Install theme in push scripts

.bazelrc

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

.ci/aarch64_linux/build_aarch64_wheel.py

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

.ci/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
@@ -144,18 +149,56 @@ case "$tag" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=11
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
     VISION=yes
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.13
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -176,7 +219,31 @@ case "$tag" in
     VISION=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-jammy-rocm-n-py3-benchmarks | pytorch-linux-noble-rocm-n-py3)
+  pytorch-linux-jammy-py3.11-clang12)
+    ANACONDA_PYTHON_VERSION=3.11
+    CLANG_VERSION=12
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3.9-gcc9)
+    ANACONDA_PYTHON_VERSION=3.9
+    GCC_VERSION=9
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-rocm-n-1-py3)
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=11
+    VISION=yes
+    ROCM_VERSION=6.3
+    NINJA_VERSION=1.9.0
+    TRITON=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-noble-rocm-n-py3)
     if [[ $tag =~ "jammy" ]]; then
       ANACONDA_PYTHON_VERSION=3.10
     else
@@ -190,21 +257,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    if [[ $tag =~ "benchmarks" ]]; then
-      INDUCTOR_BENCHMARKS=yes
-    fi
-    ;;
-  pytorch-linux-noble-rocm-alpha-py3)
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=11
-    VISION=yes
-    ROCM_VERSION=7.0
-    NINJA_VERSION=1.9.0
-    TRITON=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
+    INDUCTOR_BENCHMARKS=yes
     ;;
   pytorch-linux-jammy-xpu-2025.0-py3)
     ANACONDA_PYTHON_VERSION=3.9
@@ -222,7 +275,7 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
+    pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
@@ -234,6 +287,7 @@ case "$tag" in
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
     ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
@@ -312,6 +366,7 @@ case "$tag" in
     fi
     if [[ "$image" == *cuda* ]]; then
       extract_version_from_image_name cuda CUDA_VERSION
+      extract_version_from_image_name cudnn CUDNN_VERSION
     fi
     if [[ "$image" == *rocm* ]]; then
       extract_version_from_image_name rocm ROCM_VERSION
@@ -363,6 +418,9 @@ 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/ci_commit_pins/triton.txt

@@ -1 +1 @@
-f7888497a1eb9e98d4c07537f0d0bcfe180d1363
+ae848267bebc65c6181e8cc5e64a6357d2679260

.ci/docker/common/install_conda.sh

@@ -4,8 +4,12 @@ set -ex
 
 # Optionally install conda
 if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
-  BASE_URL="https://github.com/conda-forge/miniforge/releases/latest/download"  # @lint-ignore
-  CONDA_FILE="Miniforge3-Linux-$(uname -m).sh"
+  BASE_URL="https://repo.anaconda.com/miniconda"
+  CONDA_FILE="Miniconda3-latest-Linux-x86_64.sh"
+  if [[ $(uname -m) == "aarch64" ]] || [[ "$BUILD_ENVIRONMENT" == *xpu* ]] || [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
+    BASE_URL="https://github.com/conda-forge/miniforge/releases/latest/download"  # @lint-ignore
+    CONDA_FILE="Miniforge3-Linux-$(uname -m).sh"
+  fi
 
   MAJOR_PYTHON_VERSION=$(echo "$ANACONDA_PYTHON_VERSION" | cut -d . -f 1)
   MINOR_PYTHON_VERSION=$(echo "$ANACONDA_PYTHON_VERSION" | cut -d . -f 2)
@@ -17,6 +21,7 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
       exit 1
       ;;
   esac
+
   mkdir -p /opt/conda
   chown jenkins:jenkins /opt/conda
 

.ci/docker/common/install_cpython.sh

@@ -66,9 +66,8 @@ function do_cpython_build {
         ln -s pip3 ${prefix}/bin/pip
     fi
     # install setuptools since python 3.12 is required to use distutils
-    # packaging is needed to create symlink since wheel no longer provides needed information
-    ${prefix}/bin/pip install packaging==25.0 wheel==0.45.1 setuptools==80.9.0
-    local abi_tag=$(${prefix}/bin/python -c "from packaging.tags import interpreter_name, interpreter_version; import sysconfig ; from sysconfig import get_config_var; print('{0}{1}-{0}{1}{2}'.format(interpreter_name(), interpreter_version(), 't' if sysconfig.get_config_var('Py_GIL_DISABLED') else ''))")
+    ${prefix}/bin/pip install wheel==0.45.1 setuptools==80.9.0
+    local abi_tag=$(${prefix}/bin/python -c "from wheel.pep425tags import get_abbr_impl, get_impl_ver, get_abi_tag; print('{0}{1}-{2}'.format(get_abbr_impl(), get_impl_ver(), get_abi_tag()))")
     ln -sf ${prefix} /opt/python/${abi_tag}
 }
 

.ci/docker/common/install_cuda.sh

@@ -68,8 +68,8 @@ function install_nvshmem {
   # download, unpack, install
   wget -q "${url}"
   tar xf "${filename}.tar.gz"
-  cp -a "libnvshmem/include/"* /usr/local/cuda/include/
-  cp -a "libnvshmem/lib/"*     /usr/local/cuda/lib64/
+  cp -a "libnvshmem/include/"* /usr/local/include/
+  cp -a "libnvshmem/lib/"*     /usr/local/lib/
 
   # cleanup
   cd ..

.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_inductor_benchmark_deps.sh

@@ -15,37 +15,11 @@ function install_timm() {
   commit=$(get_pinned_commit timm)
 
   pip_install "git+https://github.com/huggingface/pytorch-image-models@${commit}"
-}
-
-function install_torchbench() {
-  local commit
-  commit=$(get_pinned_commit torchbench)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  python install.py --continue_on_fail
-
-  # TODO (huydhn): transformers-4.44.2 added by https://github.com/pytorch/benchmark/pull/2488
-  # is regressing speedup metric. This needs to be investigated further
-  pip install transformers==4.38.1
-
-  echo "Print all dependencies after TorchBench is installed"
-  python -mpip freeze
-  popd
-
-  chown -R jenkins torchbench
+  # Clean up
+  conda_run pip uninstall -y torch torchvision triton
 }
 
 # Pango is needed for weasyprint which is needed for doctr
 conda_install pango
-
-# Stable packages are ok here, just to satisfy TorchBench check
-pip_install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
-
-install_torchbench
 install_huggingface
 install_timm
-
-# Clean up
-conda_run pip uninstall -y torch torchvision torchaudio triton torchao

.ci/docker/common/install_rocm.sh

@@ -30,25 +30,16 @@ EOF
 
     # we want the patch version of 6.4 instead
     if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-        ROCM_VERSION="${ROCM_VERSION}.2"
-    fi
-
-    # Default url values
-    rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
-    amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
-
-    # Special case for ROCM_VERSION == 7.0
-    if [[ $(ver "$ROCM_VERSION") -eq $(ver 7.0) ]]; then
-        rocm_baseurl="https://repo.radeon.com/rocm/apt/7.0_alpha2"
-        amdgpu_baseurl="https://repo.radeon.com/amdgpu/30.10_alpha2/ubuntu"
+        ROCM_VERSION="${ROCM_VERSION}.1"
     fi
 
     # Add amdgpu repository
     UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
-    echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
+    echo "deb [arch=amd64] https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
 
     # Add rocm repository
     wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
+    local rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
     echo "deb [arch=amd64] ${rocm_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/rocm.list
     apt-get update --allow-insecure-repositories
 
@@ -82,33 +73,30 @@ EOF
     done
 
     # ROCm 6.3 had a regression where initializing static code objects had significant overhead
-    # CI no longer builds for ROCm 6.3, but
     # ROCm 6.4 did not yet fix the regression, also HIP branch names are different
-    if [[ $(ver $ROCM_VERSION) -ge $(ver 6.4) ]] && [[ $(ver $ROCM_VERSION) -lt $(ver 7.0) ]]; then
-        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.2) ]]; then
-            HIP_TAG=rocm-6.4.2
-            CLR_HASH=74d78ba3ac4bac235d02bcb48511c30b5cfdd457  # branch release/rocm-rel-6.4.2-statco-hotfix
-        elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.1) ]]; then
-            HIP_TAG=rocm-6.4.1
-            CLR_HASH=efe6c35790b9206923bfeed1209902feff37f386  # branch release/rocm-rel-6.4.1-statco-hotfix
+    if [[ $(ver $ROCM_VERSION) -ge $(ver 6.3) ]] && [[ $(ver $ROCM_VERSION) -lt $(ver 7.0) ]]; then
+        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.1) ]]; then
+            HIP_BRANCH=release/rocm-rel-6.4
+            VER_STR=6.4
+            VER_PATCH=.1
         elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-            HIP_TAG=rocm-6.4.0
-            CLR_HASH=600f5b0d2baed94d5121e2174a9de0851b040b0c  # branch release/rocm-rel-6.4-statco-hotfix
+            HIP_BRANCH=release/rocm-rel-6.4
+            VER_STR=6.4
+        elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.3) ]]; then
+            HIP_BRANCH=rocm-6.3.x
+            VER_STR=6.3
         fi
         # clr build needs CppHeaderParser but can only find it using conda's python
         python -m pip install CppHeaderParser
-        git clone https://github.com/ROCm/HIP -b $HIP_TAG
+        git clone https://github.com/ROCm/HIP -b $HIP_BRANCH
         HIP_COMMON_DIR=$(readlink -f HIP)
-        git clone https://github.com/jeffdaily/clr
-        pushd clr
-        git checkout $CLR_HASH
-        popd
+        git clone https://github.com/jeffdaily/clr -b release/rocm-rel-${VER_STR}${VER_PATCH}-statco-hotfix
         mkdir -p clr/build
         pushd clr/build
         # Need to point CMake to the correct python installation to find CppHeaderParser
         cmake .. -DPython3_EXECUTABLE=/opt/conda/envs/py_${ANACONDA_PYTHON_VERSION}/bin/python3 -DCLR_BUILD_HIP=ON -DHIP_COMMON_DIR=$HIP_COMMON_DIR
         make -j
-        cp hipamd/lib/libamdhip64.so.6.4.* /opt/rocm/lib/libamdhip64.so.6.4.*
+        cp hipamd/lib/libamdhip64.so.${VER_STR}.* /opt/rocm/lib/libamdhip64.so.${VER_STR}.*
         popd
         rm -rf HIP clr
     fi

