt/conda /opt/conda

README.md

@@ -294,12 +294,14 @@ Install PyTorch
 
 ```bash
 export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
+python -m pip install -r requirements-build.txt
 python -m pip install --no-build-isolation -v -e .
 ```
 
 **On macOS**
 
 ```bash
+python -m pip install -r requirements-build.txt
 python -m pip install --no-build-isolation -v -e .
 ```
 

aten/src/ATen/CMakeLists.txt

@@ -247,50 +247,6 @@ if(USE_MEM_EFF_ATTENTION)
   list(APPEND ATen_ATTENTION_KERNEL_SRCS ${mem_eff_attention_cuda_kernels_cu})
 endif()
 
-IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-  message(WARNING "Unsupported ROCM arch for FBGEMM GenAI, will set USE_FBGEMM_GENAI to OFF")
-  set(USE_FBGEMM_GENAI off)
-endif()
-
-# FBGEMM GenAI
-IF(USE_FBGEMM_GENAI)
-  set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
-  set(FBGEMM_GENAI_DIR ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
-
-  if(USE_ROCM)
-    # Only include the kernels we want to build to avoid increasing binary size.
-    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-      "${FBGEMM_GENAI_DIR}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-      "${FBGEMM_GENAI_DIR}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
-
-    # Add additional HIPCC compiler flags for performance
-    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-      -mllvm
-      -amdgpu-coerce-illegal-types=1
-      -mllvm
-      -enable-post-misched=0
-      -mllvm
-      -greedy-reverse-local-assignment=1
-      -fhip-new-launch-api)
-
-    hip_add_library(
-      fbgemm_genai STATIC
-      ${fbgemm_genai_native_rocm_hip}
-      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
-    set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
-
-    target_include_directories(fbgemm_genai PUBLIC
-      # FBGEMM version of Composable Kernel is used due to some customizations
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-      ${FBGEMM_GENAI_DIR}/include/
-      ${FBGEMM_GENAI_DIR}/common/include/
-    )
-  endif()
-endif()
-
 # XNNPACK
 file(GLOB native_xnnpack "native/xnnpack/*.cpp")
 
@@ -630,10 +586,17 @@ if(USE_CUDA AND NOT USE_ROCM)
       CUDA::cufft_static_nocallback
     )
    if(NOT BUILD_LAZY_CUDA_LINALG)
-     list(APPEND ATen_CUDA_DEPENDENCY_LIBS
-       CUDA::cusolver_static
-       ${CUDAToolkit_LIBRARY_DIR}/libcusolver_lapack_static.a     # needed for libcusolver_static
-     )
+     if(CUDA_VERSION_MAJOR LESS_EQUAL 11)
+       list(APPEND ATen_CUDA_DEPENDENCY_LIBS
+         CUDA::cusolver_static
+         ${CUDAToolkit_LIBRARY_DIR}/liblapack_static.a     # needed for libcusolver_static
+       )
+     elseif(CUDA_VERSION_MAJOR GREATER_EQUAL 12)
+       list(APPEND ATen_CUDA_DEPENDENCY_LIBS
+         CUDA::cusolver_static
+         ${CUDAToolkit_LIBRARY_DIR}/libcusolver_lapack_static.a     # needed for libcusolver_static
+       )
+     endif()
    endif()
   else()
     list(APPEND ATen_CUDA_DEPENDENCY_LIBS

aten/src/ATen/Context.cpp

@@ -14,9 +14,7 @@
 #include <ATen/cpu/FlushDenormal.h>
 
 #ifdef USE_FBGEMM
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wextra-semi")
 #include <fbgemm/Fbgemm.h>
-C10_DIAGNOSTIC_POP()
 #endif // USE_FBGEMM
 #if defined(__aarch64__) && !defined(C10_MOBILE)
 #include <cpuinfo.h>
@@ -334,14 +332,6 @@ void Context::setBenchmarkLimitCuDNN(int b) {
   benchmark_limit_cudnn = b;
 }
 
-bool Context::immediateMiopen() const {
-  return immediate_miopen;
-}
-
-void Context::setImmediateMiopen(bool b) {
-  immediate_miopen = b;
-}
-
 bool Context::allowTF32CuBLAS() const {
 #ifdef USE_ROCM
     const auto allow_tf32 = c10::utils::check_env(hipblaslt_allow_tf32);
@@ -512,7 +502,7 @@ at::BlasBackend Context::blasPreferredBackend() {
       static const std::vector<std::string> archs = {
           "gfx90a", "gfx942",
 #if ROCM_VERSION >= 60300
-          "gfx1100", "gfx1101", "gfx1200", "gfx1201", "gfx908",
+          "gfx1100", "gfx1101", "gfx1200", "gfx1201",
 #endif
 #if ROCM_VERSION >= 60500
           "gfx950"