.ci/docker/common/install_xpu.sh

@@ -34,27 +34,18 @@ function install_ubuntu() {
 
     # The xpu-smi packages
     apt-get install -y flex bison xpu-smi
-
-    if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-        # Compute and Media Runtimes
-        apt-get install -y \
-            intel-opencl-icd intel-level-zero-gpu level-zero \
-            intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
-        # Development Packages
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
-    else # rolling driver
-        apt-get install -y \
-            intel-opencl-icd libze-intel-gpu1 libze1 \
-            intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev
+    # Compute and Media Runtimes
+    apt-get install -y \
+        intel-opencl-icd intel-level-zero-gpu level-zero \
+        intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
+        libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
+        libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
+        mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
+    if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+        apt-get install -y intel-ocloc
     fi
-
+    # Development Packages
+    apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
     # Install Intel Support Packages
     apt-get install -y ${XPU_PACKAGES}
 
@@ -139,11 +130,11 @@ function install_sles() {
 
 }
 
-# Default use GPU driver rolling releases
-XPU_DRIVER_VERSION=""
-if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-    # Use GPU driver LTS releases
-    XPU_DRIVER_VERSION="/lts/2350"
+# Default use GPU driver LTS releases
+XPU_DRIVER_VERSION="/lts/2350"
+if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+    # Use GPU driver rolling releases
+    XPU_DRIVER_VERSION=""
 fi
 
 # Default use Intel® oneAPI Deep Learning Essentials 2025.0

.ci/docker/libtorch/build.sh

@@ -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/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

@@ -50,7 +50,7 @@ flatbuffers==24.12.23
 hypothesis==5.35.1
 # Pin hypothesis to avoid flakiness: https://github.com/pytorch/pytorch/issues/31136
 #Description: advanced library for generating parametrized tests
-#Pinned versions: 5.35.1
+#Pinned versions: 3.44.6, 4.53.2
 #test that import: test_xnnpack_integration.py, test_pruning_op.py, test_nn.py
 
 junitparser==2.1.1
@@ -63,12 +63,11 @@ lark==0.12.0
 #Pinned versions: 0.12.0
 #test that import:
 
-librosa>=0.6.2 ; python_version < "3.11" and platform_machine != "s390x"
-librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
+librosa>=0.6.2 ; python_version < "3.11"
+librosa==0.10.2 ; python_version == "3.12"
 #Description: A python package for music and audio analysis
 #Pinned versions: >=0.6.2
 #test that import: test_spectral_ops.py
-#librosa depends on numba; disable it for s390x while numba is disabled too
 
 #mkl #this breaks linux-bionic-rocm4.5-py3.7
 #Description: Intel oneAPI Math Kernel Library
@@ -111,15 +110,14 @@ ninja==1.11.1.3
 #Pinned versions: 1.11.1.3
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
 
-numba==0.49.0 ; python_version < "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
-numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
+numba==0.49.0 ; python_version < "3.9"
+numba==0.55.2 ; python_version == "3.9"
+numba==0.55.2 ; python_version == "3.10"
+numba==0.60.0 ; python_version == "3.12"
 #Description: Just-In-Time Compiler for Numerical Functions
 #Pinned versions: 0.54.1, 0.49.0, <=0.49.1
 #test that import: test_numba_integration.py
 #For numba issue see https://github.com/pytorch/pytorch/issues/51511
-#Need release > 0.61.2 for s390x due to https://github.com/numba/numba/pull/10073
 
 #numpy
 #Description: Provides N-dimensional arrays and linear algebra
@@ -223,9 +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
@@ -235,7 +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"
@@ -309,7 +307,7 @@ pytest-cpp==2.3.0
 #Pinned versions: 2.3.0
 #test that import:
 
-z3-solver==4.15.1.0 ; platform_machine != "s390x"
+z3-solver==4.12.6.0
 #Description: The Z3 Theorem Prover Project
 #Pinned versions:
 #test that import:
@@ -363,6 +361,7 @@ pwlf==2.2.1
 #Pinned versions: 2.2.1
 #test that import: test_sac_estimator.py
 
+
 # To build PyTorch itself
 pyyaml
 pyzstd
@@ -390,9 +389,3 @@ tlparse==0.3.30
 cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"
 #Description: required for testing CUDAGraph::raw_cuda_graph(). See https://nvidia.github.io/cuda-python/cuda-bindings/latest/support.html for how this version was chosen. Note "Any fix in the latest bindings would be backported to the prior major version" means that only the newest version of cuda-bindings will get fixes. Depending on the latest version of 12.x is okay because all 12.y versions will be supported via "CUDA minor version compatibility". Pytorch builds against 13.z versions of cuda toolkit work with 12.x versions of cuda-bindings as well because newer drivers work with old toolkits.
 #test that import: test_cuda.py
-
-setuptools-git-versioning==2.1.0
-scikit-build==0.18.1
-pyre-extensions==0.0.32
-tabulate==0.9.0
-#Description: These package are needed to build FBGEMM and torchrec on PyTorch CI

.ci/docker/requirements-docs.txt

@@ -1,10 +1,10 @@
 sphinx==5.3.0
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 5.3.0
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git@722b7e6f9ca512fcc526ad07d62b3d28c50bb6cd#egg=pytorch_sphinx_theme2
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
 
 # TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
-# but it doesn't seem to work and hangs around idly. The initial thought that it is probably
+# but it doesn't seem to work and hangs around idly. The initial thought is probably
 # something related to Docker setup. We can investigate this later.
 
 sphinxcontrib.katex==0.8.6
@@ -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

.ci/docker/ubuntu-rocm/Dockerfile

@@ -98,9 +98,8 @@ COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -98,9 +98,8 @@ COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 ARG TRITON
 ARG TRITON_CPU

.ci/manywheel/build_common.sh

@@ -97,7 +97,8 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
     exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -qUr requirements-build.txt
+retry pip install -q "setuptools>=70.1.0" packaging
+retry pip install -qU cmake ninja
 python setup.py clean
 retry pip install -qr requirements.txt
 case ${DESIRED_PYTHON} in
@@ -138,11 +139,28 @@ fi
 
 echo "Calling setup.py bdist at $(date)"
 
-time CMAKE_ARGS=${CMAKE_ARGS[@]} \
-    EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 \
     BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
     USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
     python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 \
+    BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+    USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+    CMAKE_FRESH=1 python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    time CMAKE_ARGS=${CMAKE_ARGS[@]} \
+        EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+        BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+        USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+        python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+fi
 echo "Finished setup.py bdist at $(date)"
 
 # Build libtorch packages
@@ -255,6 +273,10 @@ ls /tmp/$WHEELHOUSE_DIR
 mkdir -p "/$WHEELHOUSE_DIR"
 mv /tmp/$WHEELHOUSE_DIR/torch*linux*.whl /$WHEELHOUSE_DIR/
 
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    mv /tmp/$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/ || true
+fi
+
 if [[ -n "$BUILD_PYTHONLESS" ]]; then
     mkdir -p /$LIBTORCH_HOUSE_DIR
     mv /tmp/$LIBTORCH_HOUSE_DIR/*.zip /$LIBTORCH_HOUSE_DIR
@@ -431,8 +453,16 @@ if [[ -z "$BUILD_PYTHONLESS" ]]; then
   pushd $PYTORCH_ROOT/test
 
   # Install the wheel for this Python version
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip uninstall -y "$TORCH_NO_PYTHON_PACKAGE_NAME" || true
+  fi
+
   pip uninstall -y "$TORCH_PACKAGE_NAME"
 
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip install "$TORCH_NO_PYTHON_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
+  fi
+
   pip install "$TORCH_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
 
   # Print info on the libraries installed in this wheel

.ci/manywheel/build_libtorch.sh

@@ -92,7 +92,8 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
     exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -qUr requirements-build.txt
+retry pip install -q "setuptools>=70.1.0" packaging
+retry pip install -qU cmake ninja
 python setup.py clean
 retry pip install -qr requirements.txt
 retry pip install -q numpy==2.0.1

.ci/manywheel/build_rocm.sh

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

.ci/pytorch/build-mobile.sh

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

.ci/pytorch/build.sh

@@ -11,6 +11,10 @@ source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
 # shellcheck source=./common-build.sh
 source "$(dirname "${BASH_SOURCE[0]}")/common-build.sh"
 
+if [[ "$BUILD_ENVIRONMENT" == *-mobile-*build* ]]; then
+  exec "$(dirname "${BASH_SOURCE[0]}")/build-mobile.sh" "$@"
+fi
+
 echo "Python version:"
 python --version
 
@@ -50,6 +54,9 @@ if [[ ${BUILD_ENVIRONMENT} == *"parallelnative"* ]]; then
   export ATEN_THREADING=NATIVE
 fi
 
+# Enable LLVM dependency for TensorExpr testing
+export USE_LLVM=/opt/llvm
+export LLVM_DIR=/opt/llvm/lib/cmake/llvm
 
 if ! which conda; then
   # In ROCm CIs, we are doing cross compilation on build machines with
@@ -117,8 +124,26 @@ if [[ "$BUILD_ENVIRONMENT" == *libtorch* ]]; then
 fi
 
 # Use special scripts for Android builds
+if [[ "${BUILD_ENVIRONMENT}" == *-android* ]]; then
+  export ANDROID_NDK=/opt/ndk
+  build_args=()
+  if [[ "${BUILD_ENVIRONMENT}" == *-arm-v7a* ]]; then
+    build_args+=("-DANDROID_ABI=armeabi-v7a")
+  elif [[ "${BUILD_ENVIRONMENT}" == *-arm-v8a* ]]; then
+    build_args+=("-DANDROID_ABI=arm64-v8a")
+  elif [[ "${BUILD_ENVIRONMENT}" == *-x86_32* ]]; then
+    build_args+=("-DANDROID_ABI=x86")
+  elif [[ "${BUILD_ENVIRONMENT}" == *-x86_64* ]]; then
+    build_args+=("-DANDROID_ABI=x86_64")
+  fi
+  if [[ "${BUILD_ENVIRONMENT}" == *vulkan* ]]; then
+    build_args+=("-DUSE_VULKAN=ON")
+  fi
+  build_args+=("-DUSE_LITE_INTERPRETER_PROFILER=OFF")
+  exec ./scripts/build_android.sh "${build_args[@]}" "$@"
+fi
 