aten/src/ATen/Context.h

@@ -205,8 +205,6 @@ class TORCH_API Context {
   void setBenchmarkCuDNN(bool);
   int benchmarkLimitCuDNN() const;
   void setBenchmarkLimitCuDNN(int);
-  bool immediateMiopen() const;
-  void setImmediateMiopen(bool);
   bool deterministicCuDNN() const;
   void setDeterministicCuDNN(bool);
   bool deterministicMkldnn() const;
@@ -442,7 +440,6 @@ class TORCH_API Context {
   bool enabled_overrideable = true;
   bool allow_fp16_bf16_reduction_mathSDP = false;
   bool benchmark_cudnn = false;
-  bool immediate_miopen = false;
   Float32MatmulPrecision float32_matmul_precision =
       c10::utils::check_env("TORCH_ALLOW_TF32_CUBLAS_OVERRIDE") == true
       ? at::Float32MatmulPrecision::HIGH

aten/src/ATen/DLConvertor.cpp

@@ -132,9 +132,6 @@ DLDevice torchDeviceToDLDevice(at::Device device) {
     case DeviceType::PrivateUse1:
       ctx.device_type = DLDeviceType::kDLExtDev;
       break;
-    case DeviceType::MPS:
-      ctx.device_type = DLDeviceType::kDLMetal;
-      break;
     default:
       TORCH_CHECK_BUFFER(false, "Cannot pack tensors on " + device.str());
   }
@@ -167,8 +164,6 @@ static Device getATenDevice(DLDeviceType type, c10::DeviceIndex index, void* dat
       return at::Device(DeviceType::MAIA, index);
     case DLDeviceType::kDLExtDev:
       return at::Device(DeviceType::PrivateUse1, index);
-    case DLDeviceType::kDLMetal:
-      return at::Device(DeviceType::MPS, index);
     default:
       TORCH_CHECK_BUFFER(
           false, "Unsupported device_type: ", std::to_string(type));

aten/src/ATen/DeviceAccelerator.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <c10/core/CachingDeviceAllocator.h>
 #include <c10/core/DeviceType.h>
 #include <c10/macros/Macros.h>
 
@@ -72,6 +73,27 @@ TORCH_API c10::DeviceIndex exchangeDevice(c10::DeviceIndex device_index);
 // original device index that was active before the change.
 TORCH_API c10::DeviceIndex maybeExchangeDevice(c10::DeviceIndex device_index);
 
+TORCH_API inline void emptyCache() {
+  const auto device_type = getAccelerator(true).value();
+  at::getDeviceAllocator(device_type)->emptyCache();
+}
+
+TORCH_API inline at::CachingDeviceAllocator::DeviceStats getDeviceStats(
+    c10::DeviceIndex device_index) {
+  const auto device_type = getAccelerator(true).value();
+  return at::getDeviceAllocator(device_type)->getDeviceStats(device_index);
+}
+
+TORCH_API inline void resetAccumulatedStats(c10::DeviceIndex device_index) {
+  const auto device_type = getAccelerator(true).value();
+  at::getDeviceAllocator(device_type)->resetAccumulatedStats(device_index);
+}
+
+TORCH_API inline void resetPeakStats(c10::DeviceIndex device_index) {
+  const auto device_type = getAccelerator(true).value();
+  at::getDeviceAllocator(device_type)->resetPeakStats(device_index);
+}
+
 } // namespace at::accelerator
 
 namespace at {

aten/src/ATen/ZeroTensorFallback.cpp

@@ -9,36 +9,7 @@
 
 namespace at {
 
- /*
-  * Design:
-  * 1. ZeroTensors are regular tensors with TensorOptions, a storage
-  *    pointing to nullptr and a ZeroTensor dispatch key set.
-  *
-  * 2. ZeroTensors are immutable. This is done to prevent data race in the case of multithreading
-  *    (when two threads try to read the same zero tensor and materialize it in-place).
-  *
-  * 3. ZeroTensor has a boxed fallback that will be dispatched to any ops that don't
-  *    have special ZeroTensor handling. This fallback materializes each ZeroTensor to
-  *    `at::zeros({}, tensor.options()).expand(tensor.sizes())`.
-
-  * 4. ZeroTensors are handled above autograd. This is necessary because fallback
-  *    operations are not differentiable.
-  *     - Example: Consider add in the case it was using the fallback: zerotensor_a + b.
-  *       zerotensor_a would be materialized to c=torch.zeros_like(zerotensor_a) after
-  *       passing through the fallback. If this happens above the autograd, then the
-  *       gradients would be populated on c instead of zerotensor_a.
-  *
-  * 5. The grad field is always populated with an honest to goodness tensor. This
-  *    materialization of ZeroTensors will happen in:
-  *     - AccumulateGrad for Backward Mode AD.
-  *     - will never be required for ForwardMode AD.
-  *       - This is because if all the tangents were undefined (efficient ZeroTensors),
-  *         no computation will be performed (this is ensured via an existing pre-check).
-  *
-  * Today ZeroTensors are primarily used to represent undefined gradients in forward AD,
-  * it does not perfectly handle NaNs and Infs as we don't check the actual values
-  * and assume that they are non-zero, non-inf, non-NaN etc.
-  */
+  // TODO: add a note explaining the design decisions
   // ZeroTensors are designed to be immutable. Thus, we error out when an in-place operation is performed on ZeroTensors
   static void zeroTensorFallback(const c10::OperatorHandle& op, DispatchKeySet dispatch_keys, torch::jit::Stack* stack) {
     const auto& arguments = op.schema().arguments();

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

@@ -1 +1,55 @@
-#include <torch/headeronly/cpu/vec/intrinsics.h>
+#pragma once
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
+/* GCC or clang-compatible compiler, targeting x86/x86-64 */
+#include <x86intrin.h>
+#elif defined(__clang__) && (defined(__ARM_NEON__) || defined(__aarch64__))
+/* Clang-compatible compiler, targeting arm neon */
+#include <arm_neon.h>
+#if defined(__ARM_FEATURE_SVE)
+/* CLANG-compatible compiler, targeting ARM with SVE */
+#include <arm_sve.h>
+#endif
+#elif defined(_MSC_VER)
+/* Microsoft C/C++-compatible compiler */
+#include <intrin.h>
+#if _MSC_VER <= 1900
+#define _mm256_extract_epi64(X, Y) \
+  (_mm_extract_epi64(_mm256_extractf128_si256(X, Y >> 1), Y % 2))
+#define _mm256_extract_epi32(X, Y) \
+  (_mm_extract_epi32(_mm256_extractf128_si256(X, Y >> 2), Y % 4))
+#define _mm256_extract_epi16(X, Y) \
+  (_mm_extract_epi16(_mm256_extractf128_si256(X, Y >> 3), Y % 8))
+#define _mm256_extract_epi8(X, Y) \
+  (_mm_extract_epi8(_mm256_extractf128_si256(X, Y >> 4), Y % 16))
+#endif
+#elif defined(__GNUC__) && (defined(__ARM_NEON__) || defined(__aarch64__))
+/* GCC-compatible compiler, targeting ARM with NEON */
+#include <arm_neon.h>
+#if defined(__ARM_FEATURE_SVE)
+/* GCC-compatible compiler, targeting ARM with SVE */
+#include <arm_sve.h>
+#endif
+#if defined(MISSING_ARM_VLD1)
+#include <ATen/cpu/vec/vec256/missing_vld1_neon.h>
+#elif defined(MISSING_ARM_VST1)
+#include <ATen/cpu/vec/vec256/missing_vst1_neon.h>
+#endif
+#elif defined(__GNUC__) && defined(__IWMMXT__)
+/* GCC-compatible compiler, targeting ARM with WMMX */
+#include <mmintrin.h>
+#elif defined(__s390x__)
+// targets Z/architecture
+// we will include vecintrin later
+#elif (defined(__GNUC__) || defined(__xlC__)) && \
+    (defined(__VEC__) || defined(__ALTIVEC__))
+/* XLC or GCC-compatible compiler, targeting PowerPC with VMX/VSX */
+#include <altivec.h>
+/* We need to undef those tokens defined by <altivec.h> to avoid conflicts
+   with the C++ types. => Can still use __bool/__vector */
+#undef bool
+#undef vector
+#undef pixel
+#elif defined(__GNUC__) && defined(__SPE__)
+/* GCC-compatible compiler, targeting PowerPC with SPE */
+#include <spe.h>
+#endif

aten/src/ATen/cpu/vec/sve/vec_bfloat16.h

@@ -5,7 +5,6 @@
 #include <ATen/cpu/vec/sve/vec_common_sve.h>
 #include <ATen/cpu/vec/sve/vec_float.h>
 #include <ATen/cpu/vec/vec_base.h>
-#include <c10/util/bit_cast.h>
 #include <cmath>
 namespace at {
 namespace vec {
@@ -37,7 +36,7 @@ class Vectorized<BFloat16> {
     return VECTOR_WIDTH / sizeof(BFloat16);
   }
 
-  Vectorized();
+  Vectorized() {}
   Vectorized(svbfloat16_t v) : values(v) {}
   Vectorized(int val);
   Vectorized(BFloat16 val);
@@ -307,11 +306,6 @@ Vectorized<c10::BFloat16> inline operator/(
   return binary_operator_via_float(std::divides<Vectorized<float>>(), a, b);
 }
 
-inline Vectorized<BFloat16>::Vectorized() {
-  const short zero = 0;
-  values = svdup_n_bf16(c10::bit_cast<bfloat16_t>(zero));
-}
-
 inline Vectorized<BFloat16>::Vectorized(int val) {
   auto vals_f = svdup_n_f32(val);
   values = convert_float_bfloat16(vals_f, vals_f);

aten/src/ATen/cpu/vec/sve/vec_double.h

@@ -38,9 +38,7 @@ class Vectorized<double> {
   static constexpr size_type size() {
     return VECTOR_WIDTH / sizeof(double);
   }
-  Vectorized() {
-    values = svdup_n_f64(0);
-  }
+  Vectorized() {}
   Vectorized(svfloat64_t v) : values(v) {}
   Vectorized(double val) {
     values = svdup_n_f64(val);
@@ -587,30 +585,6 @@ Vectorized<double> inline fmadd(
   return svmad_f64_x(ptrue, a, b, c);
 }
 