-if [[ "$BUILD_ENVIRONMENT" == *vulkan* ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *android* && "$BUILD_ENVIRONMENT" == *vulkan* ]]; then
   export USE_VULKAN=1
   # shellcheck disable=SC1091
   source /var/lib/jenkins/vulkansdk/setup-env.sh
@@ -173,7 +198,7 @@ fi
 
 # We only build FlashAttention files for CUDA 8.0+, and they require large amounts of
 # memory to build and will OOM
-if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && echo "${TORCH_CUDA_ARCH_LIST}" | tr ' ' '\n' | sed 's/$/>= 8.0/' | bc | grep -q 1; then
+if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]]; then
   export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
 fi
 
@@ -189,6 +214,7 @@ if [[ "$BUILD_ENVIRONMENT" == *-clang*-asan* ]]; then
   export USE_ASAN=1
   export REL_WITH_DEB_INFO=1
   export UBSAN_FLAGS="-fno-sanitize-recover=all"
+  unset USE_LLVM
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *no-ops* ]]; then
@@ -199,7 +225,7 @@ if [[ "${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
 
@@ -261,32 +287,25 @@ else
 
       WERROR=1 python setup.py clean
 
-      WERROR=1 python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        python3 tools/packaging/split_wheel.py bdist_wheel
+      else
+        WERROR=1 python setup.py bdist_wheel
+      fi
     else
       python setup.py clean
       if [[ "$BUILD_ENVIRONMENT" == *xla* ]]; then
         source .ci/pytorch/install_cache_xla.sh
       fi
-      python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        echo "USE_SPLIT_BUILD cannot be used with xla or rocm"
+        exit 1
+      else
+        python setup.py bdist_wheel
+      fi
     fi
     pip_install_whl "$(echo dist/*.whl)"
 
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *vision* ]]; then
-      install_torchvision
-    fi
-
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *audio* ]]; then
-      install_torchaudio
-    fi
-
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *torchrec* || "${BUILD_ADDITIONAL_PACKAGES:-}" == *fbgemm* ]]; then
-      install_torchrec_and_fbgemm
-    fi
-
-    if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *torchao* ]]; then
-      install_torchao
-    fi
-
     if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
       echo "Checking that xpu is compiled"
       pushd dist/

.ci/pytorch/common_utils.sh

@@ -78,34 +78,6 @@ function pip_install_whl() {
   fi
 }
 
-function pip_build_and_install() {
-  local build_target=$1
-  local wheel_dir=$2
-
-  local found_whl=0
-  for file in "${wheel_dir}"/*.whl
-  do
-    if [[ -f "${file}" ]]; then
-      found_whl=1
-      break
-    fi
-  done
-
-  # Build the wheel if it doesn't exist
-  if [ "${found_whl}" == "0" ]; then
-    python3 -m pip wheel \
-      --no-build-isolation \
-      --no-deps \
-      --no-use-pep517 \
-      -w "${wheel_dir}" \
-      "${build_target}"
-  fi
-
-  for file in "${wheel_dir}"/*.whl
-  do
-    pip_install_whl "${file}"
-  done
-}
 
 function pip_install() {
   # retry 3 times
@@ -152,7 +124,14 @@ function get_pinned_commit() {
 function install_torchaudio() {
   local commit
   commit=$(get_pinned_commit audio)
-  pip_build_and_install "git+https://github.com/pytorch/audio.git@${commit}" dist/audio
+  if [[ "$1" == "cuda" ]]; then
+    # TODO: This is better to be passed as a parameter from _linux-test workflow
+    # so that it can be consistent with what is set in build
+    TORCH_CUDA_ARCH_LIST="8.0;8.6" pip_install --no-use-pep517 "git+https://github.com/pytorch/audio.git@${commit}"
+  else
+    pip_install --no-use-pep517 "git+https://github.com/pytorch/audio.git@${commit}"
+  fi
+
 }
 
 function install_torchtext() {
@@ -160,8 +139,8 @@ function install_torchtext() {
   local text_commit
   data_commit=$(get_pinned_commit data)
   text_commit=$(get_pinned_commit text)
-  pip_build_and_install "git+https://github.com/pytorch/data.git@${data_commit}" dist/data
-  pip_build_and_install "git+https://github.com/pytorch/text.git@${text_commit}" dist/text
+  pip_install --no-use-pep517 "git+https://github.com/pytorch/data.git@${data_commit}"
+  pip_install --no-use-pep517 "git+https://github.com/pytorch/text.git@${text_commit}"
 }
 
 function install_torchvision() {
@@ -174,14 +153,7 @@ function install_torchvision() {
     echo 'char* dlerror(void) { return "";}'|gcc -fpic -shared -o "${HOME}/dlerror.so" -x c -
     LD_PRELOAD=${orig_preload}:${HOME}/dlerror.so
   fi
-
-  if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]]; then
-    # Not sure if both are needed, but why not
-    export FORCE_CUDA=1
-    export WITH_CUDA=1
-  fi
-  pip_build_and_install "git+https://github.com/pytorch/vision.git@${commit}" dist/vision
-
+  pip_install --no-use-pep517 "git+https://github.com/pytorch/vision.git@${commit}"
   if [ -n "${LD_PRELOAD}" ]; then
     LD_PRELOAD=${orig_preload}
   fi
@@ -201,71 +173,25 @@ function install_torchrec_and_fbgemm() {
 
   if [[ "$BUILD_ENVIRONMENT" == *rocm* ]] ; then
     # install torchrec first because it installs fbgemm nightly on top of rocm fbgemm
-    pip_build_and_install "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}" dist/torchrec
+    pip_install --no-use-pep517 "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}"
     pip_uninstall fbgemm-gpu-nightly
 
-    # Set ROCM_HOME isn't available, use ROCM_PATH if set or /opt/rocm
-    ROCM_HOME="${ROCM_HOME:-${ROCM_PATH:-/opt/rocm}}"
-
-    # Find rocm_version.h header file for ROCm version extract
-    rocm_version_h="${ROCM_HOME}/include/rocm-core/rocm_version.h"
-    if [ ! -f "$rocm_version_h" ]; then
-        rocm_version_h="${ROCM_HOME}/include/rocm_version.h"
-    fi
-
-    # Error out if rocm_version.h not found
-    if [ ! -f "$rocm_version_h" ]; then
-        echo "Error: rocm_version.h not found in expected locations." >&2
-        exit 1
-    fi
-
-    # Extract major, minor and patch ROCm version numbers
-    MAJOR_VERSION=$(grep 'ROCM_VERSION_MAJOR' "$rocm_version_h" | awk '{print $3}')
-    MINOR_VERSION=$(grep 'ROCM_VERSION_MINOR' "$rocm_version_h" | awk '{print $3}')
-    PATCH_VERSION=$(grep 'ROCM_VERSION_PATCH' "$rocm_version_h" | awk '{print $3}')
-    ROCM_INT=$((MAJOR_VERSION * 10000 + MINOR_VERSION * 100 + PATCH_VERSION))
-    echo "ROCm version: $ROCM_INT"
-    export BUILD_ROCM_VERSION="$MAJOR_VERSION.$MINOR_VERSION"
-
     pip_install tabulate  # needed for newer fbgemm
     pip_install patchelf  # needed for rocm fbgemm
-
-    local wheel_dir=dist/fbgemm_gpu
-    local found_whl=0
-    for file in "${wheel_dir}"/*.whl
-    do
-      if [[ -f "${file}" ]]; then
-        found_whl=1
-        break
-      fi
-    done
-
-    # Build the wheel if it doesn't exist
-    if [ "${found_whl}" == "0" ]; then
-      git clone --recursive https://github.com/pytorch/fbgemm
-      pushd fbgemm/fbgemm_gpu
-      git checkout "${fbgemm_commit}" --recurse-submodules
-      python setup.py bdist_wheel \
-        --build-variant=rocm \
-        -DHIP_ROOT_DIR="${ROCM_PATH}" \
-        -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
-        -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
-      popd
-
-      # Save the wheel before cleaning up
-      mkdir -p dist/fbgemm_gpu
-      cp fbgemm/fbgemm_gpu/dist/*.whl dist/fbgemm_gpu
-    fi
-
-    for file in "${wheel_dir}"/*.whl
-    do
-      pip_install_whl "${file}"
-    done
-
+    git clone --recursive https://github.com/pytorch/fbgemm
+    pushd fbgemm/fbgemm_gpu
+    git checkout "${fbgemm_commit}"
+    python setup.py install \
+      --package_variant=rocm \
+      -DHIP_ROOT_DIR="${ROCM_PATH}" \
+      -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
+      -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
+    popd
     rm -rf fbgemm
   else
-    pip_build_and_install "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}" dist/torchrec
-    pip_build_and_install "git+https://github.com/pytorch/FBGEMM.git@${fbgemm_commit}#subdirectory=fbgemm_gpu" dist/fbgemm_gpu
+    # See https://github.com/pytorch/pytorch/issues/106971
+    CUDA_PATH=/usr/local/cuda-12.1 pip_install --no-use-pep517 "git+https://github.com/pytorch/FBGEMM.git@${fbgemm_commit}#egg=fbgemm-gpu&subdirectory=fbgemm_gpu"
+    pip_install --no-use-pep517 "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}"
   fi
 }
 