-template <>
-Vectorized<double> inline fnmadd(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return svmsb_f64_x(ptrue, a, b, c);
-}
-
-template <>
-Vectorized<double> inline fmsub(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return svnmsb_f64_x(ptrue, a, b, c);
-}
-
-template <>
-Vectorized<double> inline fnmsub(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return svnmad_f64_x(ptrue, a, b, c);
-}
-
 #endif // defined(CPU_CAPABILITY_SVE)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/sve/vec_float.h

@@ -38,9 +38,7 @@ class Vectorized<float> {
   static constexpr size_type size() {
     return VECTOR_WIDTH / sizeof(float);
   }
-  Vectorized() {
-    values = svdup_n_f32(0);
-  }
+  Vectorized() {}
   Vectorized(svfloat32_t v) : values(v) {}
   Vectorized(float val) {
     values = svdup_n_f32(val);
@@ -758,30 +756,6 @@ Vectorized<float> inline fmadd(
   return svmad_f32_x(ptrue, a, b, c);
 }
 
-template <>
-Vectorized<float> inline fnmadd(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return svmsb_f32_x(ptrue, a, b, c);
-}
-
-template <>
-Vectorized<float> inline fmsub(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return svnmsb_f32_x(ptrue, a, b, c);
-}
-
-template <>
-Vectorized<float> inline fnmsub(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return svnmad_f32_x(ptrue, a, b, c);
-}
-
 #endif // defined(CPU_CAPABILITY_SVE)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/sve/vec_int.h

@@ -32,9 +32,7 @@ inline namespace CPU_CAPABILITY {
     static constexpr size_type size() {                                       \
       return vl;                                                              \
     }                                                                         \
-    Vectorized() {                                                            \
-      values = svdup_n_s##bit(0);                                             \
-    }                                                                         \
+    Vectorized() {}                                                           \
     Vectorized(svint##bit##_t v) : values(v) {}                               \
     Vectorized(int##bit##_t val) {                                            \
       values = svdup_n_s##bit(val);                                           \

aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h

@@ -552,15 +552,6 @@ Vectorized<c10::BFloat16> inline fmadd(
   return a * b + c;
 }
 
-template <>
-Vectorized<c10::BFloat16> inline fnmadd(
-    const Vectorized<c10::BFloat16>& a,
-    const Vectorized<c10::BFloat16>& b,
-    const Vectorized<c10::BFloat16>& c) {
-  // See NOTE [BF16 FMA] above.
-  return -a * b + c;
-}
-
 template <>
 Vectorized<c10::BFloat16> inline fmsub(
     const Vectorized<c10::BFloat16>& a,
@@ -570,15 +561,6 @@ Vectorized<c10::BFloat16> inline fmsub(
   return a * b - c;
 }
 
-template <>
-Vectorized<c10::BFloat16> inline fnmsub(
-    const Vectorized<c10::BFloat16>& a,
-    const Vectorized<c10::BFloat16>& b,
-    const Vectorized<c10::BFloat16>& c) {
-  // See NOTE [BF16 FMA] above.
-  return -a * b - c;
-}
-
 #endif // !defined(C10_MOBILE) && defined(__aarch64__)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h

@@ -83,9 +83,7 @@ class Vectorized<float> {
   static constexpr size_type size() {
     return 4;
   }
-  Vectorized() {
-    values = vmovq_n_f32(0);
-  }
+  Vectorized() {}
   Vectorized(float32x4_t v) : values(v) {}
   Vectorized(float val) : values{vdupq_n_f32(val)} {}
   Vectorized(float val0, float val1, float val2, float val3)
@@ -584,14 +582,6 @@ Vectorized<float> inline fmadd(
   return Vectorized<float>(vfmaq_f32(c, a, b));
 }
 
-template <>
-Vectorized<float> inline fnmadd(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return Vectorized<float>(vfmsq_f32(c, a, b));
-}
-
 template <>
 Vectorized<float> inline fmsub(
     const Vectorized<float>& a,
@@ -600,14 +590,6 @@ Vectorized<float> inline fmsub(
   return Vectorized<float>(vnegq_f32(vfmsq_f32(c, a, b)));
 }
 
-template <>
-Vectorized<float> inline fnmsub(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return Vectorized<float>(vnegq_f32(vfmaq_f32(c, a, b)));
-}
-
 inline Vectorized<float> Vectorized<float>::erf() const {
   // constants
   const Vectorized<float> neg_zero_vec(-0.f);

aten/src/ATen/cpu/vec/vec128/vec128_half_neon.h

@@ -621,18 +621,6 @@ Vectorized<c10::Half> inline fmadd(
 #endif
 }
 
-template <>
-Vectorized<c10::Half> inline fnmadd(
-    const Vectorized<c10::Half>& a,
-    const Vectorized<c10::Half>& b,
-    const Vectorized<c10::Half>& c) {
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-  return Vectorized<c10::Half>(vfmsq_f16(c, a, b));
-#else
-  return -a * b + c;
-#endif
-}
-
 template <>
 Vectorized<c10::Half> inline fmsub(
     const Vectorized<c10::Half>& a,
@@ -644,18 +632,6 @@ Vectorized<c10::Half> inline fmsub(
   return a * b - c;
 #endif
 }
-
-template <>
-Vectorized<c10::Half> inline fnmsub(
-    const Vectorized<c10::Half>& a,
-    const Vectorized<c10::Half>& b,
-    const Vectorized<c10::Half>& c) {
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-  return Vectorized<c10::Half>(vnegq_f16(vfmaq_f16(c, a, b)));
-#else
-  return -a * b - c;
-#endif
-}
 #endif // !defined(C10_MOBILE) && defined(__aarch64__)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/vec256/missing_vld1_neon.h