@@ -281,10 +207,34 @@ function clone_pytorch_xla() {
   fi
 }
 
+function checkout_install_torchbench() {
+  local commit
+  commit=$(get_pinned_commit torchbench)
+  git clone https://github.com/pytorch/benchmark torchbench
+  pushd torchbench
+  git checkout "$commit"
+
+  if [ "$1" ]; then
+    python install.py --continue_on_fail models "$@"
+  else
+    # Occasionally the installation may fail on one model but it is ok to continue
+    # to install and test other models
+    python install.py --continue_on_fail
+  fi
+
+  # TODO (huydhn): transformers-4.44.2 added by https://github.com/pytorch/benchmark/pull/2488
+  # is regressing speedup metric. This needs to be investigated further
+  pip install transformers==4.38.1
+
+  echo "Print all dependencies after TorchBench is installed"
+  python -mpip freeze
+  popd
+}
+
 function install_torchao() {
   local commit
   commit=$(get_pinned_commit torchao)
-  pip_build_and_install "git+https://github.com/pytorch/ao.git@${commit}" dist/ao
+  pip_install --no-use-pep517 "git+https://github.com/pytorch/ao.git@${commit}"
 }
 
 function print_sccache_stats() {

.ci/pytorch/cpp_doc_push_script.sh

@@ -58,6 +58,7 @@ time python tools/setup_helpers/generate_code.py \
 
 # Build the docs
 pushd docs/cpp
+pip install -e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
 time make VERBOSE=1 html -j
 
 popd

.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/functorch_doc_push_script.sh

@@ -12,6 +12,7 @@ echo "version: $version"
 
 # Build functorch docs
 pushd $pt_checkout/functorch/docs
+pip install -e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
 make html
 popd
 

.ci/pytorch/macos-test.sh

@@ -157,29 +157,6 @@ test_jit_hooks() {
   assert_git_not_dirty
 }
 
-# Shellcheck doesn't like it when you pass no arguments to a function
-# that can take args. See https://www.shellcheck.net/wiki/SC2120
-# shellcheck disable=SC2120
-checkout_install_torchbench() {
-  local commit
-  commit=$(cat .ci/docker/ci_commit_pins/torchbench.txt)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  if [ "$1" ]; then
-    python install.py --continue_on_fail models "$@"
-  else
-    # Occasionally the installation may fail on one model but it is ok to continue
-    # to install and test other models
-    python install.py --continue_on_fail
-  fi
-
-  echo "Print all dependencies after TorchBench is installed"
-  python -mpip freeze
-  popd
-}
-
 torchbench_setup_macos() {
   git clone --recursive https://github.com/pytorch/vision torchvision
   git clone --recursive https://github.com/pytorch/audio torchaudio
@@ -202,6 +179,8 @@ torchbench_setup_macos() {
   USE_OPENMP=0 python setup.py develop
   popd
 
+  # Shellcheck doesn't like it when you pass no arguments to a function that can take args. See https://www.shellcheck.net/wiki/SC2120
+  # shellcheck disable=SC2119,SC2120
   checkout_install_torchbench
 }
 

.ci/pytorch/python_doc_push_script.sh

@@ -46,7 +46,7 @@ echo "error: python_doc_push_script.sh: branch (arg3) not specified"
 fi
 
 echo "install_path: $install_path  version: $version"
-
+pip install -e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
 
 build_docs () {
   set +e

.ci/pytorch/run_glootls_test.sh

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

.ci/pytorch/run_tests.sh

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

.ci/pytorch/smoke_test/smoke_test.py

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

.ci/pytorch/test.sh

@@ -289,12 +289,6 @@ elif [[ $TEST_CONFIG == 'nogpu_AVX512' ]]; then
   export ATEN_CPU_CAPABILITY=avx2
 fi
 
-if [[ "${TEST_CONFIG}" == "legacy_nvidia_driver" ]]; then
-  # Make sure that CUDA can be initialized
-  (cd test && python -c "import torch; torch.rand(2, 2, device='cuda')")
-  export USE_LEGACY_DRIVER=1
-fi
-
 test_python_legacy_jit() {
   time python test/run_test.py --include test_jit_legacy test_jit_fuser_legacy --verbose
   assert_git_not_dirty
@@ -345,12 +339,6 @@ 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 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-}
-
 test_lazy_tensor_meta_reference_disabled() {
   export TORCH_DISABLE_FUNCTIONALIZATION_META_REFERENCE=1
   echo "Testing lazy tensor operations without meta reference"
@@ -365,6 +353,7 @@ test_dynamo_wrapped_shard() {
     exit 1
   fi
   python tools/dynamo/verify_dynamo.py
+  python tools/dynamo/gb_id_mapping.py verify
   # PLEASE DO NOT ADD ADDITIONAL EXCLUDES HERE.
   # Instead, use @skipIfTorchDynamo on your tests.
   time python test/run_test.py --dynamo \
@@ -462,7 +451,7 @@ test_inductor_aoti() {
   # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
   /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${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() {
@@ -627,8 +616,6 @@ test_perf_for_dashboard() {
     device=cuda_a10g
   elif [[ "${TEST_CONFIG}" == *h100* ]]; then
     device=cuda_h100
-  elif [[ "${TEST_CONFIG}" == *b200* ]]; then
-    device=cuda_b200
   elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
     device=rocm
   fi
@@ -803,16 +790,6 @@ test_dynamo_benchmark() {
   if [[ "${TEST_CONFIG}" == *perf_compare* ]]; then
     test_single_dynamo_benchmark "training" "$suite" "$shard_id" --training --amp "$@"
   elif [[ "${TEST_CONFIG}" == *perf* ]]; then
-    # TODO (huydhn): Just smoke test some sample models
-    if [[ "${TEST_CONFIG}" == *b200* ]]; then
-      if [[ "${suite}" == "huggingface" ]]; then
-        export TORCHBENCH_ONLY_MODELS="DistillGPT2"
-      elif [[ "${suite}" == "timm_models" ]]; then
-        export TORCHBENCH_ONLY_MODELS="inception_v3"
-      elif [[ "${suite}" == "torchbench" ]]; then
-        export TORCHBENCH_ONLY_MODELS="hf_Bert"
-      fi
-    fi
     test_single_dynamo_benchmark "dashboard" "$suite" "$shard_id" "$@"
   else
     if [[ "${TEST_CONFIG}" == *cpu* ]]; then
@@ -940,6 +917,12 @@ test_torchbench_gcp_smoketest(){
   popd
 }
 
+test_python_gloo_with_tls() {
+  source "$(dirname "${BASH_SOURCE[0]}")/run_glootls_test.sh"
+  assert_git_not_dirty
+}
+
+
 test_aten() {
   # Test ATen
   # The following test(s) of ATen have already been skipped by caffe2 in rocm environment:
@@ -986,8 +969,6 @@ test_without_numpy() {
   if [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then
     python -c "import sys;sys.path.insert(0, 'fake_numpy');import torch;torch.compile(lambda x:print(x))('Hello World')"
   fi
-  # Regression test for https://github.com/pytorch/pytorch/pull/157734 (torch.onnx should be importable without numpy)
-  python -c "import sys;sys.path.insert(0, 'fake_numpy');import torch; import torch.onnx"
   popd
 }
 
@@ -1051,10 +1032,20 @@ test_libtorch_api() {
     mkdir -p $TEST_REPORTS_DIR
 
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" "$TORCH_BIN_DIR"/test_api --gtest_filter='-IMethodTest.*' --gtest_output=xml:$TEST_REPORTS_DIR/test_api.xml
+    "$TORCH_BIN_DIR"/test_tensorexpr --gtest_output=xml:$TEST_REPORTS_DIR/test_tensorexpr.xml
   else
     # Exclude IMethodTest that relies on torch::deploy, which will instead be ran in test_deploy
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_api -k "not IMethodTest"
 
+    # On s390x, pytorch is built without llvm.
+    # Even if it would be built with llvm, llvm currently doesn't support used features on s390x and
+    # test fails with errors like:
+    # JIT session error: Unsupported target machine architecture in ELF object pytorch-jitted-objectbuffer
+    # unknown file: Failure
+    # C++ exception with description "valOrErr INTERNAL ASSERT FAILED at "/var/lib/jenkins/workspace/torch/csrc/jit/tensorexpr/llvm_jit.h":34, please report a bug to PyTorch. Unexpected failure in LLVM JIT: Failed to materialize symbols: { (main, { func }) }
+    if [[ "${BUILD_ENVIRONMENT}" != *s390x* ]]; then
+      python test/run_test.py --cpp --verbose -i cpp/test_tensorexpr
+    fi
   fi
 