@@ -1 +1,396 @@
-#include <torch/headeronly/cpu/vec/vec256/missing_vld1_neon.h>
+/* Workaround for missing vld1_*_x2 and vst1_*_x2 intrinsics in gcc-7.  */
+
+__extension__ extern __inline uint8x8x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_u8_x2(const uint8_t* __a) {
+  uint8x8x2_t ret;
+  asm volatile("ld1 {%S0.8b - %T0.8b}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int8x8x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_s8_x2(const int8_t* __a) {
+  int8x8x2_t ret;
+  asm volatile("ld1 {%S0.8b - %T0.8b}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline uint16x4x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_u16_x2(const uint16_t* __a) {
+  uint16x4x2_t ret;
+  asm volatile("ld1 {%S0.4h - %T0.4h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int16x4x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_s16_x2(const int16_t* __a) {
+  int16x4x2_t ret;
+  asm volatile("ld1 {%S0.4h - %T0.4h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline uint32x2x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_u32_x2(const uint32_t* __a) {
+  uint32x2x2_t ret;
+  asm volatile("ld1 {%S0.2s - %T0.2s}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int32x2x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_s32_x2(const int32_t* __a) {
+  int32x2x2_t ret;
+  asm volatile("ld1 {%S0.2s - %T0.2s}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline uint64x1x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_u64_x2(const uint64_t* __a) {
+  uint64x1x2_t ret;
+  asm volatile("ld1 {%S0.1d - %T0.1d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int64x1x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_s64_x2(const int64_t* __a) {
+  int64x1x2_t ret;
+  __builtin_aarch64_simd_oi __o;
+  asm volatile("ld1 {%S0.1d - %T0.1d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline float16x4x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_f16_x2(const float16_t* __a) {
+  float16x4x2_t ret;
+  asm volatile("ld1 {%S0.4h - %T0.4h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline float32x2x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_f32_x2(const float32_t* __a) {
+  float32x2x2_t ret;
+  asm volatile("ld1 {%S0.2s - %T0.2s}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline float64x1x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_f64_x2(const float64_t* __a) {
+  float64x1x2_t ret;
+  asm volatile("ld1 {%S0.1d - %T0.1d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline poly8x8x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_p8_x2(const poly8_t* __a) {
+  poly8x8x2_t ret;
+  asm volatile("ld1 {%S0.8b - %T0.8b}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline poly16x4x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_p16_x2(const poly16_t* __a) {
+  poly16x4x2_t ret;
+  asm volatile("ld1 {%S0.4h - %T0.4h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline poly64x1x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1_p64_x2(const poly64_t* __a) {
+  poly64x1x2_t ret;
+  asm volatile("ld1 {%S0.1d - %T0.1d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline uint8x16x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_u8_x2(const uint8_t* __a) {
+  uint8x16x2_t ret;
+  asm volatile("ld1 {%S0.16b - %T0.16b}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int8x16x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_s8_x2(const int8_t* __a) {
+  int8x16x2_t ret;
+  asm volatile("ld1 {%S0.16b - %T0.16b}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline uint16x8x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_u16_x2(const uint16_t* __a) {
+  uint16x8x2_t ret;
+  asm volatile("ld1 {%S0.8h - %T0.8h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int16x8x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_s16_x2(const int16_t* __a) {
+  int16x8x2_t ret;
+  asm volatile("ld1 {%S0.8h - %T0.8h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline uint32x4x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_u32_x2(const uint32_t* __a) {
+  uint32x4x2_t ret;
+  asm volatile("ld1 {%S0.4s - %T0.4s}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int32x4x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_s32_x2(const int32_t* __a) {
+  int32x4x2_t ret;
+  asm volatile("ld1 {%S0.4s - %T0.4s}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline uint64x2x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_u64_x2(const uint64_t* __a) {
+  uint64x2x2_t ret;
+  asm volatile("ld1 {%S0.2d - %T0.2d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline int64x2x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_s64_x2(const int64_t* __a) {
+  int64x2x2_t ret;
+  asm volatile("ld1 {%S0.2d - %T0.2d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline float16x8x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_f16_x2(const float16_t* __a) {
+  float16x8x2_t ret;
+  asm volatile("ld1 {%S0.8h - %T0.8h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline float32x4x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_f32_x2(const float32_t* __a) {
+  float32x4x2_t ret;
+  asm volatile("ld1 {%S0.4s - %T0.4s}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline float64x2x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_f64_x2(const float64_t* __a) {
+  float64x2x2_t ret;
+  asm volatile("ld1 {%S0.2d - %T0.2d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline poly8x16x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_p8_x2(const poly8_t* __a) {
+  poly8x16x2_t ret;
+  asm volatile("ld1 {%S0.16b - %T0.16b}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline poly16x8x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_p16_x2(const poly16_t* __a) {
+  poly16x8x2_t ret;
+  asm volatile("ld1 {%S0.8h - %T0.8h}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+__extension__ extern __inline poly64x2x2_t
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_p64_x2(const poly64_t* __a) {
+  poly64x2x2_t ret;
+  asm volatile("ld1 {%S0.2d - %T0.2d}, %1" : "=w"(ret) : "Q"(*__a));
+  return ret;
+}
+
+/* vst1x2 */
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_s64_x2(int64_t* __a, int64x1x2_t val) {
+  asm volatile("st1 {%S1.1d - %T1.1d}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_u64_x2(uint64_t* __a, uint64x1x2_t val) {
+  asm volatile("st1 {%S1.1d - %T1.1d}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_f64_x2(float64_t* __a, float64x1x2_t val) {
+  asm volatile("st1 {%S1.1d - %T1.1d}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_s8_x2(int8_t* __a, int8x8x2_t val) {
+  asm volatile("st1 {%S1.8b - %T1.8b}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_p8_x2(poly8_t* __a, poly8x8x2_t val) {
+  asm volatile("st1 {%S1.8b - %T1.8b}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_s16_x2(int16_t* __a, int16x4x2_t val) {
+  asm volatile("st1 {%S1.4h - %T1.4h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_p16_x2(poly16_t* __a, poly16x4x2_t val) {
+  asm volatile("st1 {%S1.4h - %T1.4h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_s32_x2(int32_t* __a, int32x2x2_t val) {
+  asm volatile("st1 {%S1.2s - %T1.2s}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_u8_x2(uint8_t* __a, uint8x8x2_t val) {
+  asm volatile("st1 {%S1.8b - %T1.8b}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_u16_x2(uint16_t* __a, uint16x4x2_t val) {
+  asm volatile("st1 {%S1.4h - %T1.4h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_u32_x2(uint32_t* __a, uint32x2x2_t val) {
+  asm volatile("st1 {%S1.2s - %T1.2s}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_f16_x2(float16_t* __a, float16x4x2_t val) {
+  asm volatile("st1 {%S1.4h - %T1.4h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_f32_x2(float32_t* __a, float32x2x2_t val) {
+  asm volatile("st1 {%S1.2s - %T1.2s}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1_p64_x2(poly64_t* __a, poly64x1x2_t val) {
+  asm volatile("st1 {%S1.1d - %T1.1d}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_s8_x2(int8_t* __a, int8x16x2_t val) {
+  asm volatile("st1 {%S1.16b - %T1.16b}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_p8_x2(poly8_t* __a, poly8x16x2_t val) {
+  asm volatile("st1 {%S1.16b - %T1.16b}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_s16_x2(int16_t* __a, int16x8x2_t val) {
+  asm volatile("st1 {%S1.8h - %T1.8h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_p16_x2(poly16_t* __a, poly16x8x2_t val) {
+  asm volatile("st1 {%S1.8h - %T1.8h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_s32_x2(int32_t* __a, int32x4x2_t val) {
+  asm volatile("st1 {%S1.4s - %T1.4s}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_s64_x2(int64_t* __a, int64x2x2_t val) {
+  asm volatile("st1 {%S1.2d - %T1.2d}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_u8_x2(uint8_t* __a, uint8x16x2_t val) {
+  asm volatile("st1 {%S1.16b - %T1.16b}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_u16_x2(uint16_t* __a, uint16x8x2_t val) {
+  asm volatile("st1 {%S1.8h - %T1.8h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_u32_x2(uint32_t* __a, uint32x4x2_t val) {
+  asm volatile("st1 {%S1.4s - %T1.4s}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_u64_x2(uint64_t* __a, uint64x2x2_t val) {
+  asm volatile("st1 {%S1.2d - %T1.2d}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_f16_x2(float16_t* __a, float16x8x2_t val) {
+  asm volatile("st1 {%S1.8h - %T1.8h}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_f32_x2(float32_t* __a, float32x4x2_t val) {
+  asm volatile("st1 {%S1.4s - %T1.4s}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_f64_x2(float64_t* __a, float64x2x2_t val) {
+  asm volatile("st1 {%S1.2d - %T1.2d}, %0" : "=Q"(*__a) : "w"(val));
+}
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_p64_x2(poly64_t* __a, poly64x2x2_t val) {
+  asm volatile("st1 {%S1.2d - %T1.2d}, %0" : "=Q"(*__a) : "w"(val));
+}

aten/src/ATen/cpu/vec/vec256/missing_vst1_neon.h

@@ -1 +1,7 @@
-#include <torch/headeronly/cpu/vec/vec256/missing_vst1_neon.h>
+/* Workaround for missing vst1q_f32_x2 in gcc-8.  */
+
+__extension__ extern __inline void
+    __attribute__((__always_inline__, __gnu_inline__, __artificial__))
+    vst1q_f32_x2(float32_t* __a, float32x4x2_t val) {
+  asm volatile("st1 {%S1.4s - %T1.4s}, %0" : "=Q"(*__a) : "w"(val));
+}

aten/src/ATen/cpu/vec/vec256/vec256_complex_double.h

@@ -34,9 +34,7 @@ class Vectorized<c10::complex<double>> {
   static constexpr size_type size() {
     return 2;
   }
-  Vectorized() {
-    values = _mm256_setzero_pd();
-  }
+  Vectorized() {}
   Vectorized(__m256d v) : values(v) {}
   Vectorized(c10::complex<double> val) {
     double real_value = val.real();

aten/src/ATen/cpu/vec/vec256/vec256_complex_float.h

@@ -33,9 +33,7 @@ class Vectorized<c10::complex<float>> {
   static constexpr size_type size() {
     return 4;
   }
-  Vectorized() {
-    values = _mm256_setzero_ps();
-  }
+  Vectorized() {}
   Vectorized(__m256 v) : values(v) {}
   Vectorized(c10::complex<float> val) {
     float real_value = val.real();

aten/src/ATen/cpu/vec/vec256/vec256_double.h

@@ -31,9 +31,7 @@ class Vectorized<double> {
   static constexpr size_type size() {
     return 4;
   }
-  Vectorized() {
-    values = _mm256_setzero_pd();
-  }
+  Vectorized() {}
   Vectorized(__m256d v) : values(v) {}
   Vectorized(double val) {
     values = _mm256_set1_pd(val);
@@ -495,14 +493,6 @@ Vectorized<double> inline fmadd(
   return _mm256_fmadd_pd(a, b, c);
 }
 
-template <>
-Vectorized<double> inline fnmadd(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return _mm256_fnmadd_pd(a, b, c);
-}
-
 template <>
 Vectorized<double> inline fmsub(
     const Vectorized<double>& a,
@@ -510,14 +500,6 @@ Vectorized<double> inline fmsub(
     const Vectorized<double>& c) {
   return _mm256_fmsub_pd(a, b, c);
 }
-
-template <>
-Vectorized<double> inline fnmsub(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return _mm256_fnmsub_pd(a, b, c);
-}
 #endif
 
 #endif

aten/src/ATen/cpu/vec/vec256/vec256_float.h

@@ -30,9 +30,7 @@ class Vectorized<float> {
   static constexpr size_type size() {
     return 8;
   }
-  Vectorized() {
-    values = _mm256_setzero_ps();
-  }
+  Vectorized() {}
   Vectorized(__m256 v) : values(v) {}
   Vectorized(float val) {
     values = _mm256_set1_ps(val);
@@ -696,14 +694,6 @@ Vectorized<float> inline fmadd(
   return _mm256_fmadd_ps(a, b, c);
 }
 
-template <>
-Vectorized<float> inline fnmadd(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return _mm256_fnmadd_ps(a, b, c);
-}
-
 template <>
 Vectorized<float> inline fmsub(
     const Vectorized<float>& a,
@@ -712,14 +702,6 @@ Vectorized<float> inline fmsub(
   return _mm256_fmsub_ps(a, b, c);
 }
 
-template <>
-Vectorized<float> inline fnmsub(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return _mm256_fnmsub_ps(a, b, c);
-}
-
 // TODO: rewrite with ATEN vectorized (need to add unpack and shuffle)
 // Used by Inductor CPP codegen for micro gemm
 inline void transpose_block(at::vec::VectorizedN<float, 8>& input) {

aten/src/ATen/cpu/vec/vec256/vec256_int.h

@@ -23,9 +23,7 @@ struct Vectorizedi {
   }
 
  public:
-  Vectorizedi() {
-    values = _mm256_setzero_si256();
-  }
+  Vectorizedi() {}
   Vectorizedi(__m256i v) : values(v) {}
   operator __m256i() const {
     return values;
@@ -55,9 +53,7 @@ class Vectorized<int64_t> : public Vectorizedi {
     return 4;
   }
   using Vectorizedi::Vectorizedi;
-  Vectorized() {
-    values = _mm256_setzero_si256();
-  }
+  Vectorized() {}
   Vectorized(int64_t v) {
     values = _mm256_set1_epi64x(v);
   }

aten/src/ATen/cpu/vec/vec256/vec256_qint.h

@@ -54,9 +54,7 @@ struct Vectorizedqi {
 #endif
 
  public:
-  Vectorizedqi() {
-    vals = _mm256_setzero_si256();
-  }
+  Vectorizedqi() {}
   Vectorizedqi(__m256i v) : vals(v) {}
   operator __m256i() const {
     return vals;

aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h

@@ -192,9 +192,7 @@ class Vectorized16 {
   static constexpr size_type size() {
     return 32;
   }
-  Vectorized16() {
-    values = _mm512_setzero_si512();
-  }
+  Vectorized16() {}
   Vectorized16(__m512i v) : values(v) {}
   Vectorized16(T val) {
     value_type uw = val.x;

aten/src/ATen/cpu/vec/vec512/vec512_complex_double.h

@@ -34,9 +34,7 @@ class Vectorized<c10::complex<double>> {
   static constexpr size_type size() {
     return 4;
   }
-  Vectorized() {
-    values = _mm512_setzero_pd();
-  }
+  Vectorized() {}
   Vectorized(__m512d v) : values(v) {}
   Vectorized(c10::complex<double> val) {
     double real_value = val.real();

aten/src/ATen/cpu/vec/vec512/vec512_complex_float.h

@@ -34,9 +34,7 @@ class Vectorized<c10::complex<float>> {
   static constexpr size_type size() {
     return 8;
   }
-  Vectorized() {
-    values = _mm512_setzero_ps();
-  }
+  Vectorized() {}
   Vectorized(__m512 v) : values(v) {}
   Vectorized(c10::complex<float> val) {
     float real_value = val.real();

aten/src/ATen/cpu/vec/vec512/vec512_double.h

@@ -34,9 +34,7 @@ class Vectorized<double> {
   static constexpr size_type size() {
     return 8;
   }
-  Vectorized() {
-    values = _mm512_setzero_pd();
-  }
+  Vectorized() {}
   Vectorized(__m512d v) : values(v) {}
   Vectorized(double val) {
     values = _mm512_set1_pd(val);
@@ -536,14 +534,6 @@ Vectorized<double> inline fmadd(
   return _mm512_fmadd_pd(a, b, c);
 }
 
-template <>
-Vectorized<double> inline fnmadd(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return _mm512_fnmadd_pd(a, b, c);
-}
-
 template <>
 Vectorized<double> inline fmsub(
     const Vectorized<double>& a,
@@ -552,14 +542,6 @@ Vectorized<double> inline fmsub(
   return _mm512_fmsub_pd(a, b, c);
 }
 
-template <>
-Vectorized<double> inline fnmsub(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return _mm512_fnmsub_pd(a, b, c);
-}
-
 #endif
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/vec512/vec512_float.h

@@ -32,9 +32,7 @@ class Vectorized<float> {
   static constexpr size_type size() {
     return 16;
   }
-  Vectorized() {
-    values = _mm512_setzero_ps();
-  }
+  Vectorized() {}
   Vectorized(__m512 v) : values(v) {}
   Vectorized(float val) {
     values = _mm512_set1_ps(val);
@@ -749,14 +747,6 @@ Vectorized<float> inline fmadd(
   return _mm512_fmadd_ps(a, b, c);
 }
 
-template <>
-Vectorized<float> inline fnmadd(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return _mm512_fnmadd_ps(a, b, c);
-}
-
 template <>
 Vectorized<float> inline fmsub(
     const Vectorized<float>& a,
@@ -765,14 +755,6 @@ Vectorized<float> inline fmsub(
   return _mm512_fmsub_ps(a, b, c);
 }
 
-template <>
-Vectorized<float> inline fnmsub(
-    const Vectorized<float>& a,
-    const Vectorized<float>& b,
-    const Vectorized<float>& c) {
-  return _mm512_fnmsub_ps(a, b, c);
-}
-
 // TODO: rewrite with ATEN vectorized (need to add unpack and shuffle)
 // Used by Inductor CPP codegen for micro gemm
 // Code referred to FBGEMM:

aten/src/ATen/cpu/vec/vec512/vec512_int.h

@@ -53,9 +53,7 @@ class Vectorized<int64_t> : public Vectorizedi {
     return 8;
   }
   using Vectorizedi::Vectorizedi;
-  Vectorized() {
-    values = _mm512_setzero_si512();
-  }
+  Vectorized() {}
   Vectorized(int64_t v) {
     values = _mm512_set1_epi64(v);
   }

aten/src/ATen/cpu/vec/vec512/vec512_qint.h

@@ -55,9 +55,7 @@ struct Vectorizedqi {
 #endif
 
  public:
-  Vectorizedqi() {
-    vals = _mm512_setzero_si512();
-  }
+  Vectorizedqi() {}
   Vectorizedqi(__m512i v) : vals(v) {}
   operator __m512i() const {
     return vals;