   # quantization is not fully supported on s390x yet
@@ -1322,13 +1313,10 @@ 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 \
@@ -1562,7 +1550,7 @@ 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
@@ -1612,13 +1600,7 @@ if ! [[ "${BUILD_ENVIRONMENT}" == *libtorch* || "${BUILD_ENVIRONMENT}" == *-baze
 fi
 if [[ "${TEST_CONFIG}" == *numpy_2* ]]; then
   # Install numpy-2.0.2 and compatible scipy & numba versions
-  # Force re-install of pandas to avoid error where pandas checks numpy version from initial install and fails upon import
-  TMP_PANDAS_VERSION=$(python -c "import pandas; print(pandas.__version__)" 2>/dev/null)
-  if [ -n "$TMP_PANDAS_VERSION" ]; then
-    python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0 pandas=="$TMP_PANDAS_VERSION" --force-reinstall
-  else
-    python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
-  fi
+  python -mpip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
   python test/run_test.py --include dynamo/test_functions.py dynamo/test_unspec.py test_binary_ufuncs.py test_fake_tensor.py test_linalg.py test_numpy_interop.py test_tensor_creation_ops.py test_torch.py torch_np/test_basic.py
 elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
   test_linux_aarch64
@@ -1672,37 +1654,49 @@ elif [[ "${TEST_CONFIG}" == *timm* ]]; then
   id=$((SHARD_NUMBER-1))
   test_dynamo_benchmark timm_models "$id"
 elif [[ "${TEST_CONFIG}" == cachebench ]]; then
-  install_torchaudio
+  install_torchaudio cuda
   install_torchvision
-  PYTHONPATH=/torchbench test_cachebench
+  checkout_install_torchbench nanogpt BERT_pytorch resnet50 hf_T5 llama moco
+  PYTHONPATH=$(pwd)/torchbench test_cachebench
 elif [[ "${TEST_CONFIG}" == verify_cachebench ]]; then
-  install_torchaudio
+  install_torchaudio cpu
   install_torchvision
-  PYTHONPATH=/torchbench test_verify_cachebench
+  checkout_install_torchbench nanogpt
+  PYTHONPATH=$(pwd)/torchbench test_verify_cachebench
 elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
-  install_torchaudio
+  if [[ "${TEST_CONFIG}" == *cpu* ]]; then
+    install_torchaudio cpu
+  else
+    install_torchaudio cuda
+  fi
   install_torchvision
-  install_torchao
+  TORCH_CUDA_ARCH_LIST="8.0;8.6" install_torchao
   id=$((SHARD_NUMBER-1))
   # https://github.com/opencv/opencv-python/issues/885
   pip_install opencv-python==4.8.0.74
   if [[ "${TEST_CONFIG}" == *inductor_torchbench_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_smoketest_perf
+    checkout_install_torchbench hf_Bert hf_Albert timm_vision_transformer
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *inductor_torchbench_cpu_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_cpu_smoketest_perf
+    checkout_install_torchbench timm_vision_transformer phlippe_densenet basic_gnn_edgecnn \
+      llama_v2_7b_16h resnet50 timm_efficientnet mobilenet_v3_large timm_resnest \
+      functorch_maml_omniglot yolov3 mobilenet_v2 resnext50_32x4d densenet121 mnasnet1_0
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_cpu_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *torchbench_gcp_smoketest* ]]; then
-    TORCHBENCHPATH=/torchbench test_torchbench_gcp_smoketest
+    checkout_install_torchbench
+    TORCHBENCHPATH=$(pwd)/torchbench test_torchbench_gcp_smoketest
   else
+    checkout_install_torchbench
     # Do this after checkout_install_torchbench to ensure we clobber any
     # nightlies that torchbench may pull in
     if [[ "${TEST_CONFIG}" != *cpu* ]]; then
       install_torchrec_and_fbgemm
     fi
-    PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
+    PYTHONPATH=$(pwd)/torchbench test_dynamo_benchmark torchbench "$id"
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
-  PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
+  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
     test_inductor_aoti
   fi
@@ -1767,8 +1761,6 @@ elif [[ "${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.sh

@@ -41,7 +41,7 @@ fi
 python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
 
 # Install Z3 optional dependency for Windows builds.
-python -m pip install z3-solver==4.15.1.0
+python -m pip install z3-solver==4.12.2.0
 
 # Install tlparse for test\dynamo\test_structured_trace.py UTs.
 python -m pip install tlparse==0.3.30

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

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

.ci/wheel/build_wheel.sh

@@ -184,14 +184,16 @@ tmp_env_name="wheel_py$python_nodot"
 conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
 source activate "$tmp_env_name"
 
-retry pip install -r "${pytorch_rootdir}/requirements-build.txt"
-pip install "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing-extensions
+pip install "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing_extensions
 retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
 retry brew install libomp
 
 # For USE_DISTRIBUTED=1 on macOS, need libuv, which is build as part of tensorpipe submodule
 export USE_DISTRIBUTED=1
 
+if [[ -n "$CROSS_COMPILE_ARM64" ]]; then
+    export CMAKE_OSX_ARCHITECTURES=arm64
+fi
 export USE_MKLDNN=OFF
 export USE_QNNPACK=OFF
 export BUILD_TEST=OFF
@@ -199,7 +201,16 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-python setup.py bdist_wheel -d "$whl_tmp_dir"
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 CMAKE_FRESH=1 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    python setup.py bdist_wheel -d "$whl_tmp_dir"
+fi
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.circleci/scripts/binary_linux_test.sh

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

.circleci/scripts/binary_populate_env.sh

@@ -134,6 +134,7 @@ export DESIRED_PYTHON="${DESIRED_PYTHON:-}"
 export DESIRED_CUDA="$DESIRED_CUDA"
 export LIBTORCH_VARIANT="${LIBTORCH_VARIANT:-}"
 export BUILD_PYTHONLESS="${BUILD_PYTHONLESS:-}"
+export USE_SPLIT_BUILD="${USE_SPLIT_BUILD:-}"
 if [[ "${OSTYPE}" == "msys" ]]; then
   export LIBTORCH_CONFIG="${LIBTORCH_CONFIG:-}"
   if [[ "${LIBTORCH_CONFIG:-}" == 'debug' ]]; then

.circleci/scripts/binary_upload.sh

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

.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/actions/linux-test/action.yml

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

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

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

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-e500f0cf88bc57ffd8b0029033da305eef24ae25
+6c57850358f34c47802db216b0746e4e9d08a95a

.github/ci_commit_pins/fbgemm_rocm.txt

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

.github/ci_commit_pins/vllm.txt

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

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-095faec1e7b6cc47220181e74ae9cde2605f9b00
+1c00dea2c9adb2137903c86b4191e8c247f8fda9

.github/merge_rules.yaml

@@ -76,7 +76,6 @@
   - .github/ci_commit_pins/audio.txt
   - .github/ci_commit_pins/vision.txt
   - .github/ci_commit_pins/torchdynamo.txt
-  - .github/ci_commit_pins/vllm.txt
   - .ci/docker/ci_commit_pins/triton.txt
   approved_by:
   - pytorchbot
@@ -131,6 +130,21 @@
   - Lint
   - pull
 
+- name: Mobile
+  patterns:
+  - ios/**
+  - android/**
+  - test/mobile/**
+  approved_by:
+  - linbinyu
+  - IvanKobzarev
+  - dreiss
+  - raziel
+  mandatory_checks_name:
+  - EasyCLA
+  - Lint
+  - pull
+
 - name: PrimTorch
   patterns:
   - torch/_meta_registrations.py
@@ -477,23 +491,6 @@
   - srossross
   - chillee
   - zou3519
-  - guilhermeleobas
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
-- name: Dynamo
-  patterns:
-  - torch/_dynamo/**
-  - torch/csrc/dynamo/**
-  - test/dynamo/**
-  - test/dynamo_expected_failures/**
-  - test/dynamo_skips/**
-  - test/inductor_expected_failures/**
-  - test/inductor_skips/**
-  approved_by:
-  - guilhermeleobas
   mandatory_checks_name:
   - EasyCLA
   - Lint

.github/pytorch-probot.yml

@@ -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

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

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

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

.github/scripts/generate_binary_build_matrix.py

@@ -193,7 +193,7 @@ LIBTORCH_CONTAINER_IMAGES: dict[str, str] = {
     "cpu": "libtorch-cxx11-builder:cpu",
 }
 
-FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
+FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t"]
 
 
 def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
@@ -273,6 +273,7 @@ def generate_wheels_matrix(
     os: str,
     arches: Optional[list[str]] = None,
     python_versions: Optional[list[str]] = None,
+    use_split_build: bool = False,
 ) -> list[dict[str, str]]:
     package_type = "wheel"
     if os == "linux" or os == "linux-aarch64" or os == "linux-s390x":
@@ -314,11 +315,15 @@ def generate_wheels_matrix(
             # TODO: Enable python 3.13t on cpu-s390x
             if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
                 continue
-            # TODO: Enable python 3.14 on non linux OSes
-            if os != "linux" and (
-                python_version == "3.14" or python_version == "3.14t"
+
+            if use_split_build and (
+                arch_version not in ["12.6", "12.8", "12.9", "cpu"] or os != "linux"
             ):
-                continue
+                raise RuntimeError(
+                    "Split build is only supported on linux with cuda 12* and cpu.\n"
+                    f"Currently attempting to build on arch version {arch_version} and os {os}.\n"
+                    "Please modify the matrix generation to exclude this combination."
+                )
 
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
@@ -334,6 +339,7 @@ def generate_wheels_matrix(
                         "gpu_arch_type": gpu_arch_type,
                         "gpu_arch_version": gpu_arch_version,
                         "desired_cuda": desired_cuda,
+                        "use_split_build": "True" if use_split_build else "False",
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],
@@ -366,6 +372,7 @@ def generate_wheels_matrix(
                             "desired_cuda": translate_desired_cuda(
                                 gpu_arch_type, gpu_arch_version
                             ),
+                            "use_split_build": "True" if use_split_build else "False",
                             "container_image": WHEEL_CONTAINER_IMAGES[
                                 arch_version
                             ].split(":")[0],
@@ -388,6 +395,7 @@ def generate_wheels_matrix(
                         "desired_cuda": translate_desired_cuda(
                             gpu_arch_type, gpu_arch_version
                         ),
+                        "use_split_build": "True" if use_split_build else "False",
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],

.github/scripts/generate_ci_workflows.py

@@ -59,7 +59,9 @@ class BinaryBuildWorkflow:
     is_scheduled: str = ""
     branches: str = "nightly"
     # Mainly for macos
+    cross_compile_arm64: bool = False
     macos_runner: str = "macos-14-xlarge"
+    use_split_build: bool = False
     # Mainly used for libtorch builds
     build_variant: str = ""
 
@@ -70,6 +72,9 @@ class BinaryBuildWorkflow:
                 for item in [self.os, "binary", self.package_type, self.build_variant]
                 if item != ""
             )
+        if self.use_split_build:
+            # added to distinguish concurrency groups
+            self.build_environment += "-split"
 
     def generate_workflow_file(self, workflow_template: jinja2.Template) -> None:
         output_file_path = (
@@ -112,6 +117,21 @@ LINUX_BINARY_BUILD_WORFKLOWS = [
             isolated_workflow=True,
         ),
     ),
+    # See https://github.com/pytorch/pytorch/issues/138750
+    #   BinaryBuildWorkflow(
+    #     os=OperatingSystem.LINUX,
+    #     package_type="manywheel",
+    #     build_configs=generate_binary_build_matrix.generate_wheels_matrix(
+    #         OperatingSystem.LINUX,
+    #         use_split_build=True,
+    #         arches=["11.8", "12.1", "12.4", "cpu"],
+    #     ),
+    #     ciflow_config=CIFlowConfig(
+    #         labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},
+    #         isolated_workflow=True,
+    #     ),
+    #     use_split_build=True,
+    # ),
     BinaryBuildWorkflow(
         os=OperatingSystem.LINUX,
         package_type="libtorch",
@@ -155,11 +175,27 @@ LINUX_BINARY_SMOKE_WORKFLOWS = [
         package_type="manywheel",
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
-            arches=["12.8"],
-            python_versions=["3.12"],
+            arches=["12.6", "12.8", "12.9"],
+            python_versions=["3.9"],
         ),
         branches="main",
     ),
+    # See https://github.com/pytorch/pytorch/issues/138750
+    # BinaryBuildWorkflow(
+    #     os=OperatingSystem.LINUX,
+    #     package_type="manywheel",
+    #     build_configs=generate_binary_build_matrix.generate_wheels_matrix(
+    #         OperatingSystem.LINUX,
+    #         arches=["11.8", "12.1", "12.4"],
+    #         python_versions=["3.9"],
+    #         use_split_build=True,
+    #     ),
+    #     ciflow_config=CIFlowConfig(
+    #         labels={LABEL_CIFLOW_PERIODIC},
+    #     ),
+    #     branches="main",
+    #     use_split_build=True,
+    # ),
     BinaryBuildWorkflow(
         os=OperatingSystem.LINUX,
         package_type="libtorch",
@@ -302,6 +338,7 @@ MACOS_BINARY_BUILD_WORKFLOWS = [
             generate_binary_build_matrix.RELEASE,
             libtorch_variants=["shared-with-deps"],
         ),
+        cross_compile_arm64=False,
         macos_runner="macos-14-xlarge",
         ciflow_config=CIFlowConfig(
             labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_LIBTORCH},
@@ -314,6 +351,7 @@ MACOS_BINARY_BUILD_WORKFLOWS = [
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.MACOS_ARM64
         ),
+        cross_compile_arm64=False,
         macos_runner="macos-14-xlarge",
         ciflow_config=CIFlowConfig(
             labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},

.github/scripts/lintrunner.sh

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

.github/scripts/runner_determinator.py

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

.github/scripts/trymerge.py

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

.github/templates/macos_binary_build_workflow.yml.j2

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

.github/templates/upload.yml.j2

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

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

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

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

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

.github/workflows/_binary-upload.yml

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

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

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

.github/workflows/_linux-build.yml

@@ -16,6 +16,11 @@ on:
         type: boolean
         default: true
         description: If set, upload generated build artifacts.
+      build-with-debug:
+        required: false
+        type: boolean
+        default: false
+        description: If set, build in debug mode.
       sync-tag:
         required: false
         type: string
@@ -82,6 +87,7 @@ on:
         required: false
         type: number
         default: 1
+
       allow-reuse-old-whl:
         description: |
           If set, the build try to pull an old wheel from s3 that was built on a
@@ -89,13 +95,6 @@ on:
         required: false
         type: boolean
         default: true
-      build-additional-packages:
-        description: |
-          If set, the build job will also builds these packages and saves their
-          wheels as artifacts
-        required: false
-        type: string
-        default: ""
 
     secrets:
       HUGGING_FACE_HUB_TOKEN:
@@ -107,6 +106,7 @@ on:
         description: |
           FB app token to write to scribe endpoint
 
+
     outputs:
       docker-image:
         value: ${{ jobs.build.outputs.docker-image }}
@@ -225,7 +225,7 @@ jobs:
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
         run: |
           mkdir -p ../../usage_logs
-          python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7
+          python3 -m pip install psutil==5.9.1 dataclasses_json==0.6.7
           python3 -m tools.stats.monitor \
           --log-interval "$MONITOR_LOG_INTERVAL" \
           --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" \
@@ -247,6 +247,8 @@ jobs:
         env:
           BUILD_ENVIRONMENT: ${{ inputs.build-environment }}
           BRANCH: ${{ steps.parse-ref.outputs.branch }}
+          # TODO duplicated
+          AWS_DEFAULT_REGION: us-east-1
           PR_NUMBER: ${{ github.event.pull_request.number }}
           SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
           # Do not set SCCACHE_S3_KEY_PREFIX to share the cache between all build jobs
@@ -258,10 +260,10 @@ jobs:
           DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
           DOCKER_IMAGE_S390X: ${{ inputs.docker-image-name }}
           XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
+          DEBUG: ${{ inputs.build-with-debug && '1' || '0' }}
           OUR_GITHUB_JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
           HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
           SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.SCRIBE_GRAPHQL_ACCESS_TOKEN }}
-          BUILD_ADDITIONAL_PACKAGES: ${{ inputs.build-additional-packages }}
         run: |
           START_TIME=$(date +%s)
           if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
@@ -293,6 +295,7 @@ jobs:
           container_name=$(docker run \
             -e BUILD_ENVIRONMENT \
             -e MAX_JOBS="$(nproc --ignore=2)" \
+            -e AWS_DEFAULT_REGION \
             -e PR_NUMBER \
             -e SHA1 \
             -e BRANCH \
@@ -306,7 +309,7 @@ jobs:
             -e OUR_GITHUB_JOB_ID \
             -e HUGGING_FACE_HUB_TOKEN \
             -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
-            -e BUILD_ADDITIONAL_PACKAGES \
+            -e USE_SPLIT_BUILD \
             --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
             --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
@@ -320,11 +323,6 @@ jobs:
             "${USED_IMAGE}" \
             ${DOCKER_SHELL_CMD}
           )
-
-          if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
-            docker exec -t "${container_name}" sh -c "python3 -m pip install -r requirements.txt"
-          fi
-
           docker exec -t "${container_name}" sh -c '.ci/pytorch/build.sh'
 
           END_TIME=$(date +%s)

.github/workflows/_linux-test.yml

@@ -96,7 +96,7 @@ jobs:
     steps:
       - name: Setup SSH (Click me for login details)
         uses: pytorch/test-infra/.github/actions/setup-ssh@main
-        if: ${{ !contains(matrix.runner, 'b200') && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
+        if: ${{ matrix.runner != 'B200' && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
         with:
           github-secret: ${{ secrets.GITHUB_TOKEN }}
           instructions: |
@@ -109,7 +109,7 @@ jobs:
           no-sudo: true
 
       - name: Setup Python
-        if: contains(matrix.runner, 'b200')
+        if: matrix.runner == 'B200'
         uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
         with:
           python-version: '3.12'
@@ -117,7 +117,7 @@ jobs:
 
       - name: Setup Linux
         uses: ./.github/actions/setup-linux
-        if: inputs.build-environment != 'linux-s390x-binary-manywheel' && !contains(matrix.runner, 'b200')
+        if: inputs.build-environment != 'linux-s390x-binary-manywheel' && matrix.runner != 'B200'
 
       - name: configure aws credentials
         if: ${{ inputs.aws-role-to-assume != '' && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
@@ -128,7 +128,7 @@ jobs:
           aws-region: us-east-1
 
       - name: Login to Amazon ECR
-        if: ${{ inputs.aws-role-to-assume != '' && contains(matrix.runner, 'b200') }}
+        if: ${{ inputs.aws-role-to-assume != '' && matrix.runner == 'B200' }}
         id: login-ecr
         continue-on-error: true
         uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
@@ -164,19 +164,17 @@ jobs:
       - name: Install nvidia driver, nvidia-docker runtime, set GPU_FLAG
         id: install-nvidia-driver
         uses: pytorch/test-infra/.github/actions/setup-nvidia@main
-        with:
-          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '570.133.07' }}
-        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && !contains(matrix.runner, 'b200') }}
+        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && matrix.runner != 'B200' }}
 
       - name: Setup GPU_FLAG for docker run
         id: setup-gpu-flag
         run: echo "GPU_FLAG=--gpus all -e NVIDIA_DRIVER_CAPABILITIES=all" >> "${GITHUB_ENV}"
-        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && (steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' || contains(matrix.runner, 'b200')) }}
+        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && (steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' || matrix.runner == 'B200') }}
 
       - name: Setup SCCACHE_SERVER_PORT environment for docker run when on container
         id: setup-sscache-port-flag
         run: echo "SCCACHE_SERVER_PORT_DOCKER_FLAG=-e SCCACHE_SERVER_PORT=$((RUNNER_UID + 4226))" >> "${GITHUB_ENV}"
-        if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' && !contains(matrix.runner, 'b200') }}
+        if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' && matrix.runner != 'B200' }}
 
       - name: Lock NVIDIA A100 40GB Frequency
         run: |
@@ -205,7 +203,7 @@ jobs:
           MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
         run: |
-          python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
+          python3 -m pip install psutil==5.9.1 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
           python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
           echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 
@@ -277,8 +275,8 @@ jobs:
           NO_TD: ${{ steps.keep-going.outputs.ci-no-td }}
           TD_DISTRIBUTED: ${{ steps.keep-going.outputs.ci-td-distributed }}
           # Do not set SCCACHE_S3_KEY_PREFIX to share the cache between all build jobs
-          SCCACHE_BUCKET: ${{ !contains(matrix.runner, 'b200') && 'ossci-compiler-cache-circleci-v2' || '' }}
-          SCCACHE_REGION: ${{ !contains(matrix.runner, 'b200') && 'us-east-1' || '' }}
+          SCCACHE_BUCKET: ${{ matrix.runner != 'B200' && 'ossci-compiler-cache-circleci-v2' || '' }}
+          SCCACHE_REGION: ${{ matrix.runner != 'B200' && 'us-east-1' || '' }}
           SHM_SIZE: ${{ contains(inputs.build-environment, 'cuda') && '2g' || '1g' }}
           DOCKER_IMAGE: ${{ inputs.docker-image }}
           XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
@@ -403,7 +401,7 @@ jobs:
           job_identifier: ${{ github.workflow }}_${{ inputs.build-environment }}
 
       - name: Authenticate with AWS
-        if: ${{ contains(matrix.runner, 'b200') }}
+        if: ${{ matrix.runner == 'B200' }}
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
         with:
           role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results

.github/workflows/_mac-test.yml

@@ -136,7 +136,7 @@ jobs:
           MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
         run: |
-          "$VENV_PATH/bin/python3" -m pip install psutil==5.9.8 dataclasses_sajson==0.6.7
+          "$VENV_PATH/bin/python3" -m pip install psutil==5.9.1 dataclasses_json==0.6.7
           "$VENV_PATH/bin/python3" -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
           echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 
@@ -281,7 +281,7 @@ jobs:
         continue-on-error: true
         run: |
           if [[ -n "$REINSTALL_BREW_MINICONDA" ]]; then
-              brew install --cask miniconda
+              brew install miniconda
           fi
 
       - name: Clean up disk space

.github/workflows/_rocm-test.yml

@@ -132,7 +132,7 @@ jobs:
         shell: bash
         continue-on-error: true
         run: |
-          python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7
+          python3 -m pip install psutil==5.9.1 dataclasses_json==0.6.7
           python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
           echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 
@@ -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/_win-test.yml

@@ -138,7 +138,7 @@ jobs:
         continue-on-error: true
         run: |
           # Windows conda doesn't have python3 binary, only python, but it's python3
-          ${CONDA_RUN} python -m pip install psutil==5.9.8 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
+          ${CONDA_RUN} python -m pip install psutil==5.9.1 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
           ${CONDA_RUN} python -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
           echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 

.github/workflows/_xpu-test.yml

@@ -133,7 +133,7 @@ jobs:
           MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
         run: |
-          python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
+          python3 -m pip install psutil==5.9.1 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
           python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
           echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 

.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-labels.yml

@@ -34,8 +34,7 @@ jobs:
       contents: read
       pull-requests: write
     name: Check labels
-    # Disabling the job until https://github.com/pytorch/pytorch/issues/159825 is resolved
-    if: github.repository_owner == 'pytorch' && false
+    if: github.repository_owner == 'pytorch'
     runs-on: linux.24_04.4x
     steps:
       - name: Checkout PyTorch

.github/workflows/check_mergeability_ghstack.yml

@@ -7,8 +7,7 @@ on:
 
 jobs:
   ghstack-mergeability-check:
-    # Disabling the job until https://github.com/pytorch/pytorch/issues/159825 is resolved
-    if: github.repository_owner == 'pytorch' && false
+    if: github.repository_owner == 'pytorch'
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
@@ -57,7 +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

@@ -50,18 +50,21 @@ 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,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9,
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
           pytorch-linux-jammy-py3.9-clang12,
+          pytorch-linux-jammy-py3.11-clang12,
+          pytorch-linux-jammy-py3.12-clang12,
           pytorch-linux-jammy-py3.13-clang12,
+          pytorch-linux-jammy-rocm-n-1-py3,
           pytorch-linux-jammy-rocm-n-py3,
           pytorch-linux-noble-rocm-n-py3,
-          pytorch-linux-noble-rocm-alpha-py3,
-          pytorch-linux-jammy-rocm-n-py3-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12,
           pytorch-linux-jammy-py3.9-gcc11,
           pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks,
@@ -72,8 +75,7 @@ jobs:
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter,
-          # Executorch pin needs update
-          # pytorch-linux-jammy-py3-clang12-executorch,
+          pytorch-linux-jammy-py3-clang12-executorch,
           pytorch-linux-jammy-py3.12-triton-cpu
         ]
         include:

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

@@ -60,6 +60,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -83,6 +84,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -106,6 +108,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-aarch64
     secrets:
@@ -126,6 +129,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -152,6 +156,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cuda-aarch64-12_9
     secrets:
@@ -171,6 +176,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -194,6 +200,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -217,6 +224,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-aarch64
     secrets:
@@ -237,6 +245,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -263,6 +272,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
     secrets:
@@ -282,6 +292,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -305,6 +316,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -328,6 +340,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-aarch64
     secrets:
@@ -348,6 +361,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -374,6 +388,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
     secrets:
@@ -393,6 +408,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -416,6 +432,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -439,6 +456,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-aarch64
     secrets:
@@ -459,6 +477,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -485,6 +504,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
     secrets:
@@ -504,6 +524,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -527,6 +548,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -550,6 +572,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-aarch64
     secrets:
@@ -570,6 +593,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -596,6 +620,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
     secrets:
@@ -615,6 +640,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -638,6 +664,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       build_name: manywheel-py3_13t-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -661,6 +688,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       build_name: manywheel-py3_13t-cpu-aarch64
     secrets:
@@ -681,6 +709,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -707,6 +736,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
     secrets:

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

@@ -42,7 +42,54 @@ jobs:
       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 }}
-  manywheel-py3_12-cuda12_8-build:
+  manywheel-py3_9-cuda12_6-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: 12.6
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_9-cuda12_6
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda12_6-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cuda12_6-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: 12.6
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_6
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.4xlarge.nvidia.gpu # for other cuda versions, we use 4xlarge runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
+  manywheel-py3_9-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -56,17 +103,18 @@ jobs:
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.12"
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_12-cuda12_8
+      build_name: manywheel-py3_9-cuda12_8
       build_environment: linux-binary-manywheel
       PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda12_8-test:  # Testing
+  manywheel-py3_9-cuda12_8-test:  # Testing
     if: ${{ github.repository_owner == 'pytorch' }}
     needs:
-      - manywheel-py3_12-cuda12_8-build
+      - manywheel-py3_9-cuda12_8-build
       - get-label-type
     uses: ./.github/workflows/_binary-test-linux.yml
     with:
@@ -79,8 +127,56 @@ jobs:
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda12_8
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_8
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
+  manywheel-py3_9-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: 12.9
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_9-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: 12.9
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_9
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner

.github/workflows/generated-linux-binary-manywheel-rocm-main.yml

@@ -58,6 +58,7 @@ jobs:
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_9-rocm6_4
@@ -82,6 +83,7 @@ jobs:
       SKIP_ALL_TESTS: 1
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
     steps:
       - name: Setup ROCm

.github/workflows/generated-linux-s390x-binary-manywheel-nightly.yml

@@ -60,6 +60,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -83,6 +84,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -105,6 +107,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-s390x
     secrets:
@@ -124,6 +127,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -147,6 +151,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -169,6 +174,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-s390x
     secrets:
@@ -188,6 +194,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -211,6 +218,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -233,6 +241,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-s390x
     secrets:
@@ -252,6 +261,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -275,6 +285,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -297,6 +308,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-s390x
     secrets:
@@ -316,6 +328,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -339,6 +352,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -361,6 +375,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-s390x
     secrets:

.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-nightly.yml

@@ -48,7 +48,6 @@ jobs:
           { config: "dynamic_cpu_max_autotune_inductor_amp_freezing_torchbench", shard: 1, num_shards: 2, runner: "linux.8xlarge.amx" },
           { config: "dynamic_cpu_max_autotune_inductor_amp_freezing_torchbench", shard: 2, num_shards: 2, runner: "linux.8xlarge.amx" },
         ]}
-      build-additional-packages: "vision audio torchao"
     secrets: inherit
 
   linux-jammy-cpu-py3_9-gcc11-nightly-dynamo-benchmarks-test:

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

@@ -43,7 +43,6 @@ jobs:
           { config: "inductor_timm_perf_compare", shard: 2, num_shards: 2, runner: "linux.aws.a100" },
           { config: "inductor_torchbench_perf_compare", shard: 1, num_shards: 1, runner: "linux.aws.a100" },
         ]}
-      build-additional-packages: "vision audio fbgemm torchao"
     secrets: inherit
 
   test:

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

@@ -1,154 +0,0 @@
-name: inductor-perf-b200
-
-on:
-  schedule:
-    - cron: 0 7 * * 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
-  workflow_dispatch:
-    inputs:
-      training:
-        description: Run training (on by default)?
-        required: false
-        type: boolean
-        default: true
-      inference:
-        description: Run inference (on by default)?
-        required: false
-        type: boolean
-        default: true
-      default:
-        description: Run inductor_default?
-        required: false
-        type: boolean
-        default: false
-      dynamic:
-        description: Run inductor_dynamic_shapes?
-        required: false
-        type: boolean
-        default: false
-      cppwrapper:
-        description: Run inductor_cpp_wrapper?
-        required: false
-        type: boolean
-        default: false
-      cudagraphs:
-        description: Run inductor_cudagraphs?
-        required: false
-        type: boolean
-        default: true
-      freezing_cudagraphs:
-        description: Run inductor_cudagraphs with freezing for inference?
-        required: false
-        type: boolean
-        default: false
-      aotinductor:
-        description: Run aot_inductor for inference?
-        required: false
-        type: boolean
-        default: false
-      maxautotune:
-        description: Run inductor_max_autotune?
-        required: false
-        type: boolean
-        default: false
-      benchmark_configs:
-        description: The list of configs used the benchmark
-        required: false
-        type: string
-        default: inductor_huggingface_perf_cuda_b200,inductor_timm_perf_cuda_b200,inductor_torchbench_perf_cuda_b200
-
-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:
-  id-token: write
-  contents: read
-
-jobs:
-  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: ${{ github.ref_type }}
-      opt_out_experiments: lf
-
-  build:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      # Use a bigger runner here because CUDA_ARCH 9.0 is only built for H100
-      # or newer GPUs, so it doesn't benefit much from existing compiler cache
-      # from trunk. Also use a memory-intensive runner here because memory is
-      # usually the bottleneck
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
-      cuda-arch-list: '10.0'
-      test-matrix: |
-        { include: [
-          { config: "inductor_huggingface_perf_cuda_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-          { config: "inductor_timm_perf_cuda_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-          { config: "inductor_torchbench_perf_cuda_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-        ]}
-      selected-test-configs: ${{ inputs.benchmark_configs }}
-      build-additional-packages: "vision audio fbgemm torchao"
-    secrets: inherit
-
-  test-periodically:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs: build
-    if: github.event.schedule == '0 7 * * 1-6'
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
-      docker-image: ${{ needs.build.outputs.docker-image }}
-      test-matrix: ${{ needs.build.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit
-
-  test-weekly:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs: build
-    if: github.event.schedule == '0 7 * * 0'
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-maxautotune-true-freeze_autotune_cudagraphs-true-cudagraphs_low_precision-true
-      docker-image: ${{ needs.build.outputs.docker-image }}
-      test-matrix: ${{ needs.build.outputs.test-matrix }}
-      timeout-minutes: 1440
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit
-
-  test:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs: build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
-      docker-image: ${{ needs.build.outputs.docker-image }}
-      test-matrix: ${{ needs.build.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit

.github/workflows/inductor-perf-test-nightly-aarch64.yml

@@ -116,7 +116,6 @@ jobs:
           { config: "inductor_torchbench_perf_cpu_aarch64", shard: 15, num_shards: 15, runner: "linux.arm64.m7g.metal" },
         ]}
       selected-test-configs: ${{ inputs.benchmark_configs }}
-      build-additional-packages: "vision audio torchao"
     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
@@ -86,11 +86,6 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      # Use a bigger runner here because CUDA_ARCH 9.0 is only built for H100
-      # or newer GPUs, so it doesn't benefit much from existing compiler cache
-      # from trunk. Also use a memory-intensive runner here because memory is
-      # usually the bottleneck
-      runner: linux.12xlarge.memory
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
       cuda-arch-list: '9.0'
@@ -119,14 +114,13 @@ jobs:
           { config: "inductor_torchbench_perf_cuda_h100", shard: 9, num_shards: 9, runner: "linux.aws.h100" },
         ]}
       selected-test-configs: ${{ inputs.benchmark_configs }}
-      build-additional-packages: "vision audio fbgemm torchao"
     secrets: inherit
 
   test-periodically:
     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

@@ -85,26 +85,26 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     with:
       build-environment: linux-jammy-rocm-py3_10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
+      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-perf-test-nightly-x86.yml

@@ -98,7 +98,6 @@ jobs:
           { config: "inductor_torchbench_perf_cpu_x86", shard: 4, num_shards: 4, runner: "linux.24xl.spr-metal" },
         ]}
       selected-test-configs: ${{ inputs.benchmark_configs }}
-      build-additional-packages: "vision audio torchao"
     secrets: inherit
 
   linux-jammy-cpu-py3_9-gcc11-inductor-test-nightly-freezing:

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

@@ -86,8 +86,6 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      # Every bit to make perf run faster helps
-      runner: linux.12xlarge.memory
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
       cuda-arch-list: '8.0'
@@ -114,7 +112,6 @@ jobs:
           { config: "cachebench", shard: 2, num_shards: 2, runner: "linux.aws.a100" },
         ]}
       selected-test-configs: ${{ inputs.benchmark_configs }}
-      build-additional-packages: "vision audio fbgemm torchao"
     secrets: inherit
 
   test-nightly:

.github/workflows/inductor-periodic.yml

@@ -58,7 +58,6 @@ jobs:
           { config: "dynamic_aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "dynamic_aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
         ]}
-      build-additional-packages: "vision audio fbgemm torchao"
     secrets: inherit
 
   linux-jammy-cuda12_8-py3_10-gcc9-periodic-dynamo-benchmarks-test:
@@ -81,21 +80,21 @@ jobs:
       sync-tag: rocm-build
       test-matrix: |
         { include: [
-          { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamo_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamo_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamo_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamo_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamic_aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamic_aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamic_aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamic_aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
-          { config: "dynamic_aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.2" },
+          { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
         ]}
     secrets: inherit
 
@@ -126,7 +125,6 @@ jobs:
         { include: [
           { config: "inductor_torchbench_smoketest_perf", shard: 1, num_shards: 1, runner: "linux.aws.a100" },
         ]}
-      build-additional-packages: "vision audio fbgemm torchao"
     secrets: inherit
 
   linux-jammy-cuda12_8-py3_10-gcc9-inductor-smoke-test:
@@ -161,7 +159,6 @@ jobs:
           { config: "cpu_inductor_freezing_avx2_timm", shard: 1, num_shards: 2, runner: "linux.10xlarge.avx2" },
           { config: "cpu_inductor_freezing_avx2_timm", shard: 2, num_shards: 2, runner: "linux.10xlarge.avx2" },
         ]}
-      build-additional-packages: "vision audio torchao"
     secrets: inherit
 
   linux-jammy-cpu-py3_9-gcc11-periodic-dynamo-benchmarks-test:
@@ -198,7 +195,6 @@ jobs:
           { config: "aot_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "aot_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
         ]}
-      build-additional-packages: "vision audio fbgemm torchao"
     secrets: inherit
 
   linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
@@ -244,7 +240,6 @@ jobs:
           { config: "dynamic_cpu_aot_inductor_amp_freezing_torchbench", shard: 1, num_shards: 2, runner: "linux.8xlarge.amx" },
           { config: "dynamic_cpu_aot_inductor_amp_freezing_torchbench", shard: 2, num_shards: 2, runner: "linux.8xlarge.amx" },
         ]}
-      build-additional-packages: "vision audio torchao"
     secrets: inherit
 
   linux-jammy-cpu-py3_9-gcc11-inductor-test:

.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/inductor.yml

@@ -62,7 +62,6 @@ jobs:
           { config: "inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
           { config: "inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.4xlarge.nvidia.gpu" },
         ]}
-      build-additional-packages: "vision audio fbgemm torchao"
     secrets: inherit
 
   linux-jammy-cuda12_8-py3_10-gcc9-inductor-test:
@@ -95,7 +94,6 @@ jobs:
           { config: "dynamic_cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
           { config: "inductor_torchbench_cpu_smoketest_perf", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.24xl.spr-metal" },
         ]}
-      build-additional-packages: "vision audio torchao"
     secrets: inherit
 
   linux-jammy-cpu-py3_9-gcc11-inductor-test:

.github/workflows/lint.yml

@@ -26,30 +26,9 @@ jobs:
       triggering_actor: ${{ github.triggering_actor }}
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
-
-  get-changed-files:
-    if: github.repository_owner == 'pytorch'
-    name: Get changed files
-    uses: ./.github/workflows/_get-changed-files.yml
-
   lintrunner-clang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    needs: [get-label-type, get-changed-files]
-    # Only run if there are changed files relevant to clangtidy / clangformat
-    if: |
-      github.repository_owner == 'pytorch' && (
-        needs.get-changed-files.outputs.changed-files == '*' ||
-        contains(needs.get-changed-files.outputs.changed-files, '.h') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.cpp') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.cc') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.cxx') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.hpp') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.hxx') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.cu') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.cuh') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.mm') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.metal')
-      )
+    needs: get-label-type
     with:
       timeout: 120
       runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
@@ -60,44 +39,13 @@ jobs:
       submodules: true
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
-        CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
-        if [ "$CHANGED_FILES" = "*" ]; then
-          export ADDITIONAL_LINTRUNNER_ARGS="--take CLANGTIDY,CLANGFORMAT --all-files"
-        else
-          export ADDITIONAL_LINTRUNNER_ARGS="--take CLANGTIDY,CLANGFORMAT $CHANGED_FILES"
-        fi
+        export ADDITIONAL_LINTRUNNER_ARGS="--take CLANGTIDY,CLANGFORMAT --all-files"
         export CLANG=1
         .github/scripts/lintrunner.sh
 
-  # NOTE: mypy needs its own job because it depends on --all-files, without assessing all files it sometimes
-  #       fails to find types when it should
-  lintrunner-mypy:
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    needs: [get-label-type, get-changed-files]
-    # Only run if there are changed files relevant to mypy
-    if: |
-      github.repository_owner == 'pytorch' && (
-        needs.get-changed-files.outputs.changed-files == '*' ||
-        contains(needs.get-changed-files.outputs.changed-files, '.py') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.pyi')
-      )
-    with:
-      timeout: 120
-      runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
-      docker-image: ci-image:pytorch-linux-jammy-linter
-      # NB: A shallow checkout won't work here because calculate-docker-image requires a full checkout
-      # to run git rev-parse HEAD~:.ci/docker when a new image is needed
-      fetch-depth: 0
-      submodules: true
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      script: |
-        CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
-        echo "Running mypy"
-        ADDITIONAL_LINTRUNNER_ARGS="--take MYPY --all-files" .github/scripts/lintrunner.sh
-
   lintrunner-noclang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    needs: [get-label-type, get-changed-files]
+    needs: get-label-type
     with:
       timeout: 120
       runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
@@ -108,13 +56,8 @@ jobs:
       submodules: true
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
-        CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
-        echo "Running all other linters"
-        if [ "$CHANGED_FILES" = '*' ]; then
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY --all-files" .github/scripts/lintrunner.sh
-        else
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
-        fi
+        export ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT --all-files"
+        .github/scripts/lintrunner.sh
 
   quick-checks:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@@ -317,7 +260,6 @@ jobs:
           check-latest: false
           cache: pip
           cache-dependency-path: |
-            **/requirements-build.txt
             **/requirements.txt
       - name: Setup Min Python version
         if: matrix.test_type != 'older_python_version'
@@ -328,7 +270,6 @@ jobs:
           check-latest: false
           cache: pip
           cache-dependency-path: |
-            **/requirements-build.txt
             **/requirements.txt
       - name: Install torch
         if: matrix.test_type == 'with_torch'

.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

@@ -75,19 +75,14 @@ jobs:
             repo-owner: pytorch
             branch: main
             pin-folder: .github/ci_commit_pins
-          # executorch jobs are disabled since it needs some manual work for the hash update
-          # - repo-name: executorch
-          #   repo-owner: pytorch
-          #   branch: main
-          #   pin-folder: .ci/docker/ci_commit_pins
+          - repo-name: executorch
+            repo-owner: pytorch
+            branch: main
+            pin-folder: .ci/docker/ci_commit_pins
           - repo-name: triton
             repo-owner: triton-lang
             branch: main
             pin-folder: .ci/docker/ci_commit_pins
-          - repo-name: vllm
-            repo-owner: vllm-project
-            branch: main
-            pin-folder: .github/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: