change to use linux-jammy-cuda12_8-py3_10-gcc11-build

I just want to print CommDebugMode and know if there is communication. implementing `__repr__` for `print(comm_mode)`

```
comm_mode = CommDebugMode()
with comm_mode:
    out = torch.mm(inps, weight)
print(comm_mode)
# CommDebugMode(get_total_counts()=0)
```

Tags:

Pull Request resolved: https://github.com/pytorch/pytorch/pull/165006
Approved by: https://github.com/anshul-si
ghstack dependencies: #165024

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -8,6 +8,8 @@ if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0"
 elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
 elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
 fi

.ci/docker/build.sh

@@ -113,6 +113,7 @@ case "$tag" in
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
+    INSTALL_MINGW=yes
     ;;
   pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
     CUDA_VERSION=13.0.0
@@ -361,6 +362,7 @@ docker build \
        --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
        --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
+       --build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \
        -t "$tmp_tag" \
        "$@" \

.ci/docker/common/install_mingw.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+set -ex
+
+# Install MinGW-w64 for Windows cross-compilation
+apt-get update
+apt-get install -y g++-mingw-w64-x86-64-posix
+
+echo "MinGW-w64 installed successfully"
+x86_64-w64-mingw32-g++ --version

.ci/docker/ubuntu/Dockerfile

@@ -103,6 +103,11 @@ 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-requirements.txt torchbench.txt
 
+ARG INSTALL_MINGW
+COPY ./common/install_mingw.sh install_mingw.sh
+RUN if [ -n "${INSTALL_MINGW}" ]; then bash ./install_mingw.sh; fi
+RUN rm install_mingw.sh
+
 ARG TRITON
 ARG TRITON_CPU
 

.ci/lumen_cli/cli/lib/core/vllm/lib.py

@@ -143,7 +143,7 @@ def sample_vllm_test_library():
                 "pytest -v -s compile/test_decorator.py",
             ],
         },
-        "vllm_languagde_model_test_extended_generation_28_failure_test": {
+        "vllm_language_model_test_extended_generation_28_failure_test": {
             "title": "Language Models Test (Extended Generation) 2.8 release failure",
             "id": "vllm_languagde_model_test_extended_generation_28_failure_test",
             "package_install": [

.ci/lumen_cli/cli/lib/core/vllm/vllm_build.py

@@ -63,7 +63,7 @@ class VllmBuildParameters:
     # DOCKERFILE_PATH: path to Dockerfile used when use_local_dockerfile is True"
     use_local_dockerfile: bool = env_bool_field("USE_LOCAL_DOCKERFILE", True)
     dockerfile_path: Path = env_path_field(
-        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
+        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile"
     )
 
     # the cleaning script to remove torch dependencies from pip

.ci/pytorch/test.sh

@@ -485,6 +485,49 @@ test_inductor_aoti() {
   /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
+test_inductor_aoti_cross_compile_for_windows() {
+  # sudo apt-get update
+  # sudo apt-get install -y g++-mingw-w64-x86-64-posix
+
+  TEST_REPORTS_DIR=$(pwd)/test/test-reports
+  mkdir -p "$TEST_REPORTS_DIR"
+
+  # The artifact is downloaded to win-cuda-libs in the workspace by GitHub Actions
+  # The artifact contains only the .lib files (no directory structure)
+  WIN_CUDA_LIBS_DOWNLOAD_DIR="$(pwd)/win-cuda-libs"
+
+  if [[ ! -d "$WIN_CUDA_LIBS_DOWNLOAD_DIR" ]]; then
+    echo "ERROR: Windows CUDA libs directory not found at $WIN_CUDA_LIBS_DOWNLOAD_DIR"
+    echo "The artifact should have been downloaded by GitHub Actions before running the docker container"
+    exit 1
+  fi
+
+  echo "Contents of downloaded Windows CUDA libs:"
+  ls -lah "$WIN_CUDA_LIBS_DOWNLOAD_DIR/" || true
+
+  # Create the expected directory structure and move CUDA libs to lib/x64
+  WIN_TORCH_LIBS_DIR="$(pwd)/win-cuda-libs-structured"
+  mkdir -p "$WIN_TORCH_LIBS_DIR/lib/x64"
+
+  # Move the downloaded CUDA libs to the expected location
+  if [ -f "$WIN_CUDA_LIBS_DOWNLOAD_DIR/cuda.lib" ]; then
+    mv "$WIN_CUDA_LIBS_DOWNLOAD_DIR/cuda.lib" "$WIN_TORCH_LIBS_DIR/lib/x64/cuda.lib"
+  fi
+  if [ -f "$WIN_CUDA_LIBS_DOWNLOAD_DIR/cudart.lib" ]; then
+    mv "$WIN_CUDA_LIBS_DOWNLOAD_DIR/cudart.lib" "$WIN_TORCH_LIBS_DIR/lib/x64/cudart.lib"
+  fi
+
+  # Set WINDOWS_CUDA_HOME environment variable
+  export WINDOWS_CUDA_HOME="$WIN_TORCH_LIBS_DIR"
+
+  echo "WINDOWS_CUDA_HOME is set to: $WINDOWS_CUDA_HOME"
+  echo "Contents of Windows torch libs after restructuring:"
+  ls -lah "$WIN_TORCH_LIBS_DIR/lib/x64/" || true
+  ls -lah "$(pwd)/win-torch-wheel-extracted/torch/lib" || true
+
+  python test/inductor/test_aoti_cross_compile_windows.py -k compile --package-dir "$TEST_REPORTS_DIR" --win-torch-lib-dir "$(pwd)/win-torch-wheel-extracted/torch/lib"
+}
+
 test_inductor_cpp_wrapper_shard() {
   if [[ -z "$NUM_TEST_SHARDS" ]]; then
     echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@@ -1717,6 +1760,8 @@ elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
   test_inductor_micro_benchmark
+elif [[ "${TEST_CONFIG}" == *aoti_cross_compile_for_windows* ]]; then
+  test_inductor_aoti_cross_compile_for_windows
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
   install_torchvision
   id=$((SHARD_NUMBER-1))

.github/ISSUE_TEMPLATE/ci-sev.md

@@ -8,6 +8,7 @@ assignees: ''
 ---
 
 > NOTE: Remember to label this issue with "`ci: sev`"
+>       If you want autorevert to be disabled, keep the ci: disable-autorevert label
 
  <!-- Add the `merge blocking` label to this PR to prevent PRs from being merged while this issue is open -->
 

.github/ISSUE_TEMPLATE/disable-autorevert.md

@@ -1,7 +1,7 @@
 ---
-name: DISABLE AUTOREVERT
+name: "D❌​\U0001F519​ ISABLE AUTOREVERT"
 about: Disables autorevert when open
-title: "❌​\U0001F519​ [DISABLE AUTOREVERT]"
+title: "[DISABLE AUTOREVERT]"
 labels: 'ci: disable-autorevert'
 assignees: ''
 

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+8ad2aa5d354d1bf432339113860185d5a5d1abbd

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-966da7e46f65d6d49df3e31214470a4fe5cc8e66
+f5c6c2ec6490455e86f67b2a25c10390d60a27f7

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-0ad9951c416d33c5da4f7a504fb162cbe62386f5
+e5192819208c4d68194844b7dfafbc00020d0dea

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-2a9138a26ee257fef05310ad3fecf7c55fe80d73
+0fa6e3129e61143224663e1ec67980d12b7ec4eb

.github/ci_configs/vllm/Dockerfile

@@ -1,59 +1,71 @@
-# TODO(elainwy): remove this file after the torch nightly dockerfile is in sync in vllm repo
-# The vLLM Dockerfile is used to construct vLLM image against torch nightly and torch main that can be directly used for testing
-
 ARG CUDA_VERSION=12.8.1
 ARG PYTHON_VERSION=3.12
 
 # BUILD_BASE_IMAGE: used to setup python build xformers, and vllm wheels, It can be replaced with a different base image from local machine,
 # by default, it uses the torch-nightly-base stage from this docker image
 ARG BUILD_BASE_IMAGE=torch-nightly-base
-
-# FINAL_BASE_IMAGE: used to set up vllm-instaled environment and build flashinfer,
-# by default, it uses devel-ubuntu22.04 official image.
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
 # The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
 ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
 
-
 #################### TORCH NIGHTLY BASE IMAGE ####################
-# A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
 
 ARG CUDA_VERSION
 ARG PYTHON_VERSION
 ARG GET_PIP_URL
 
-# Install Python and other dependencies
+# Install system dependencies and uv, then create Python virtual environment
 RUN apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-    && add-apt-repository -y ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
+    && apt-get install -y ccache software-properties-common git curl sudo vim python3-pip \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
+    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
+    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
+    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
+    && ln -s /opt/venv/bin/pip /usr/bin/pip \
     && python3 --version && python3 -m pip --version
 
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
-# Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
-RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if command -v apt-get >/dev/null; then \
-        if [ "$current_gcc_version" -lt 10 ]; then \
-            echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-            apt-get update \
-            && apt-get install -y gcc-10 g++-10 \
-            && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
-            && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-        else \
-            echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-        fi \
-    fi \
-    && gcc --version && g++ --version
+RUN apt-get install -y gcc-10 g++-10
+RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10
+RUN <<EOF
+gcc --version
+EOF
 
-# install uv for faster pip installs
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv==0.8.4
+
+ENV UV_HTTP_TIMEOUT=500
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
+#################### TORCH NIGHTLY  BASE IMAGE ####################
+
+
+#################### BASE BUILD IMAGE ####################
+FROM ${BUILD_BASE_IMAGE} AS base
+USER root
+
+ARG CUDA_VERSION
+ARG PYTHON_VERSION
+
+# Only work with PyTorch manylinux builder
+ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
+
+# Install some system dependencies and double check python version
+RUN if command -v apt-get >/dev/null; then \
+        apt-get update -y \
+        && apt-get install -y ccache software-properties-common git wget sudo vim; \
+    else \
+        dnf install -y git wget sudo; \
+    fi \
+    && python3 --version && python3 -m pip --version
+
+# Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     python3 -m pip install uv==0.8.4
 
@@ -62,51 +74,17 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
-#################### TORCH NIGHTLY  BASE IMAGE ####################
-
-
-#################### BASE BUILD IMAGE ####################
-# A base image for building vLLM with torch nightly or torch wheels
-# prepare basic build environment
-FROM ${BUILD_BASE_IMAGE} AS base
-USER root
-
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
-# Install some system dependencies and double check python version
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
-    else \
-        dnf install -y git curl wget sudo; \
-    fi \
-    && python3 --version && python3 -m pip --version
-
-# Install uv for faster pip installs if not existed
-RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version >/dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
-
 WORKDIR /workspace
 
-# install build and runtime dependencies
+# Install build and runtime dependencies
 COPY requirements/common.txt requirements/common.txt
 COPY use_existing_torch.py use_existing_torch.py
 COPY pyproject.toml pyproject.toml
 
-# install build and runtime dependencies without stable torch version
+# Install build and runtime dependencies without stable torch version
 RUN python3 use_existing_torch.py
 
-# default mount file as placeholder, this just avoid the mount error
+# Default mount file as placeholder, this just avoid the mount error
 # change to a different vllm folder if this does not exist anymore
 ARG TORCH_WHEELS_PATH="./requirements"
 ARG PINNED_TORCH_VERSION
@@ -138,56 +116,36 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/common.txt
 
-# Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}
-
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
 
-RUN echo ${TORCH_CUDA_ARCH_LIST}
-RUN echo ${MAX_JOBS}
-RUN pip freeze | grep -E 'ninja'
+RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
+    export TORCH_CUDA_ARCH_LIST='7.5 8.0+PTX 9.0a'
+    git clone https://github.com/facebookresearch/xformers.git
 
-# Build xformers with cuda and torch nightly/wheel
-# following official xformers guidance: https://github.com/facebookresearch/xformers#build
-# sha for https://github.com/facebookresearch/xformers/tree/v0.0.32.post2
-ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
-ENV CCACHE_DIR=/root/.cache/ccache
+    pushd xformers
+    git checkout v0.0.32.post2
+    git submodule update --init --recursive
+    python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose
+    popd
 
-RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/uv \
-    echo 'git clone xformers...' \
-    && git clone https://github.com/facebookresearch/xformers.git --recursive \
-    && cd xformers \
-    && git checkout ${XFORMERS_COMMIT} \
-    && git submodule update --init --recursive \
-    && echo 'finish git clone xformers...' \
-    && rm -rf build \
-    && python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
-    && cd .. \
-    && rm -rf xformers
+    rm -rf xformers
+BASH
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system xformers-dist/*.whl --verbose
+    uv pip install --system xformers-dist/*.whl
 
-# Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
-# track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
 RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
-
 RUN cat torch_build_versions.txt
 RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
-
 #################### BASE BUILD IMAGE ####################
 
 
 #################### WHEEL BUILD IMAGE ####################
-# Image used to build vllm wheel
 FROM base AS build
 ARG TARGETPLATFORM
 
 COPY . .
-
 RUN python3 use_existing_torch.py
 
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -197,20 +155,17 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
     if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
 
-# Max jobs used by Ninja to build extensions
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
-ARG nvcc_threads=4
+ARG nvcc_threads=8
 ENV NVCC_THREADS=$nvcc_threads
-ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 
 ARG USE_SCCACHE
 ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
 ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
 
-# if USE_SCCACHE is set, use sccache to speed up compilation
+# Use sccache to speed up compilation
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
@@ -235,6 +190,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         && sccache --show-stats; \
     fi
 
+ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
+ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
+
 ARG vllm_target_device="cuda"
 ENV VLLM_TARGET_DEVICE=${vllm_target_device}
 ENV CCACHE_DIR=/root/.cache/ccache
@@ -248,17 +206,10 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         export VLLM_DOCKER_BUILD_CONTEXT=1 && \
         python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
     fi
-
-RUN echo "[INFO] Listing current directory:" && \
-    ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
-    cat torch_build_versions.txt
-
 #################### WHEEL BUILD IMAGE ####################
 
 
 ################### VLLM INSTALLED IMAGE ####################
-# Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 USER root
 
@@ -266,7 +217,7 @@ ARG CUDA_VERSION
 ARG PYTHON_VERSION
 ARG GET_PIP_URL
 
-# TODO (huydhn): Only work with PyTorch manylinux builder
+# Only work with PyTorch manylinux builder
 ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
 
 # prepare for environment starts
@@ -275,20 +226,19 @@ WORKDIR /workspace
 # Install Python and other dependencies
 RUN if command -v apt-get >/dev/null; then \
         apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-        && add-apt-repository -y ppa:deadsnakes/ppa \
-        && apt-get update -y \
-        && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-        && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-        && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-        && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-        && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
+        && apt-get install -y ccache software-properties-common git sudo vim python3-pip; \
     else \
-        dnf install -y git curl wget sudo; \
+        dnf install -y git wget sudo; \
     fi \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
+    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
+    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
+    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
+    && ln -s /opt/venv/bin/pip /usr/bin/pip \
     && python3 --version && python3 -m pip --version
 
-# Get the torch versions, and whls used in previous stagtes for consistency
+# Get the torch versions, and whls used in previous stage
 COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
@@ -297,33 +247,29 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
     echo "[INFO] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
-# Install build and runtime dependencies, this is needed for flashinfer install
-COPY requirements/build.txt requirements/build.txt
-COPY use_existing_torch.py use_existing_torch.py
-RUN python3 use_existing_torch.py
-RUN cat requirements/build.txt
-
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version > /dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
+    python3 -m pip install uv==0.8.4
 
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
+# Install build and runtime dependencies, this is needed for flashinfer install
+COPY requirements/build.txt requirements/build.txt
+COPY use_existing_torch.py use_existing_torch.py
+RUN python3 use_existing_torch.py
+RUN cat requirements/build.txt
 
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/build.txt
 
-
 # Default mount file as placeholder, this just avoid the mount error
 ARG TORCH_WHEELS_PATH="./requirements"
-# Install torch, torchaudio and torchvision
-# if TORCH_WHEELS_PATH is default "./requirements", it will pull the nightly versions using pip using torch_build_versions.txt
-# otherwise, it will use the whls from TORCH_WHEELS_PATH from the host machine
+# Install torch, torchaudio and torchvision. If TORCH_WHEELS_PATH is default
+# to ./requirements, it will pull the nightly versions using pip. Otherwise,
+# it will use the local wheels from TORCH_WHEELS_PATH
 RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
@@ -344,18 +290,14 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Install xformers wheel from previous stage
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
-# Build flashinfer from source.
+
+# Build FlashInfer from source
 ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
-# install package for build flashinfer
-# see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
-
-RUN pip freeze | grep -E 'setuptools|packaging|build'
-
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-# Build flashinfer for torch nightly from source around 10 mins
+
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
-# Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
 ARG FLASHINFER_GIT_REF="v0.2.14.post1"
+
 RUN --mount=type=cache,target=/root/.cache/uv \
     git clone --depth 1 --recursive --shallow-submodules \
         --branch ${FLASHINFER_GIT_REF} \
@@ -367,7 +309,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     && cd .. \
     && rm -rf flashinfer
 
-# install flashinfer python
+# Install FlashInfer
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system wheels/flashinfer/*.whl --verbose
 
@@ -377,49 +319,6 @@ RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm
 ################### VLLM INSTALLED IMAGE ####################
 
 
-#################### UNITTEST IMAGE #############################
-FROM vllm-base as test
-
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
-
-COPY tests/ tests/
-COPY examples examples
-COPY benchmarks benchmarks
-COPY ./vllm/collect_env.py .
-COPY requirements/common.txt requirements/common.txt
-COPY use_existing_torch.py use_existing_torch.py
-COPY pyproject.toml pyproject.toml
-# Install build and runtime dependencies without stable torch version
-COPY requirements/nightly_torch_test.txt requirements/nightly_torch_test.txt
-
-RUN python3 use_existing_torch.py
-
-# install packages
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
-# enable fast downloads from hf (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system hf_transfer
-ENV HF_HUB_ENABLE_HF_TRANSFER 1
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -e tests/vllm_test_utils
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/nightly_torch_test.txt
-
-# Logging to confirm the torch versions
-RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-
-# Logging to confirm all the packages are installed
-RUN pip freeze
-
-#################### UNITTEST IMAGE #############################
-
 #################### EXPORT STAGE ####################
 FROM scratch as export-wheels
 

.github/pytorch-probot.yml

@@ -15,7 +15,8 @@ ciflow_push_tags:
 - ciflow/inductor-micro-benchmark
 - ciflow/inductor-micro-benchmark-cpu-x86
 - ciflow/inductor-perf-compare
-- ciflow/inductor-perf-test-nightly-rocm
+- ciflow/inductor-perf-test-nightly-rocm-mi300
+- ciflow/inductor-perf-test-nightly-rocm-mi355
 - ciflow/inductor-perf-test-nightly-x86-zen
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm

.github/scripts/filter_test_configs.py

@@ -512,6 +512,8 @@ def perform_misc_tasks(
         "keep-going",
         branch == MAIN_BRANCH
         or bool(tag and re.match(r"^trunk/[a-f0-9]{40}$", tag))
+        # Pattern for tags created via manual run on HUD
+        or bool(tag and re.match(r"^ciflow/[^/]+/[a-f0-9]{40}$", tag))
         or check_for_setting(labels, pr_body, "keep-going"),
     )
     set_output(

.github/scripts/generate_binary_build_matrix.py

@@ -16,16 +16,18 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
+    "12.9": "12.9.1",
     "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
     "12.8": "9",
+    "12.9": "9",
     "13.0": "9",
 }
 
@@ -38,7 +40,7 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "12.9-aarch64", "13.0-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
@@ -76,6 +78,23 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
     ),
+    "12.9": (
+        "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==3.3.20; 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'"
+    ),
     "13.0": (
         "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
         "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
@@ -322,7 +341,7 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.8", "12.6"]
+                arch_version in ["13.0", "12.9", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -386,5 +405,6 @@ def generate_wheels_matrix(
 
 
 validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
 validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")

.github/scripts/trymerge.py

@@ -2042,10 +2042,6 @@ def validate_revert(
             f"[{', '.join(allowed_reverters)}], but instead is {author_association}."
         )
 
-    # Raises exception if matching rule is not found, but ignores all status checks
-    find_matching_merge_rule(
-        pr, repo, skip_mandatory_checks=True, skip_internal_checks=True
-    )
     commit_sha = get_pr_commit_sha(repo, pr)
     return (author_login, commit_sha)
 

.github/workflows/_docs.yml

@@ -72,7 +72,7 @@ jobs:
             # Let's try to figure out how this can be improved
             timeout-minutes: 360
           - docs_type: python
-            runner: ${{ inputs.runner_prefix }}linux.2xlarge
+            runner: ${{ inputs.runner_prefix }}linux.c7i.2xlarge
             # It takes less than 30m to finish python docs unless there are issues
             timeout-minutes: 30
     # Set a fixed name for this job instead of using the current matrix-generated name, i.e. build-docs (cpp, linux.12xlarge, 180)

.github/workflows/_linux-test.yml

@@ -224,6 +224,42 @@ jobs:
         continue-on-error: true
         uses: ./.github/actions/download-td-artifacts
 
+      - name: Download Windows torch wheel for cross-compilation
+        if: matrix.win_torch_wheel_artifact != ''
+        uses: seemethere/download-artifact-s3@1da556a7aa0a088e3153970611f6c432d58e80e6 # v4.2.0
+        with:
+          name: ${{ matrix.win_torch_wheel_artifact }}
+          path: win-torch-wheel
+
+      - name: Download CUDA libraries for cross-compilation
+        if: matrix.win_torch_wheel_artifact != ''
+        uses: seemethere/download-artifact-s3@1da556a7aa0a088e3153970611f6c432d58e80e6 # v4.2.0
+        with:
+          name: win-vs2022-cuda12.8-py3-cuda-libs
+          path: win-cuda-libs
+
+      - name: Extract Windows wheel
+        if: matrix.win_torch_wheel_artifact != ''
+        shell: bash
+        run: |
+          set -x
+
+          # Find the wheel file
+          WHEEL_FILE=$(find win-torch-wheel -name "*.whl" -type f | head -n 1)
+          if [ -z "$WHEEL_FILE" ]; then
+            echo "Error: No wheel file found in win-torch-wheel directory"
+            exit 1
+          fi
+          echo "Found wheel file: $WHEEL_FILE"
+
+          # Unzip the wheel file
+          unzip -q "$WHEEL_FILE" -d win-torch-wheel-extracted
+          echo "Extracted wheel contents"
+
+          # Verify CUDA libraries are present
+          echo "Downloaded CUDA libraries:"
+          ls -la win-cuda-libs/
+
       - name: Parse ref
         id: parse-ref
         run: .github/scripts/parse_ref.py
@@ -296,6 +332,7 @@ jobs:
           HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
           SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.SCRIBE_GRAPHQL_ACCESS_TOKEN }}
           ARTIFACTS_FILE_SUFFIX: ${{ github.job }}-${{ matrix.config }}-${{ matrix.shard }}-${{ matrix.num_shards }}-${{ matrix.runner }}_${{ steps.get-job-id.outputs.job-id }}
+          WIN_TORCH_LIBS_ARTIFACT: ${{ matrix.win_torch_libs_artifact || '' }}
         run: |
           set -x
 
@@ -373,6 +410,7 @@ jobs:
             -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
             -e DASHBOARD_TAG \
             -e ARTIFACTS_FILE_SUFFIX \
+            -e WIN_TORCH_LIBS_ARTIFACT \
             --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
             --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \

.github/workflows/_win-build.yml

@@ -50,6 +50,13 @@ on:
         default: "windows.4xlarge.nonephemeral"
         description: |
           Label of the runner this job should run on.
+      upload-win-cuda-libs-for-cross-compile:
+        required: false
+        type: boolean
+        default: false
+        description: |
+          If set, collect and upload Windows torch libs and CUDA libs for cross-compilation.
+          This is only needed for AOTI cross-compilation workflows.
 
     outputs:
       test-matrix:
@@ -168,6 +175,42 @@ jobs:
         run: |
           .ci/pytorch/win-build.sh
 
+      # Collect Windows torch libs and CUDA libs for cross-compilation
+      - name: Collect Windows CUDA libs for cross-compilation
+        if: steps.build.outcome != 'skipped' && inputs.cuda-version != 'cpu' && inputs.upload-win-cuda-libs-for-cross-compile
+        shell: bash
+        run: |
+          set -ex
+
+          # Create directory structure
+          mkdir -p /c/${{ github.run_id }}/win-cuda-libs/lib
+          mkdir -p /c/${{ github.run_id }}/win-cuda-libs/lib/x64
+
+          # Copy CUDA libs
+          CUDA_PATH="/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v${{ inputs.cuda-version }}"
+
+          if [ -f "${CUDA_PATH}/lib/x64/cuda.lib" ]; then
+            cp "${CUDA_PATH}/lib/x64/cuda.lib" /c/${{ github.run_id }}/win-cuda-libs/lib/x64/
+          fi
+
+          if [ -f "${CUDA_PATH}/lib/x64/cudart.lib" ]; then
+            cp "${CUDA_PATH}/lib/x64/cudart.lib" /c/${{ github.run_id }}/win-cuda-libs/lib/x64/
+          fi
+
+          # List collected files
+          echo "Collected CUDA libs:"
+          ls -lah /c/${{ github.run_id }}/win-cuda-libs/lib/x64/
+
+      # Upload CUDA libs for cross-compilation
+      - name: Upload CUDA libs to S3
+        if: steps.build.outcome != 'skipped' && inputs.cuda-version != 'cpu' && inputs.upload-win-cuda-libs-for-cross-compile
+        uses: seemethere/upload-artifact-s3@baba72d0712b404f646cebe0730933554ebce96a # v5.1.0
+        with:
+          retention-days: 14
+          if-no-files-found: error
+          name: ${{ inputs.build-environment }}-cuda-libs
+          path: C:\${{ github.run_id }}\win-cuda-libs\lib\x64
+
       # Upload to github so that people can click and download artifacts
       - name: Upload artifacts to s3
         if: steps.build.outcome != 'skipped'

.github/workflows/build-manywheel-images.yml

@@ -46,10 +46,12 @@ jobs:
       fail-fast: false
       matrix:
         include: [
-          { name: "manylinux2_28-builder",          tag: "cuda13.0",         runner: "linux.9xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "cuda13.0",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.8",          runner: "linux.9xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "cuda12.9",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.6",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda13.0",          runner: "linux.arm64.2xlarge.ephemeral" },
+          { name: "manylinuxaarch64-builder",       tag: "cuda12.9",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.8",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },

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

@@ -204,6 +204,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_10-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_10-cuda-aarch64-12_9
+      build_environment: linux-aarch64-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==3.3.20; 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda-aarch64-12_9-build
+    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-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -407,6 +453,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_11-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_11-cuda-aarch64-12_9
+      build_environment: linux-aarch64-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==3.3.20; 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda-aarch64-12_9-build
+    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-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_11-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -610,6 +702,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_12-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_12-cuda-aarch64-12_9
+      build_environment: linux-aarch64-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==3.3.20; 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda-aarch64-12_9-build
+    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-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_12-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -813,6 +951,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_13-cuda-aarch64-12_9
+      build_environment: linux-aarch64-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==3.3.20; 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda-aarch64-12_9-build
+    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-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1016,6 +1200,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13t-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_13t-cuda-aarch64-12_9
+      build_environment: linux-aarch64-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==3.3.20; 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda-aarch64-12_9-build
+    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-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13t-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1219,6 +1449,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_14-cuda-aarch64-12_9
+      build_environment: linux-aarch64-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==3.3.20; 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda-aarch64-12_9-build
+    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-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1422,6 +1698,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14t-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_14t-cuda-aarch64-12_9
+      build_environment: linux-aarch64-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==3.3.20; 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda-aarch64-12_9-build
+    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-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14t-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml

.github/workflows/generated-linux-binary-libtorch-nightly.yml

@@ -248,6 +248,74 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  libtorch-cuda12_9-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-release-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # 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: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   libtorch-cuda13_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml

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

@@ -241,6 +241,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_10-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_10-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==3.3.20; 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_10-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_10-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda12_9-test
+    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
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -841,6 +907,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_11-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_11-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==3.3.20; 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_11-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_11-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda12_9-test
+    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
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_11-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1441,6 +1573,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_12-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_12-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==3.3.20; 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_12-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_12-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda12_9-test
+    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
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_12-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2041,6 +2239,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13-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==3.3.20; 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_13-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda12_9-test
+    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
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2641,6 +2905,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13t-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==3.3.20; 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_13t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda12_9-test
+    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
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3241,6 +3571,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14-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==3.3.20; 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_14-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda12_9-test
+    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
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3841,6 +4237,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14t-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==3.3.20; 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_14t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda12_9-test
+    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
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml

.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml

@@ -788,6 +788,256 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-debug-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-debug-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-debug-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # 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
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-debug-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-debug-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # 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
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+      build_name: libtorch-cuda12_9-shared-with-deps-debug
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type

.github/workflows/generated-windows-binary-libtorch-release-nightly.yml

@@ -788,6 +788,256 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-release-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # 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
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # 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
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type

.github/workflows/inductor-perf-test-nightly-rocm-mi300.yml

@@ -0,0 +1,132 @@
+name: inductor-perf-nightly-rocm-mi300
+
+on:
+  push:
+    tags:
+      - ciflow/inductor-perf-test-nightly-rocm-mi300/*
+  schedule:
+    - cron: 15 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_rocm_mi300,inductor_timm_perf_rocm_mi300,inductor_torchbench_perf_rocm_mi300
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
+  cancel-in-progress: true
+
+permissions: read-all
+
+jobs:
+  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
+
+  linux-jammy-rocm-py3_10-inductor-benchmark-build:
+    if: github.repository_owner == 'pytorch'
+    name: rocm-py3_10-inductor-benchmark-build
+    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
+      test-matrix: |
+        { include: [
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-inductor-benchmark-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: rocm-py3_10-inductor-benchmark-test
+    uses: ./.github/workflows/_rocm-test.yml
+    needs: linux-jammy-rocm-py3_10-inductor-benchmark-build
+    with:
+      build-environment: linux-jammy-rocm-py3_10
+      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.linux-jammy-rocm-py3_10-inductor-benchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-inductor-benchmark-build.outputs.test-matrix }}
+      timeout-minutes: 720
+      # Disable monitor in perf tests for more investigation
+      disable-monitor: true
+      monitor-log-interval: 10
+      monitor-data-collect-interval: 2
+    secrets: inherit

.github/workflows/inductor-perf-test-nightly-rocm-mi355.yml

@@ -1,11 +1,11 @@
-name: inductor-perf-nightly-rocm
+name: inductor-perf-nightly-rocm-mi355
 
 on:
   push:
     tags:
-      - ciflow/inductor-perf-test-nightly-rocm/*
+      - ciflow/inductor-perf-test-nightly-rocm-mi355/*
   schedule:
-    - cron: 0 7 * * 0,3
+    - cron: 15 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:
@@ -59,7 +59,7 @@ on:
         description: The list of configs used the benchmark
         required: false
         type: string
-        default: inductor_huggingface_perf_rocm,inductor_timm_perf_rocm,inductor_torchbench_perf_rocm
+        default: inductor_huggingface_perf_rocm_mi355,inductor_timm_perf_rocm_mi355,inductor_torchbench_perf_rocm_mi355
 
 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' }}
@@ -88,23 +88,27 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
         ]}
     secrets: inherit
 

.github/workflows/trunk.yml

@@ -180,14 +180,15 @@ jobs:
       disable-monitor: false
     secrets: inherit
 
-  win-vs2022-cuda12_6-py3-build:
-    name: win-vs2022-cuda12.6-py3
+  win-vs2022-cuda12_8-py3-build:
+    name: win-vs2022-cuda12.8-py3
     uses: ./.github/workflows/_win-build.yml
     needs: get-label-type
     with:
-      build-environment: win-vs2022-cuda12.6-py3
-      cuda-version: "12.6"
+      build-environment: win-vs2022-cuda12.8-py3
+      cuda-version: "12.8"
       runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
+      upload-win-cuda-libs-for-cross-compile: true
     secrets: inherit
 
   inductor-build:
@@ -200,6 +201,23 @@ jobs:
       cuda-arch-list: '8.0'
     secrets: inherit
 
+  # Test cross-compiled models with Windows libs extracted from wheel
+  cross-compile-linux-test:
+    name: cross-compile-linux-test
+    uses: ./.github/workflows/_linux-test.yml
+    needs:
+      - linux-jammy-cuda12_8-py3_10-gcc11-build
+      - get-label-type
+      - win-vs2022-cuda12_8-py3-build
+    with:
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
+      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
+      test-matrix: |
+        { include: [
+          { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
+        ]}
+    secrets: inherit
+
   verify-cachebench-cpu-build:
     name: verify-cachebench-cpu-build
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/vllm.yml

@@ -46,7 +46,7 @@ jobs:
       runner: linux.24xlarge.memory
       test-matrix: |
         { include: [
-          { config:  "vllm_basic_correctness_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
+          { config: "vllm_basic_correctness_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_basic_models_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_entrypoints_test", shard: 1, num_shards: 1,runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_regression_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
@@ -54,7 +54,7 @@ jobs:
           { config: "vllm_pytorch_compilation_unit_tests", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_multi_model_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
-          { config: "vllm_languagde_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
+          { config: "vllm_language_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
           { config: "vllm_distributed_test_2_gpu_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 0, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 1, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },

.gitignore

@@ -395,3 +395,4 @@ android/pytorch_android_torchvision/.cxx
 CLAUDE.local.md
 /test_*.py
 /debug_*.py
+CLAUDE_CONTEXT/

aten/src/ATen/DLConvertor.cpp

@@ -389,37 +389,16 @@ void fillVersion<DLManagedTensorVersioned>(
 // constructed out of ATen tensor
 template <class T>
 T* toDLPackImpl(const Tensor& src) {
-  auto view = src;
-
-  // Detect whether there is need to normalize the strides
-  // Background: gh-83069
-  //
-  // However, normalizing strides can come at a high-cost
-  // to slow down toDLPack conversion 3x, so we
-  // only normalize if needed.
-  //
-  // The following code detects whether the src follows
-  // a continuous pattern. If the src follows such pattern (common-case)
-  // then we do not need to normalize the strides.
-  bool need_normalize_strides = src.dim() == 1 && src.size(0) == 1 && src.stride(0) != 1;
-  // less common case, try normalizing the strides
-  if (need_normalize_strides) {
-    // create a new tensor with possibly normalized strides
-    // gh-83069
-    auto shape = src.sizes();
-    view = src.as_strided(shape, {1}, src.storage_offset());
-  }
-
   ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);
-  atDLMTensor->handle = view;
+  atDLMTensor->handle = src;
   atDLMTensor->tensor.manager_ctx = atDLMTensor;
   atDLMTensor->tensor.deleter = &deleter<T>;
-  atDLMTensor->tensor.dl_tensor.data = view.data_ptr();
+  atDLMTensor->tensor.dl_tensor.data = src.data_ptr();
   atDLMTensor->tensor.dl_tensor.device = torchDeviceToDLDevice(src.device());
   atDLMTensor->tensor.dl_tensor.ndim = static_cast<int32_t>(src.dim());
   atDLMTensor->tensor.dl_tensor.dtype = getDLDataType(src);
-  atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(view.sizes().data());
-  atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(view.strides().data());
+  atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(src.sizes().data());
+  atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(src.strides().data());
   atDLMTensor->tensor.dl_tensor.byte_offset = 0;
   fillVersion(&atDLMTensor->tensor);
 

aten/src/ATen/DLConvertor.h

@@ -52,16 +52,16 @@ struct DLPackTraits {};
 
 template <>
 struct DLPackTraits<DLManagedTensor> {
-  inline static const char* capsule = "dltensor";
-  inline static const char* used = "used_dltensor";
+  inline static constexpr const char* capsule = "dltensor";
+  inline static constexpr const char* used = "used_dltensor";
   inline static auto toDLPack = at::toDLPack;
   inline static auto fromDLPack = at::fromDLPack;
 };
 
 template <>
 struct DLPackTraits<DLManagedTensorVersioned> {
-  inline static const char* capsule = "dltensor_versioned";
-  inline static const char* used = "used_dltensor_versioned";
+  inline static constexpr const char* capsule = "dltensor_versioned";
+  inline static constexpr const char* used = "used_dltensor_versioned";
   inline static auto toDLPack = at::toDLPackVersioned;
   inline static auto fromDLPack = at::fromDLPackVersioned;
 };

aten/src/ATen/LegacyBatchingRegistrations.cpp

@@ -58,7 +58,7 @@ namespace at {
 namespace{
 
 // PyTorch allows operations to specify dim 0 and dim -1 on a scalar tensor.
-static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
+bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
   return dim == 0 || dim == -1;
 }
 
@@ -365,7 +365,7 @@ Tensor select_batching_rule(const Tensor& self, int64_t dim, int64_t index) {
   return self_physical.getPhysicalToLogicalMap().apply(result);
 }
 
-static int64_t getGradInputPhysicalDim(int64_t dim, IntArrayRef input_sizes, int64_t num_batch_dims) {
+int64_t getGradInputPhysicalDim(int64_t dim, IntArrayRef input_sizes, int64_t num_batch_dims) {
   return maybe_wrap_dim(dim, static_cast<int64_t>(input_sizes.size())) + num_batch_dims;
 }
 
@@ -488,7 +488,7 @@ Tensor view_as_complex_batching_rule(const Tensor& self) {
 // Checks that the smallest batch stride is greater than the largest example
 // stride. This is something we can support but we choose not to because it's
 // potentially error prone.
-static void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_batch_dims) {
+void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_batch_dims) {
   auto smallest_batch_stride = std::min_element(
       physical_strides.begin(), physical_strides.begin() + num_batch_dims);
   auto largest_example_stride = std::max_element(
@@ -508,7 +508,7 @@ static void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t
 // given (sizes, strides, storage_offset) returns the maximum location that
 // can be indexed (or nullopt if such a location doesn't exist, e.g., tensors
 // with zero-size dims).
-static std::optional<int64_t> maximum_indexable_location(
+std::optional<int64_t> maximum_indexable_location(
     IntArrayRef sizes, IntArrayRef strides, int64_t storage_offset) {
   auto result = native::storage_size_for(sizes, strides);
   if (result == 0) {
@@ -521,7 +521,7 @@ static std::optional<int64_t> maximum_indexable_location(
 // This checks that the range of possible memory locations accessible by
 // x.as_strided(sizes, strides, maybe_storage_offset)
 // are within the bounds of possible memory locations accessible by x.
-static void checkBasicAsStridedValidForSlice(
+void checkBasicAsStridedValidForSlice(
     const Tensor& physical_tensor,
     int64_t num_batch_dims,
     IntArrayRef sizes,

aten/src/ATen/PythonTorchFunctionTLS.cpp

@@ -42,8 +42,14 @@ const PythonTorchFunctionTLS& PythonTorchFunctionTLS::get_state() {
 }
 
 bool torch_function_mode_enabled() {
-  return PythonTorchFunctionTLS::get_disabled_state() != TorchFunctionDisabledState::ALL_DISABLED &&
-         PythonTorchFunctionTLS::stack_len() > 0;
+  // Manually flatten because gcc is refusing to inline here.  Note
+  // that we are still calling __tls_get_addr twice here with GCC,
+  // presumably because of
+  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81501 (which says
+  // the fix ships in GCC 16), but forcing inlining still improves
+  // performance.
+  const auto& ptfs = pythonTorchFunctionState;
+  return ptfs.disabled_state_ != TorchFunctionDisabledState::ALL_DISABLED && !ptfs.stack_.empty();
 }
 
 // This is needed to disambiguate the ternary torch function disabled states

aten/src/ATen/PythonTorchFunctionTLS.h

@@ -27,6 +27,7 @@ struct TORCH_API PythonTorchFunctionTLS {
   TorchFunctionDisabledState disabled_state_ =
       TorchFunctionDisabledState::ENABLED;
   std::vector<std::shared_ptr<c10::SafePyObject>> stack_;
+  friend TORCH_API bool torch_function_mode_enabled();
 };
 
 TORCH_API bool torch_function_mode_enabled();

aten/src/ATen/SavedTensorHooks.cpp

@@ -13,7 +13,7 @@ namespace {
   // and left at true for the rest of the execution.
   // It's an optimization so that users who never use default hooks don't need to
   // read the thread_local variables pack_hook_ and unpack_hook_.
-  static bool is_initialized(false);
+  bool is_initialized(false);
 }
 
 static void assertSavedTensorHooksNotDisabled() {

aten/src/ATen/TensorIterator.cpp

@@ -56,7 +56,7 @@ inline void get_strides(int64_t* strides, ArrayRef<OperandInfo> operands, int64_
   }
 }
 
-static OptionalTensorRef make_otr(const TensorBase &tensor) {
+OptionalTensorRef make_otr(const TensorBase &tensor) {
   if (tensor.defined()) {
     return OptionalTensorRef(tensor);
   } else {

aten/src/ATen/autocast_mode.cpp

@@ -36,7 +36,7 @@ namespace {
 using weakref_type = c10::weak_intrusive_ptr<TensorImpl, UndefinedTensorImpl>;
 using val_type = std::tuple<weakref_type, Tensor>;
 
-static ska::flat_hash_map<TensorImpl*, val_type>& get_cached_casts() {
+ska::flat_hash_map<TensorImpl*, val_type>& get_cached_casts() {
   static ska::flat_hash_map<TensorImpl*, val_type> cached_casts;
   return cached_casts;
 }

aten/src/ATen/core/CachingHostAllocator.cpp

@@ -6,9 +6,9 @@ namespace at {
 
 namespace {
 
-static std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
+std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
     allocator_array{};
-static std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
+std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
     allocator_priority{};
 
 } // anonymous namespace

aten/src/ATen/core/ivalue.h

@@ -624,7 +624,14 @@ struct TORCH_API IValue final {
   IValue(const c10::SymBool& i) {
     if (auto mi = i.maybe_as_bool()) {
       tag = Tag::Bool;
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
       payload.u.as_int = *mi;
+#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+      /* due to byteorder if value assigned as_int, as_bool actually is not set correctly */
+      payload.u.as_bool = *mi;
+#else
+#error Unexpected or undefined __BYTE_ORDER__
+#endif
     } else {
       tag = Tag::SymBool;
       payload.u.as_intrusive_ptr = i.toSymNodeImpl().release();

aten/src/ATen/cuda/CUDABlas.cpp

@@ -108,7 +108,7 @@ static hipblasStatus_t rocBLASStatusToHIPStatus(rocblas_status error)
 
 namespace {
 
-static cublasOperation_t _cublasOpFromChar(char op) {
+cublasOperation_t _cublasOpFromChar(char op) {
   // NOLINTNEXTLINE(bugprone-switch-missing-default-case)
   switch (op) {
     case 'n':
@@ -128,7 +128,7 @@ static cublasOperation_t _cublasOpFromChar(char op) {
       "_cublasOpFromChar input should be 't', 'n' or 'c' but got `", op, "`");
 }
 
-static void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
+void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
   // Note: leading dimensions generally are checked that they are > 0
   // and at least as big the result requires (even if the value won't
   // be used).
@@ -142,7 +142,7 @@ static void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
     *lda = std::max<int64_t>(m, 1);
 }
 
-static void _cublasAdjustLdLevel3(
+void _cublasAdjustLdLevel3(
     char transa,
     char transb,
     int64_t m,

aten/src/ATen/cuda/CUDAGeneratorImpl.cpp

@@ -15,19 +15,19 @@ namespace cuda::detail {
 namespace {
 
 // Total number of gpus in the system.
-static int64_t num_gpus;
+int64_t num_gpus;
 
 // Ensures default_gens_cuda is initialized once.
-static std::deque<c10::once_flag> cuda_gens_init_flag;
+std::deque<c10::once_flag> cuda_gens_init_flag;
 
 // Default, global CUDA generators, one per GPU.
-static std::vector<Generator> default_gens_cuda;
+std::vector<Generator> default_gens_cuda;
 
 /*
  * Populates the global variables related to CUDA generators
  * Warning: this function must only be called once!
  */
-static void initCUDAGenVector() {
+void initCUDAGenVector() {
   // Ensures we only call cudaGetDeviceCount only once.
   static bool num_gpu_init_flag [[maybe_unused]] = []() {
     num_gpus = static_cast<int32_t>(c10::cuda::device_count());

aten/src/ATen/functorch/BatchRulesLinearAlgebra.cpp

@@ -39,7 +39,7 @@ Tensor vdot_decomp(const Tensor& A, const Tensor& B) {
 // NB: I wrote this like this because we *might* want its for a future matmul
 // batch rule that isn't decomposed...
 // "tv" = tensor @ vector
-static std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
+std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
     const Tensor& self, std::optional<int64_t> self_bdim,
     const Tensor& other, std::optional<int64_t> other_bdim) {
   if (self_bdim && other_bdim) {
@@ -66,7 +66,7 @@ static std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
   TORCH_INTERNAL_ASSERT(false, "can't get here");
 }
 
-static std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
+std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
     const Tensor& self, std::optional<int64_t> self_bdim,
     const Tensor& other, std::optional<int64_t> other_bdim) {
   auto self_logical_rank = rankWithoutBatchDim(self, self_bdim);
@@ -79,7 +79,7 @@ static std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
   return tv_batch_rule(self, self_bdim, other, other_bdim);
 }
 
-static std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
+std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
     const Tensor& self, std::optional<int64_t> self_bdim,
     const Tensor& other, std::optional<int64_t> other_bdim) {
   auto self_logical_rank = rankWithoutBatchDim(self, self_bdim);
@@ -94,7 +94,7 @@ static std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
   return std::make_tuple( at::matmul(self_, other_), 0 );
 }
 
-static std::tuple<Tensor, std::optional<int64_t>> bmm_batch_rule(
+std::tuple<Tensor, std::optional<int64_t>> bmm_batch_rule(
     const Tensor& self, std::optional<int64_t> self_bdim,
     const Tensor& other, std::optional<int64_t> other_bdim) {
   auto self_logical_rank = rankWithoutBatchDim(self, self_bdim);
@@ -250,7 +250,7 @@ struct LinalgCheckMatrixBinaryRuleHelper<op_name, F, Func, typelist<A, B, T...>>
   }
 };
 
-static void expect_at_least_rank(
+void expect_at_least_rank(
     const Tensor& tensor,
     std::optional<int64_t> tensor_bdim,
     int64_t expected_rank,
@@ -472,7 +472,7 @@ atol_rtol_tensor_batch_rule(
   return std::make_tuple(Func(input_, atol_, rtol_, hermitian), 0);
 }
 
-static std::tuple<Tensor, std::optional<int64_t>>
+std::tuple<Tensor, std::optional<int64_t>>
 pinv_batch_rule(
     const Tensor& input, std::optional<int64_t> input_bdim, const std::optional<Tensor>& atol,
     const std::optional<int64_t> atol_bdim, const std::optional<Tensor>& rtol,

aten/src/ATen/functorch/BatchRulesScatterOps.cpp

@@ -19,7 +19,7 @@
 namespace at::functorch {
 
 namespace {
-static bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
+bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
   for (const auto& bdim : bdims) {
     if (bdim.has_value()) {
       return true;
@@ -28,7 +28,7 @@ static bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
   return false;
 }
 
-static int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
+int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
   int64_t result = 0;
   for (const auto& idx : indices) {
     if (!idx.has_value() || !idx->defined()) {
@@ -40,7 +40,7 @@ static int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
   return result;
 }
 
-static int64_t get_max_index_logical_dim(
+int64_t get_max_index_logical_dim(
     ArrayRef<std::optional<Tensor>> indices,
     ArrayRef<std::optional<int64_t>> indices_bdims) {
   int64_t max_logical_dim = -1;
@@ -57,7 +57,7 @@ static int64_t get_max_index_logical_dim(
   return max_logical_dim;
 }
 
-static std::vector<std::optional<Tensor>> batchIndices(
+std::vector<std::optional<Tensor>> batchIndices(
   at::TensorOptions options,
   ArrayRef<std::optional<Tensor>> indices,
   ArrayRef<std::optional<int64_t>> indices_bdims,
@@ -126,7 +126,7 @@ static std::vector<std::optional<Tensor>> batchIndices(
 
 // Define an "advanced index" to be a selection object that is
 // a non-trivial Tensor (i.e. it does not represent :).
-static bool is_advanced_index(const std::optional<Tensor>& idx) {
+bool is_advanced_index(const std::optional<Tensor>& idx) {
   if (!idx.has_value()) {
     return false;
   }
@@ -137,7 +137,7 @@ static bool is_advanced_index(const std::optional<Tensor>& idx) {
 }
 
 // See NOTE: [advanced indices adjacent] for definition
-static bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indices) {
+bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indices) {
   int64_t num_advanced_indices_regions = 0;
   bool in_advanced_indices_region = false;
   for (const auto& idx : indices) {
@@ -165,7 +165,7 @@ static bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indice
 // - result: Tensor[B, 4, 5, 6, 2, 3, 7, 8]
 //                     -------  ----
 //                     region2  region1
-static Tensor swap_regions(const Tensor& tensor, int64_t first_region_size, int64_t second_region_size) {
+Tensor swap_regions(const Tensor& tensor, int64_t first_region_size, int64_t second_region_size) {
   VmapDimVector permutation(tensor.dim(), 0);
   std::iota(permutation.begin(), permutation.end(), 0);
   std::rotate(
@@ -553,7 +553,7 @@ Tensor &_index_put_impl__plumbing(Tensor &self, const List<std::optional<Tensor>
   return self;
 }
 
-static Tensor maybe_permute_values(
+Tensor maybe_permute_values(
     const Tensor& values,
     ArrayRef<std::optional<Tensor>> orig_indices,
     ArrayRef<std::optional<int64_t>> orig_indices_bdims) {
@@ -1052,7 +1052,7 @@ std::tuple<Tensor, std::optional<int64_t>> index_add_batch_rule(
                                    other, other_bdim, alpha, false);
 }
 
-static std::tuple<Tensor,Tensor> binary_pointwise_align(
+std::tuple<Tensor,Tensor> binary_pointwise_align(
     const Tensor & self,
     std::optional<int64_t> self_bdim,
     const Tensor & mask,

aten/src/ATen/functorch/BatchRulesViews.cpp

@@ -346,7 +346,7 @@ std::tuple<Tensor, std::optional<int64_t>> slice_batch_rule(
   return std::make_tuple(std::move(result), 0);
 }
 
-static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
+bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
   return dim == 0 || dim == -1;
 }
 

aten/src/ATen/functorch/LegacyBatchingRegistrations.cpp

@@ -68,18 +68,18 @@ namespace at::functorch {
 
 namespace{
 // PyTorch allows operations to specify dim 0 and dim -1 on a scalar tensor.
-static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
+bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
   return dim == 0 || dim == -1;
 }
 
-static int64_t get_current_level() {
+int64_t get_current_level() {
   auto maybe_level = maybeCurrentDynamicLayer();
   TORCH_INTERNAL_ASSERT(maybe_level.has_value());
   return maybe_level->layerId();
 }
 
 // This check should probably go into the dispatcher...
-static bool participatesInCurrentLevel(const Tensor& self) {
+bool participatesInCurrentLevel(const Tensor& self) {
   auto current_level = get_current_level();
   auto* maybe_batched_impl = maybeGetBatchedImpl(self);
   if (!maybe_batched_impl) {
@@ -90,7 +90,7 @@ static bool participatesInCurrentLevel(const Tensor& self) {
   return self_level == current_level;
 }
 
-static bool participatesInCurrentLevel(ITensorListRef self) {
+bool participatesInCurrentLevel(ITensorListRef self) {
   for (const Tensor& tensor : self) {
     if (participatesInCurrentLevel(tensor)) {
       return true;
@@ -285,7 +285,7 @@ std::vector<Tensor> unbind_batching_rule(const Tensor& self, int64_t dim) {
 // given (sizes, strides, storage_offset) returns the maximum location that
 // can be indexed (or nullopt if such a location doesn't exist, e.g., tensors
 // with zero-size dims).
-static std::optional<c10::SymInt> maximum_indexable_location(
+std::optional<c10::SymInt> maximum_indexable_location(
     c10::SymIntArrayRef sizes, c10::SymIntArrayRef strides, const c10::SymInt& storage_offset) {
   auto result = native::storage_size_for(sizes, strides);
   if (result == 0) {
@@ -298,7 +298,7 @@ static std::optional<c10::SymInt> maximum_indexable_location(
 // This checks that the range of possible memory locations accessible by
 // x.as_strided(sizes, strides, maybe_storage_offset)
 // are within the bounds of possible memory locations accessible by x.
-static void checkBasicAsStridedValidForSlice(
+void checkBasicAsStridedValidForSlice(
     const Tensor& physical_tensor,
     int64_t num_batch_dims,
     c10::SymIntArrayRef sizes,

aten/src/ATen/functorch/PyTorchOperatorHacks.cpp

@@ -71,7 +71,7 @@ Tensor linear_hack(const Tensor& input, const Tensor& weight, const std::optiona
   return output;
 }
 
-static inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
+inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
   if (reduction == at::Reduction::Mean) {
     return unreduced.mean();
   } else if (reduction == at::Reduction::Sum) {
@@ -127,7 +127,7 @@ namespace {
 template<bool inplace>
 using Ctype = std::conditional_t<inplace, Tensor&, Tensor>;
 
-static Tensor make_feature_noise(const Tensor& input) {
+Tensor make_feature_noise(const Tensor& input) {
   auto input_sizes = input.sizes();
   TORCH_CHECK(input.dim() >= 2, "Feature dropout requires at least 2 dimensions in the input");
   std::vector<int64_t> sizes;
@@ -141,7 +141,7 @@ static Tensor make_feature_noise(const Tensor& input) {
   return at::empty(sizes, input.options());
 }
 
-static bool is_fused_kernel_acceptable(const Tensor& input, double p) {
+bool is_fused_kernel_acceptable(const Tensor& input, double p) {
   return (input.is_cuda() || input.is_xpu() || input.is_lazy() || input.is_privateuseone()) && p > 0 && p < 1 && input.numel() > 0;
 }
 
@@ -210,7 +210,7 @@ ALIAS_SPECIALIZATION(_feature_dropout,       true,  false)
 ALIAS_SPECIALIZATION(_alpha_dropout,         false, true )
 ALIAS_SPECIALIZATION(_feature_alpha_dropout, true,  true )
 
-static Tensor dropout(const Tensor& input, double p, bool train) {
+Tensor dropout(const Tensor& input, double p, bool train) {
   auto result = [&]() {
     NoNamesGuard guard;
     if (train && is_fused_kernel_acceptable(input, p)) {

aten/src/ATen/native/AdaptiveAveragePooling3d.cpp

@@ -24,7 +24,7 @@ namespace at::native {
 namespace {
 
 template <typename scalar_t>
-static void adaptive_avg_pool3d_out_frame(
+void adaptive_avg_pool3d_out_frame(
     const scalar_t* input_p,
     scalar_t* output_p,
     int64_t sizeD,
@@ -176,7 +176,7 @@ void adaptive_avg_pool3d_out_cpu_template(
 }
 
 template <typename scalar_t>
-static void adaptive_avg_pool3d_backward_out_frame(
+void adaptive_avg_pool3d_backward_out_frame(
     scalar_t* gradInput_p,
     const scalar_t* gradOutput_p,
     int64_t sizeD,

aten/src/ATen/native/AdaptiveMaxPooling3d.cpp

@@ -93,7 +93,7 @@ namespace {
 // 5d tensor B x D x T x H x W
 
 template <typename scalar_t>
-static void adaptive_max_pool3d_single_out_frame(
+void adaptive_max_pool3d_single_out_frame(
           const scalar_t *input_p,
           scalar_t *output_p,
           int64_t *ind_p,
@@ -170,7 +170,7 @@ static void adaptive_max_pool3d_single_out_frame(
 }
 
 template <typename scalar_t>
-static void adaptive_max_pool3d_out_frame(
+void adaptive_max_pool3d_out_frame(
           const scalar_t *input_data,
           scalar_t *output_data,
           int64_t *indices_data,
@@ -202,7 +202,7 @@ static void adaptive_max_pool3d_out_frame(
 }
 
 template <typename scalar_t>
-static void adaptive_max_pool3d_backward_single_out_frame(
+void adaptive_max_pool3d_backward_single_out_frame(
           scalar_t *gradInput_p,
           const scalar_t *gradOutput_p,
           const int64_t *ind_p,
@@ -241,7 +241,7 @@ static void adaptive_max_pool3d_backward_single_out_frame(
 }
 
 template <typename scalar_t>
-static void adaptive_max_pool3d_backward_out_frame(
+void adaptive_max_pool3d_backward_out_frame(
           scalar_t *gradInput_data,
           const scalar_t *gradOutput_data,
           const int64_t *indices_data,

aten/src/ATen/native/AveragePool3d.cpp

@@ -153,7 +153,7 @@ namespace at::native {
 namespace {
 
 template <typename scalar_t>
-static void avg_pool3d_out_frame(
+void avg_pool3d_out_frame(
           const scalar_t *input_p,
           scalar_t *output_p,
           int64_t nslices,
@@ -333,7 +333,7 @@ TORCH_IMPL_FUNC(avg_pool3d_out_cpu) (
 namespace {
 
 template <typename scalar_t>
-static void avg_pool3d_backward_out_frame(
+void avg_pool3d_backward_out_frame(
           scalar_t *gradInput_p,
           const scalar_t *gradOutput_p,
           int64_t nslices,

aten/src/ATen/native/BatchLinearAlgebraKernel.cpp

@@ -143,13 +143,13 @@ Tensor& cholesky_inverse_kernel_impl(Tensor& result, Tensor& infos, bool upper)
  For more info see https://github.com/pytorch/pytorch/issues/145801#issuecomment-2631781776
 */
 template <typename T>
-static inline
+inline
 std::enable_if_t<std::is_floating_point_v<T>, int> lapack_work_to_int(const T val) {
     const auto next_after = std::nextafter(val, std::numeric_limits<T>::infinity());
     return std::max<int>(1, std::ceil(next_after));
 }
 template <typename T>
-static inline
+inline
 std::enable_if_t<c10::is_complex<T>::value, int> lapack_work_to_int(const T val) {
     return lapack_work_to_int(val.real());
 }
@@ -343,7 +343,7 @@ void linalg_eigh_kernel(const Tensor& eigenvalues, const Tensor& eigenvectors, c
   For further details, please see the LAPACK documentation for GEQRF.
 */
 template <typename scalar_t>
-static void apply_geqrf(const Tensor& input, const Tensor& tau) {
+void apply_geqrf(const Tensor& input, const Tensor& tau) {
 #if !AT_BUILD_WITH_LAPACK()
   TORCH_CHECK(
       false,
@@ -1039,7 +1039,7 @@ void lu_solve_kernel(const Tensor& LU, const Tensor& pivots, const Tensor& B, Tr
 }
 
 template <typename scalar_t>
-static void apply_svd(const Tensor& A,
+void apply_svd(const Tensor& A,
                       const bool full_matrices,
                       const bool compute_uv,
                       const Tensor& U,

aten/src/ATen/native/Col2Im.cpp

@@ -71,7 +71,7 @@
 namespace at::native {
 namespace {
 
-static void col2im_out_cpu_template(
+void col2im_out_cpu_template(
     Tensor& output,
     const Tensor& input_,
     IntArrayRef output_size,

aten/src/ATen/native/ConvolutionMM2d.cpp

@@ -25,7 +25,7 @@ namespace at::native {
 
 namespace {
 
-static Tensor compute_columns2d(
+Tensor compute_columns2d(
     const Tensor& input,
     IntArrayRef padding,
     IntArrayRef stride,
@@ -93,7 +93,7 @@ static Tensor compute_columns2d(
   return columns.contiguous();
 }
 
-static inline void slow_conv2d_shape_check(
+inline void slow_conv2d_shape_check(
     const Tensor& input,
     const Tensor& grad_output,
     const Tensor& weight,
@@ -205,7 +205,7 @@ static inline void slow_conv2d_shape_check(
   }
 }
 
-static inline Tensor view_weight_2d(const Tensor& weight_,
+inline Tensor view_weight_2d(const Tensor& weight_,
     at::MemoryFormat memory_format = at::MemoryFormat::Contiguous) {
   Tensor weight = weight_.contiguous(memory_format);
   if (weight.dim() == 4) {
@@ -220,7 +220,7 @@ static inline Tensor view_weight_2d(const Tensor& weight_,
 }
 
 template <typename scalar_t>
-static void slow_conv2d_update_output_frame(
+void slow_conv2d_update_output_frame(
     TensorAccessor<const scalar_t, 3> input,
     TensorAccessor<scalar_t, 3> output,
     TensorAccessor<const scalar_t, 2> weight,
@@ -480,7 +480,7 @@ void slow_conv2d_backward_weight_frame(
   }
 }
 
-static void slow_conv2d_backward_weight_out_cpu_template(
+void slow_conv2d_backward_weight_out_cpu_template(
     Tensor& grad_weight,
     const Tensor& input,
     const Tensor& grad_output_,

aten/src/ATen/native/ConvolutionMM3d.cpp

@@ -28,7 +28,7 @@ namespace at::native {
 
 namespace {
 
-static Tensor compute_columns3d(
+Tensor compute_columns3d(
     const Tensor& input_,
     IntArrayRef stride,
     IntArrayRef padding,
@@ -108,7 +108,7 @@ static Tensor compute_columns3d(
   return columns;
 }
 
-static inline void slow_conv3d_shape_check(
+inline void slow_conv3d_shape_check(
     const Tensor& input,
     const Tensor& grad_output,
     const Tensor& weight,
@@ -273,7 +273,7 @@ static inline void slow_conv3d_shape_check(
   }
 }
 
-static Tensor view_weight_2d(const Tensor& weight_) {
+Tensor view_weight_2d(const Tensor& weight_) {
   Tensor weight = weight_.contiguous();
   if (weight.dim() == 5) {
     const int64_t s1 = weight.size(0);
@@ -286,7 +286,7 @@ static Tensor view_weight_2d(const Tensor& weight_) {
 }
 
 template <typename scalar_t>
-static void slow_conv3d_update_output_frame(
+void slow_conv3d_update_output_frame(
     TensorAccessor<const scalar_t, 4> input,
     TensorAccessor<scalar_t, 4> output,
     TensorAccessor<const scalar_t, 2> weight,
@@ -515,7 +515,7 @@ void slow_conv3d_backward_weight_frame(
       grad_weight.data(), ldc, grad_weight.stride(0) * n);
 }
 
-static void slow_conv3d_backward_parameters_out_cpu_template(
+void slow_conv3d_backward_parameters_out_cpu_template(
     Tensor& grad_weight,
     const Tensor& input,
     const Tensor& grad_output,

aten/src/ATen/native/EmbeddingBag.cpp

@@ -108,7 +108,7 @@ bool is_fast_path(const Tensor& src, const std::optional<Tensor>& scale, Tensor&
 // index_add (using add_indices as the index), without creating an intermediary
 // tensor to hold the selected embeddings
 template <typename data_t, typename index_t>
-static std::enable_if_t<std::is_same_v<data_t, double>, void>
+std::enable_if_t<std::is_same_v<data_t, double>, void>
 index_select_add(
     const Tensor& select_indices,
     const Tensor& add_indices,
@@ -494,7 +494,7 @@ index_select_add(const Tensor &select_indices,
 // mul (scaling by per_sample_weights)
 // index_add (using add_indices as the index)
 template <typename data_t, typename index_t>
-static std::enable_if_t<std::is_same_v<data_t, double>, void>
+std::enable_if_t<std::is_same_v<data_t, double>, void>
 index_select_scale_add(
     const Tensor& select_indices,
     const Tensor& add_indices,

aten/src/ATen/native/FractionalMaxPool2d.cpp

@@ -130,7 +130,7 @@ namespace native {
 namespace {
 
 template <typename scalar_t>
-static void fractional_max_pool2d_out_single_batch_frame(
+void fractional_max_pool2d_out_single_batch_frame(
   const scalar_t* input,
   scalar_t* output,
   int64_t* indices,
@@ -188,7 +188,7 @@ static void fractional_max_pool2d_out_single_batch_frame(
 }
 
 template <typename scalar_t>
-static void fractional_max_pool2d_out_frame(
+void fractional_max_pool2d_out_frame(
   const scalar_t* input,
   scalar_t* output,
   int64_t* indices,
@@ -220,7 +220,7 @@ static void fractional_max_pool2d_out_frame(
   }
 
 template <typename scalar_t>
-static void fractional_max_pool2d_backward_out_single_batch_frame(
+void fractional_max_pool2d_backward_out_single_batch_frame(
   scalar_t* gradInput,
   const scalar_t* gradOutput,
   const int64_t* indices,
@@ -247,7 +247,7 @@ static void fractional_max_pool2d_backward_out_single_batch_frame(
 }
 
 template <typename scalar_t>
-static void fractional_max_pool2d_backward_out_frame(
+void fractional_max_pool2d_backward_out_frame(
   scalar_t* gradInput,
   const scalar_t* gradOutput,
   const int64_t* indices,

aten/src/ATen/native/FractionalMaxPool3d.cpp

@@ -99,7 +99,7 @@ namespace at::native {
 namespace {
 
 template<typename scalar_t>
-static void fractional_max_pool3d_out_single_batch_frame(
+void fractional_max_pool3d_out_single_batch_frame(
   const scalar_t* input,
   scalar_t* output,
   int64_t* indices,
@@ -169,7 +169,7 @@ static void fractional_max_pool3d_out_single_batch_frame(
 }
 
 template<typename scalar_t>
-static void fractional_max_pool3d_out_frame(
+void fractional_max_pool3d_out_frame(
   const scalar_t* input,
   scalar_t* output,
   int64_t* indices,
@@ -257,7 +257,7 @@ TORCH_IMPL_FUNC(fractional_max_pool3d_out_cpu)(
 namespace {
 
 template<typename scalar_t>
-static void fractional_max_pool3d_backward_out_single_batch_frame(
+void fractional_max_pool3d_backward_out_single_batch_frame(
   scalar_t* gradInput,
   const scalar_t* gradOutput,
   const int64_t* indices,
@@ -287,7 +287,7 @@ static void fractional_max_pool3d_backward_out_single_batch_frame(
 }
 
 template<typename scalar_t>
-static void fractional_max_pool3d_backward_out_frame(
+void fractional_max_pool3d_backward_out_frame(
   scalar_t* gradInput,
   const scalar_t* gradOutput,
   const int64_t* indices,

aten/src/ATen/native/Im2Col.cpp

@@ -19,7 +19,7 @@
 namespace at::native {
 namespace {
 
-static void im2col_out_cpu_template(
+void im2col_out_cpu_template(
     Tensor& output,
     const Tensor& input_,
     IntArrayRef kernel_size,

aten/src/ATen/native/Loss.cpp

@@ -61,7 +61,7 @@
 constexpr float EPSILON = 1e-12;
 
 namespace {
-  static inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
+  inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
     if (reduction == at::Reduction::Mean) {
       return unreduced.mean();
     } else if (reduction == at::Reduction::Sum) {

aten/src/ATen/native/LossCTC.cpp

@@ -44,7 +44,7 @@ namespace {
 
 // this ad-hoc converts from targets (l in [1]) to augmented targets (l' in [1]) note that no bound-checking is done
 template<typename target_t>
-static inline int64_t get_target_prime(target_t* target, int64_t offset, int64_t stride, int64_t idx, int64_t BLANK) {
+inline int64_t get_target_prime(target_t* target, int64_t offset, int64_t stride, int64_t idx, int64_t BLANK) {
   if (idx % 2 == 0) {
     return BLANK;
   } else {

aten/src/ATen/native/LossMultiLabelMargin.cpp

@@ -58,7 +58,7 @@ inline scalar_t multilabel_margin_loss_forward_inner_sum_cpu(
 }
 
 template <typename scalar_t>
-static void multilabel_margin_loss_forward_out_frame(
+void multilabel_margin_loss_forward_out_frame(
     const Tensor& input_contiguous,
     const Tensor& target_contiguous,
     Tensor& output,
@@ -108,7 +108,7 @@ static void multilabel_margin_loss_forward_out_frame(
   }
 }
 
-static void multilabel_margin_loss_forward_out_cpu_template(
+void multilabel_margin_loss_forward_out_cpu_template(
     const Tensor& input,
     const Tensor& target,
     Tensor& output,
@@ -153,7 +153,7 @@ static void multilabel_margin_loss_forward_out_cpu_template(
 }
 
 template <typename scalar_t>
-static void multilabel_margin_loss_backward_out_frame(
+void multilabel_margin_loss_backward_out_frame(
     Tensor& grad_input,
     const Tensor& grad_output,
     const Tensor& input_contiguous,
@@ -222,7 +222,7 @@ static void multilabel_margin_loss_backward_out_frame(
   }
 }
 
-static void multilabel_margin_loss_backward_out_cpu_template(
+void multilabel_margin_loss_backward_out_cpu_template(
     Tensor& grad_input,
     const Tensor& grad_output,
     const Tensor& input,

aten/src/ATen/native/LossMultiMargin.cpp

@@ -57,7 +57,7 @@ inline int64_t target_index_checked(
 }
 
 template <typename scalar_t>
-static inline void multi_margin_loss_cpu_kernel(
+inline void multi_margin_loss_cpu_kernel(
     Tensor& output,
     const scalar_t* input_data,
     const int64_t* target_data,
@@ -148,7 +148,7 @@ void multi_margin_loss_out_cpu_template(
 }
 
 template <typename scalar_t>
-static void multi_margin_loss_backward_cpu_kernel(
+void multi_margin_loss_backward_cpu_kernel(
     scalar_t* grad_input_data,
     const Tensor& grad_output,
     const scalar_t* input_data,

aten/src/ATen/native/LossNLL.cpp

@@ -159,7 +159,7 @@ inline scalar_t* optional_data(const Tensor& source) {
 }
 
 template <typename scalar_t, typename target_t>
-static void nll_loss_out_frame(
+void nll_loss_out_frame(
     const Tensor& output,
     const Tensor& total_weight,
     const Tensor& input,
@@ -338,7 +338,7 @@ void nll_loss_forward_out_cpu_template(
 }
 
 template <typename scalar_t, typename target_t>
-static void nll_loss_backward_out_frame(
+void nll_loss_backward_out_frame(
     const Tensor& grad_input,
     const Tensor& grad_output,
     const Tensor& input,

aten/src/ATen/native/LossNLL2d.cpp

@@ -99,7 +99,7 @@ inline void check_gradout_shape_nll_loss2d(
 
 
 template <typename scalar_t>
-static void nll_loss2d_forward_out_frame(
+void nll_loss2d_forward_out_frame(
     Tensor& output,
     Tensor& total_weight,
     const Tensor& input,
@@ -280,7 +280,7 @@ void nll_loss2d_forward_out_cpu_template(
 }
 
 template <typename scalar_t>
-static void nll_loss2d_backward_out_frame(
+void nll_loss2d_backward_out_frame(
     Tensor& grad_input,
     const Tensor& grad_output,
     const Tensor& input,

aten/src/ATen/native/NaiveConvolutionTranspose2d.cpp

@@ -24,7 +24,7 @@
 
 namespace at {
 namespace {
-static inline void slow_conv_transpose2d_shape_check(
+inline void slow_conv_transpose2d_shape_check(
     const Tensor& input,
     const Tensor& grad_output,
     const Tensor& weight,
@@ -386,7 +386,7 @@ void slow_conv_transpose2d_out_cpu_template(
   }
 }
 
-static void slow_conv_transpose2d_backward_out_cpu_template(
+void slow_conv_transpose2d_backward_out_cpu_template(
     const Tensor& input_,
     const Tensor& grad_output_,
     Tensor& grad_input,

aten/src/ATen/native/NaiveConvolutionTranspose3d.cpp

@@ -22,7 +22,7 @@ namespace at::native {
 
 namespace {
 
-static inline void slow_conv_transpose3d_shape_check(
+inline void slow_conv_transpose3d_shape_check(
     const Tensor& input,
     const Tensor& grad_output,
     const Tensor& weight,

aten/src/ATen/native/Normalization.cpp

@@ -92,7 +92,7 @@ namespace {
              arg_name, " should contain ", expected, " elements not ", actual);
   }
 
-  static inline Tensor repeat_if_defined(const Tensor& t, const SymInt& repeat) {
+  inline Tensor repeat_if_defined(const Tensor& t, const SymInt& repeat) {
     if (t.defined()) {
       return t.repeat_symint(repeat);
     }

aten/src/ATen/native/RNN.cpp

@@ -538,7 +538,7 @@ c10::intrusive_ptr<CellParamsBase> make_quantized_cell_params_fp16(
       std::move(w_ih_packed), std::move(w_hh_packed));
 }
 
-static std::unordered_map<
+std::unordered_map<
     std::string,
     c10::intrusive_ptr<CellParamsBase> (*)(CellParamsSerializationType)>
     cell_params_deserializers = {
@@ -578,7 +578,7 @@ struct QRNNCellParamsWrapper {
 
 // Gathers every two elements of a vector in a vector of pairs
 template<typename T>
-static std::vector<pair_of<T>> pair_vec(const std::vector<T>& vals) {
+std::vector<pair_of<T>> pair_vec(const std::vector<T>& vals) {
   TORCH_CHECK(vals.size() % 2 == 0, "Odd number of params or hiddens given to a bidirectional RNN");
   std::vector<pair_of<T>> result;
   result.reserve(vals.size() / 2);
@@ -590,7 +590,7 @@ static std::vector<pair_of<T>> pair_vec(const std::vector<T>& vals) {
 
 // Flattens a vector of pairs
 template<typename T>
-static std::vector<T> unpair_vec(std::vector<pair_of<T>>&& vals) {
+std::vector<T> unpair_vec(std::vector<pair_of<T>>&& vals) {
   std::vector<T> result;
   result.reserve(vals.size() * 2);
   for (const auto i : c10::irange(vals.size())) {
@@ -601,7 +601,7 @@ static std::vector<T> unpair_vec(std::vector<pair_of<T>>&& vals) {
 }
 
 // Parses a flat list of parameter tensors into a list of CellParams
-static std::vector<CellParams> gather_params(TensorList params, bool has_biases, bool has_projections = false) {
+std::vector<CellParams> gather_params(TensorList params, bool has_biases, bool has_projections = false) {
   static at::Tensor undefined;
   std::vector<CellParams> result;
   if (has_biases) {
@@ -1894,10 +1894,10 @@ static DEFINE_QUANTIZED_RNN_CELL_DYNAMIC(quantized_rnn_tanh_cell_dynamic, simple
 
 namespace {
 
-[[maybe_unused]] static auto ensure_linear_params_registered =
+[[maybe_unused]] auto ensure_linear_params_registered =
     register_linear_params();
 
-static auto cell_params_base_registry =
+auto cell_params_base_registry =
     torch::selective_class_<CellParamsBase>("rnn", TORCH_SELECTIVE_CLASS("CellParamsBase"))
         .def_pickle(
             [](const c10::intrusive_ptr<CellParamsBase>& self)

aten/src/ATen/native/TensorAdvancedIndexing.cpp

@@ -2676,7 +2676,7 @@ inline std::tuple<Tensor, Tensor, int64_t> _take_along_dim_helper(
       std::move(dim));
 }
 
-static inline void checkDevice(CheckedFrom c, const Tensor& t, Device device) {
+inline void checkDevice(CheckedFrom c, const Tensor& t, Device device) {
   TORCH_CHECK(
       !t.defined() || t.device() == device,
       "Expected tensor to have ",
@@ -2689,7 +2689,7 @@ static inline void checkDevice(CheckedFrom c, const Tensor& t, Device device) {
       ")");
 }
 
-static inline void checkDevice(
+inline void checkDevice(
     CheckedFrom c,
     at::ArrayRef<Tensor> tensors,
     Device device) {

aten/src/ATen/native/TensorShape.cpp

@@ -3641,7 +3641,7 @@ Tensor& transpose_(Tensor& self, int64_t dim0, int64_t dim1) {
 namespace {
 // Transpose implementation for sparse compressed layouts
 // NB: We assume that dim1,dim0 have already been wrapped
-static inline Tensor sparse_compressed_transpose(
+inline Tensor sparse_compressed_transpose(
     const Tensor& self,
     int64_t dim0,
     int64_t dim1) {

aten/src/ATen/native/UnfoldBackward.h

@@ -29,7 +29,7 @@ namespace {
 // grad_in does not mean that it is a gradient wrt to input,
 // grad_in/grad_out is just an input/output of unfold_backward kernel.
 
-[[maybe_unused]] static TensorIterator _make_unfold_backward_iter_over_grad_out(
+[[maybe_unused]] TensorIterator _make_unfold_backward_iter_over_grad_out(
     Tensor& grad_out,
     const Tensor& grad_in,
     int64_t dim,

aten/src/ATen/native/UpSampleBicubic2d.cpp

@@ -105,7 +105,7 @@ namespace at::native {
 namespace {
 
 template <typename scalar_t>
-static void upsample_bicubic2d_backward_out_frame(
+void upsample_bicubic2d_backward_out_frame(
     const scalar_t* odata,
     scalar_t* idata,
     int64_t input_height,
@@ -177,7 +177,7 @@ static void upsample_bicubic2d_backward_out_frame(
   });
 }
 
-static void upsample_bicubic2d_backward_kernel(
+void upsample_bicubic2d_backward_kernel(
     const Tensor& grad_input,
     const Tensor& grad_output_,
     IntArrayRef output_size,

aten/src/ATen/native/ao_sparse/quantized/cpu/fbgemm_utils.cpp

@@ -39,6 +39,6 @@ int register_linear_params() {
 }
 
 namespace {
-[[maybe_unused]] static auto linear_params = register_linear_params();
+[[maybe_unused]] auto linear_params = register_linear_params();
 } // namespace
 }  // namespace ao::sparse

aten/src/ATen/native/cpu/Activation.cpp

@@ -30,7 +30,7 @@ namespace {
 // Workaround for gcc-14.2.0 ICE during RTL pass: expand when compiling for NEON
 __attribute__((optimize("no-tree-vectorize")))
 #endif
-static void log_sigmoid_cpu_kernel(TensorBase &output, TensorBase &buffer, const TensorBase &input) {
+void log_sigmoid_cpu_kernel(TensorBase &output, TensorBase &buffer, const TensorBase &input) {
   if (at::isReducedFloatingType(input.scalar_type())) {
     AT_DISPATCH_REDUCED_FLOATING_TYPES(input.scalar_type(), "log_sigmoid_cpu", [&]() {
     using Vec = Vectorized<scalar_t>;
@@ -96,7 +96,7 @@ static void log_sigmoid_cpu_kernel(TensorBase &output, TensorBase &buffer, const
   }
 }
 
-static void log_sigmoid_backward_cpu_kernel(TensorIterator& iter) {
+void log_sigmoid_backward_cpu_kernel(TensorIterator& iter) {
   if (at::isReducedFloatingType(iter.dtype())) {
     AT_DISPATCH_REDUCED_FLOATING_TYPES(iter.dtype(), "log_sigmoid_backward_cpu", [&]() {
       using Vec = Vectorized<scalar_t>;
@@ -150,7 +150,7 @@ static void log_sigmoid_backward_cpu_kernel(TensorIterator& iter) {
   }
 }
 
-static void threshold_kernel(
+void threshold_kernel(
     TensorIteratorBase& iter,
     const Scalar& threshold_scalar,
     const Scalar& value_scalar) {
@@ -868,7 +868,7 @@ void hardswish_backward_kernel(TensorIterator& iter) {
   }
 }
 
-static void leaky_relu_kernel(TensorIteratorBase& iter, const Scalar& negval_) {
+void leaky_relu_kernel(TensorIteratorBase& iter, const Scalar& negval_) {
   if (at::isReducedFloatingType(iter.dtype())) {
     AT_DISPATCH_REDUCED_FLOATING_TYPES(iter.dtype(), "leaky_relu_cpu", [&]() {
     auto zero_vec = Vectorized<float>((float)(0));
@@ -907,7 +907,7 @@ static void leaky_relu_kernel(TensorIteratorBase& iter, const Scalar& negval_) {
   }
 }
 
-static void leaky_relu_backward_kernel(TensorIteratorBase& iter, const Scalar& negval_) {
+void leaky_relu_backward_kernel(TensorIteratorBase& iter, const Scalar& negval_) {
   if (at::isReducedFloatingType(iter.dtype())) {
     AT_DISPATCH_REDUCED_FLOATING_TYPES(iter.dtype(), "leaky_relu_backward_cpu", [&]() {
     auto zero_vec = Vectorized<float>((float)(0));

aten/src/ATen/native/cpu/BlasKernel.cpp

@@ -369,7 +369,7 @@ void gemm_notrans_(
 #endif // defined(__aarch64__) && !defined(C10_MOBILE)
 
 #if !defined(C10_MOBILE)
-static float compute_dot(const at::Half* a, const at::Half* b, int64_t len) {
+float compute_dot(const at::Half* a, const at::Half* b, int64_t len) {
   return at::native::CPU_CAPABILITY::fp16_dot_with_fp32_arith(
       a, b, len);
 }
@@ -406,7 +406,7 @@ void gemm_transa_(
   });
 }
 
-static float compute_dot(const at::BFloat16* a, const at::BFloat16* b, int64_t len) {
+float compute_dot(const at::BFloat16* a, const at::BFloat16* b, int64_t len) {
   return at::native::CPU_CAPABILITY::bf16_dot_with_fp32_arith(a, b, len);
 }
 

aten/src/ATen/native/cpu/CopyKernel.cpp

@@ -15,12 +15,12 @@ namespace at::native {
 inline namespace CPU_CAPABILITY {
 
 namespace {
-static bool reduced_input(ScalarType input_t, ScalarType output_t) {
+bool reduced_input(ScalarType input_t, ScalarType output_t) {
   return !at::isFloat8Type(input_t) && at::isReducedFloatingType(input_t) &&
       output_t == kFloat;
 }
 
-static bool reduced_output(ScalarType input_t, ScalarType output_t) {
+bool reduced_output(ScalarType input_t, ScalarType output_t) {
   return !at::isFloat8Type(output_t) && at::isReducedFloatingType(output_t) &&
       input_t == kFloat;
 }

aten/src/ATen/native/cpu/CrossKernel.cpp

@@ -15,7 +15,7 @@ namespace at::native {
 namespace {
 
 template<typename scalar_t>
-static void apply_cross(const Tensor& result, const Tensor& a, const Tensor& b, const int64_t dim) {
+void apply_cross(const Tensor& result, const Tensor& a, const Tensor& b, const int64_t dim) {
   int64_t total = a.numel() / 3;
   int64_t a_stride = a.stride(dim);
   int64_t b_stride = b.stride(dim);
@@ -68,7 +68,7 @@ static void apply_cross(const Tensor& result, const Tensor& a, const Tensor& b,
   });
 }
 
-static void cross_kernel_impl(const Tensor& result, const Tensor& a, const Tensor& b, const int64_t dim) {
+void cross_kernel_impl(const Tensor& result, const Tensor& a, const Tensor& b, const int64_t dim) {
   AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kBFloat16, kHalf, result.scalar_type(), "cross", [&]() {
     apply_cross<scalar_t>(result, a, b, dim);
   });

aten/src/ATen/native/cpu/DistanceOpsKernel.cpp

@@ -422,19 +422,19 @@ void pdist_forward_kernel_impl(Tensor& result, const Tensor& self, const double
   });
 }
 
-static void pdist_backward_kernel_impl(Tensor& result, const Tensor& grad, const Tensor& self, const double p, const Tensor& dist) {
+void pdist_backward_kernel_impl(Tensor& result, const Tensor& grad, const Tensor& self, const double p, const Tensor& dist) {
   AT_DISPATCH_FLOATING_TYPES(self.scalar_type(), "pdist_backward", [&] {
     Dist<scalar_t>::apply_backward_pdist(result, grad, self, p, dist);
   });
 }
 
-static void cdist_kernel_impl(Tensor& result, const Tensor& x1, const Tensor& x2, const double p) {
+void cdist_kernel_impl(Tensor& result, const Tensor& x1, const Tensor& x2, const double p) {
   AT_DISPATCH_FLOATING_TYPES(result.scalar_type(), "cdist", [&] {
     Dist<scalar_t>::apply_cdist(result, x1, x2, p);
   });
 }
 
-static void cdist_backward_kernel_impl(Tensor& result, const Tensor& grad, const Tensor& x1, const Tensor& x2, const double p, const Tensor& dist) {
+void cdist_backward_kernel_impl(Tensor& result, const Tensor& grad, const Tensor& x1, const Tensor& x2, const double p, const Tensor& dist) {
   AT_DISPATCH_FLOATING_TYPES(result.scalar_type(), "cdist_backward", [&] {
     Dist<scalar_t>::apply_backward_cdist(result, grad, x1, x2, p, dist);
   });

aten/src/ATen/native/cpu/DistributionKernels.cpp

@@ -27,7 +27,7 @@
 namespace at::native {
 namespace {
 
-static void cauchy_kernel(TensorIteratorBase& iter, double median, double sigma, std::optional<Generator> gen) {
+void cauchy_kernel(TensorIteratorBase& iter, double median, double sigma, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::cauchy_kernel(iter, median, sigma, generator);
 }
@@ -101,7 +101,7 @@ void bernoulli_scalar_kernel(const TensorBase &self, double p, std::optional<Gen
 }
 #endif
 
-static void exponential_kernel_default(TensorIteratorBase& iter, double lambda, std::optional<Generator> gen) {
+void exponential_kernel_default(TensorIteratorBase& iter, double lambda, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::exponential_kernel(iter, lambda, generator);
 }
@@ -198,12 +198,12 @@ void exponential_kernel(TensorIteratorBase &iter, double lambda, std::optional<G
 }
 #endif
 
-static void geometric_kernel(TensorIteratorBase& iter, double p, std::optional<Generator> gen) {
+void geometric_kernel(TensorIteratorBase& iter, double p, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::geometric_kernel(iter, p, generator);
 }
 
-static void log_normal_kernel(TensorIteratorBase& iter, double mean, double std, std::optional<Generator> gen) {
+void log_normal_kernel(TensorIteratorBase& iter, double mean, double std, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::log_normal_kernel(iter, mean, std, generator);
 }
@@ -218,12 +218,12 @@ void normal_kernel(const TensorBase &self, double mean, double std, std::optiona
   templates::cpu::normal_kernel(self, mean, std, generator);
 }
 
-static void random_from_to_kernel(TensorIteratorBase& iter, uint64_t range, int64_t base, std::optional<Generator> gen) {
+void random_from_to_kernel(TensorIteratorBase& iter, uint64_t range, int64_t base, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::random_from_to_kernel(iter, range, base, generator);
 }
 
-static void random_kernel(TensorIteratorBase& iter, std::optional<Generator> gen) {
+void random_kernel(TensorIteratorBase& iter, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::random_kernel(iter, generator);
 }
@@ -231,7 +231,7 @@ static void random_kernel(TensorIteratorBase& iter, std::optional<Generator> gen
 // This is the special kernel to handle single specific case:
 // from(inclusive) = std::numeric_limits<int64_t>::lowest()
 // to(exclusive) = None (= std::numeric_limits<int64_t>::max() + 1)
-static void random_full_64_bits_range_kernel(TensorIteratorBase& iter, std::optional<Generator> gen) {
+void random_full_64_bits_range_kernel(TensorIteratorBase& iter, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::random_full_64_bits_range_kernel(iter, generator);
 }

aten/src/ATen/native/cpu/DistributionTemplates.h

@@ -85,7 +85,7 @@ struct RandomKernel {
 // ==================================================== Normal ========================================================
 
 #ifdef CPU_CAPABILITY_AVX2
-static void normal_fill_16_AVX2(float *data,
+void normal_fill_16_AVX2(float *data,
                          const __m256* two_pi,
                          const __m256* one,
                          const __m256* minus_two,
@@ -136,7 +136,7 @@ void normal_fill_AVX2(const TensorBase &self, const float mean, const float std,
 #endif
 
 template <typename scalar_t>
-static void normal_fill_16(scalar_t *data, const scalar_t mean, const scalar_t std) {
+void normal_fill_16(scalar_t *data, const scalar_t mean, const scalar_t std) {
   for (const auto j : c10::irange(8)) {
     const scalar_t u1 = 1 - data[j]; // [0, 1) -> (0, 1] for log.
     const scalar_t u2 = data[j + 8];

aten/src/ATen/native/cpu/FlashAttentionKernel.cpp

@@ -158,14 +158,14 @@ inline void _mul_reduce_max_fusion_kernel(
 }
 
 template <typename scalar_t>
-static inline scalar_t* conditional_data_ptr(scalar_t* ptr, scalar_t* ptr2) {
+inline scalar_t* conditional_data_ptr(scalar_t* ptr, scalar_t* ptr2) {
   TORCH_CHECK(ptr2 == nullptr);
   return ptr;
 }
 
 template <typename scalar_t,
           typename std::enable_if_t<is_reduced_floating_point_v<scalar_t>, int> = 0>
-static inline scalar_t* conditional_data_ptr(float* ptr, scalar_t* ptr2) {
+inline scalar_t* conditional_data_ptr(float* ptr, scalar_t* ptr2) {
   return ptr2;
 }
 

aten/src/ATen/native/cpu/GridSamplerKernel.cpp

@@ -441,7 +441,7 @@ struct ComputeLocation<scalar_t, GridSamplerPadding::Reflection, align_corners>
 // See NOTE [ Grid Sample CPU Kernels ] for details.
 
 template<typename scalar_t>
-static inline void
+inline void
 mask_scatter_add(const scalar_t *src, scalar_t* base_addr,
                  const int_same_size_t<scalar_t> *offsets,
                  const int_same_size_t<scalar_t> *mask, int64_t len) {
@@ -1030,7 +1030,7 @@ struct ApplyGridSample<scalar_t, 2, GridSamplerInterpolation::Bicubic,
 // See NOTE [ Grid Sample CPU Kernels ] for details.
 
 template<typename scalar_t, typename ApplyFn>
-static inline void grid_sample_2d_grid_slice_iterator(
+inline void grid_sample_2d_grid_slice_iterator(
     const TensorAccessor<const scalar_t, 3>& grid_slice, const ApplyFn &apply_fn) {
   int64_t out_H = grid_slice.size(0);
   int64_t out_W = grid_slice.size(1);

aten/src/ATen/native/cpu/HistogramKernel.cpp

@@ -259,7 +259,7 @@ void histogramdd_out_cpu_template(const Tensor& self, const std::optional<Tensor
  *
  * Refer to histogramdd_out_cpu_template for more details.
  */
-static void histogramdd_kernel_impl(const Tensor& self, const std::optional<Tensor>& weight, bool density,
+void histogramdd_kernel_impl(const Tensor& self, const std::optional<Tensor>& weight, bool density,
         Tensor& hist, const TensorList& bin_edges) {
     histogramdd_out_cpu_template<BINARY_SEARCH>(self, weight, density, hist, bin_edges);
 }
@@ -269,7 +269,7 @@ static void histogramdd_kernel_impl(const Tensor& self, const std::optional<Tens
  *
  * Refer to histogramdd_out_cpu_template for more details.
  */
-static void histogramdd_linear_kernel_impl(const Tensor& self, const std::optional<Tensor>& weight,
+void histogramdd_linear_kernel_impl(const Tensor& self, const std::optional<Tensor>& weight,
         bool density, Tensor& hist, const TensorList& bin_edges, bool local_search) {
     if (local_search) {
         // histogramdd codepath: both hist and bin_edges are eventually returned as output,
@@ -298,7 +298,7 @@ void infer_bin_edges_from_input(const Tensor& input, const int64_t N,
     std::copy(max_data, max_data + N, rightmost_edges.begin());
 }
 
-static void histogram_select_outer_bin_edges_impl(const Tensor& input, const int64_t N,
+void histogram_select_outer_bin_edges_impl(const Tensor& input, const int64_t N,
         std::vector<double> &leftmost_edges, std::vector<double> &rightmost_edges) {
     AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "histogramdd", [&]() {
         infer_bin_edges_from_input<scalar_t>(input, N, leftmost_edges, rightmost_edges);

aten/src/ATen/native/cpu/MultinomialKernel.cpp

@@ -210,7 +210,7 @@ multinomial_with_replacement_apply(
   }
 }
 
-static void multinomial_with_replacement_kernel_impl(
+void multinomial_with_replacement_kernel_impl(
     Tensor& result,
     const Tensor& self,
     const int64_t n_sample,

aten/src/ATen/native/cpu/PaddingKernel.cpp

@@ -96,7 +96,7 @@ struct ReplicationPad {
 };
 
 template <typename scalar_t>
-static inline void copy_stub(scalar_t* out, const scalar_t* in, int64_t size) {
+inline void copy_stub(scalar_t* out, const scalar_t* in, int64_t size) {
   using Vec = Vectorized<scalar_t>;
   int64_t d = 0;
   for (; d < size - (size % Vec::size()); d += Vec::size()) {
@@ -112,7 +112,7 @@ static inline void copy_stub(scalar_t* out, const scalar_t* in, int64_t size) {
 }
 
 template <typename scalar_t>
-static inline void add_stub(scalar_t* grad_in, const scalar_t* grad_out, int64_t size) {
+inline void add_stub(scalar_t* grad_in, const scalar_t* grad_out, int64_t size) {
   using Vec = Vectorized<scalar_t>;
   int64_t d = 0;
   for (; d < size - (size % Vec::size()); d += Vec::size()) {

aten/src/ATen/native/cpu/PointwiseOpsKernel.cpp

@@ -9,7 +9,7 @@
 namespace at::native {
 namespace {
 
-static void addcmul_cpu_kernel(TensorIteratorBase& iter, const Scalar& value) {
+void addcmul_cpu_kernel(TensorIteratorBase& iter, const Scalar& value) {
   ScalarType dtype = iter.common_dtype();
   if (at::isReducedFloatingType(dtype)) {
     AT_DISPATCH_REDUCED_FLOATING_TYPES(dtype, "addcmul_cpu_out", [&]() {
@@ -50,7 +50,7 @@ static void addcmul_cpu_kernel(TensorIteratorBase& iter, const Scalar& value) {
   }
 }
 
-static void addcdiv_cpu_kernel(TensorIteratorBase& iter, const Scalar& value) {
+void addcdiv_cpu_kernel(TensorIteratorBase& iter, const Scalar& value) {
   ScalarType dtype = iter.common_dtype();
   if (at::isReducedFloatingType(dtype)) {
     AT_DISPATCH_REDUCED_FLOATING_TYPES(dtype, "addcdiv_cpu_out", [&]() {
@@ -90,7 +90,7 @@ static void addcdiv_cpu_kernel(TensorIteratorBase& iter, const Scalar& value) {
   }
 }
 
-static void smooth_l1_backward_cpu_kernel(TensorIterator& iter, const Scalar& norm, double beta) {
+void smooth_l1_backward_cpu_kernel(TensorIterator& iter, const Scalar& norm, double beta) {
   ScalarType dtype = iter.dtype(0);
   if (dtype == kBFloat16) {
     auto norm_val = norm.to<float>();
@@ -176,7 +176,7 @@ static void smooth_l1_backward_cpu_kernel(TensorIterator& iter, const Scalar& no
   }
 }
 
-static void huber_backward_cpu_kernel(TensorIterator& iter, const Scalar& norm, double delta) {
+void huber_backward_cpu_kernel(TensorIterator& iter, const Scalar& norm, double delta) {
   ScalarType dtype = iter.dtype(0);
   AT_DISPATCH_FLOATING_TYPES_AND2(kBFloat16, kHalf, dtype, "huber_backward_cpu_out", [&] {
     auto norm_val = norm.to<scalar_t>();
@@ -215,7 +215,7 @@ static void huber_backward_cpu_kernel(TensorIterator& iter, const Scalar& norm,
   });
 }
 
-static void mse_backward_cpu_kernel(TensorIterator& iter, const Scalar& value) {
+void mse_backward_cpu_kernel(TensorIterator& iter, const Scalar& value) {
   ScalarType dtype = iter.dtype(0);
   AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, dtype, "mse_backward_cpu_out", [&] {
     scalar_t scalar_val = value.to<scalar_t>();

aten/src/ATen/native/cpu/RangeFactoriesKernel.cpp

@@ -18,7 +18,7 @@ namespace {
 
 using namespace vec;
 
-static void arange_kernel(TensorIterator& iter, const Scalar& scalar_start, const Scalar& scalar_steps, const Scalar& scalar_step) {
+void arange_kernel(TensorIterator& iter, const Scalar& scalar_start, const Scalar& scalar_steps, const Scalar& scalar_step) {
   AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, iter.dtype(), "arange_cpu", [&]() {
     using accscalar_t = at::acc_type<scalar_t, false>;
     auto start = scalar_start.to<accscalar_t>();
@@ -42,7 +42,7 @@ static void arange_kernel(TensorIterator& iter, const Scalar& scalar_start, cons
   });
 }
 
-static void linspace_kernel(TensorIterator& iter, const Scalar& scalar_start, const Scalar& scalar_end, int64_t steps) {
+void linspace_kernel(TensorIterator& iter, const Scalar& scalar_start, const Scalar& scalar_end, int64_t steps) {
   AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kHalf, kBFloat16, iter.dtype(), "linspace_cpu", [&]() {
     // step should be of double type for all integral types
     using step_t = std::conditional_t<std::is_integral_v<scalar_t>, double, scalar_t>;

aten/src/ATen/native/cpu/ReduceAllOpsKernel.cpp

@@ -62,7 +62,7 @@ inline void reduce_all_impl(
   output.fill_(result);
 }
 
-static void min_all_kernel_impl(Tensor& result, const Tensor& input) {
+void min_all_kernel_impl(Tensor& result, const Tensor& input) {
   if (input.scalar_type() == ScalarType::Bool) {
     TensorIterator iter = TensorIteratorConfig()
       .add_input(input)
@@ -87,7 +87,7 @@ static void min_all_kernel_impl(Tensor& result, const Tensor& input) {
   }
 }
 
-static void max_all_kernel_impl(Tensor& result, const Tensor& input) {
+void max_all_kernel_impl(Tensor& result, const Tensor& input) {
   if (input.scalar_type() == ScalarType::Bool) {
     TensorIterator iter = TensorIteratorConfig()
       .add_input(input)
@@ -167,7 +167,7 @@ inline void reduce_all_impl_vec_two_outputs(
   output2.fill_(result.second);
 }
 
-static void aminmax_allreduce_kernel(
+void aminmax_allreduce_kernel(
     const Tensor& input,
     Tensor& min_result,
     Tensor& max_result) {

aten/src/ATen/native/cpu/ReduceOpsKernel.cpp

@@ -28,7 +28,7 @@ namespace at::native { namespace {
 using namespace vec;
 
 template <typename scalar_t, typename func_t>
-static inline void cpu_cum_base_kernel(const Tensor& result,
+inline void cpu_cum_base_kernel(const Tensor& result,
     const Tensor& self,
     int64_t dim,
     const func_t& f,
@@ -76,7 +76,7 @@ static inline void cpu_cum_base_kernel(const Tensor& result,
   iter.for_each(loop, grain_size);
 }
 
-static void cumsum_cpu_kernel(const Tensor& result, const Tensor& self, int64_t dim) {
+void cumsum_cpu_kernel(const Tensor& result, const Tensor& self, int64_t dim) {
   auto wrap_dim = maybe_wrap_dim(dim, self.dim());
   int64_t self_dim_size = ensure_nonempty_size(self, wrap_dim);
 
@@ -95,7 +95,7 @@ static void cumsum_cpu_kernel(const Tensor& result, const Tensor& self, int64_t
   });
 }
 
-static void cumprod_cpu_kernel(const Tensor& result, const Tensor& self, int64_t dim) {
+void cumprod_cpu_kernel(const Tensor& result, const Tensor& self, int64_t dim) {
   auto wrap_dim = maybe_wrap_dim(dim, self.dim());
   int64_t self_dim_size = ensure_nonempty_size(self, wrap_dim);
 
@@ -114,7 +114,7 @@ static void cumprod_cpu_kernel(const Tensor& result, const Tensor& self, int64_t
   });
 }
 
-static void logcumsumexp_cpu_kernel(Tensor& result, const Tensor& self, int64_t dim) {
+void logcumsumexp_cpu_kernel(Tensor& result, const Tensor& self, int64_t dim) {
   auto wrap_dim = maybe_wrap_dim(dim, self.dim());
   int64_t self_dim_size = ensure_nonempty_size(self, wrap_dim);
 
@@ -135,7 +135,7 @@ static void logcumsumexp_cpu_kernel(Tensor& result, const Tensor& self, int64_t
   });
 }
 
-static void std_var_kernel_impl(TensorIterator& iter, double correction, bool take_sqrt) {
+void std_var_kernel_impl(TensorIterator& iter, double correction, bool take_sqrt) {
   AT_DISPATCH_FLOATING_TYPES_AND2(kHalf, kBFloat16, iter.dtype(), "std_cpu", [&] {
     binary_kernel_reduce(
         iter,
@@ -148,7 +148,7 @@ static void std_var_kernel_impl(TensorIterator& iter, double correction, bool ta
   });
 }
 
-static void prod_kernel_impl(TensorIterator& iter) {
+void prod_kernel_impl(TensorIterator& iter) {
   // Workaround for the error: '*' in boolean context, suggest '&&' instead
   if (iter.dtype() == ScalarType::Bool) {
     using scalar_t = bool;
@@ -203,7 +203,7 @@ void norm_kernel_cpu_impl(TensorIterator& iter, const double& val) {
   }
 }
 
-static void norm_kernel_tensor_iterator_impl(
+void norm_kernel_tensor_iterator_impl(
     TensorIterator& iter,
     const Scalar& p) {
   double val = 0;
@@ -274,7 +274,7 @@ static void norm_kernel_tensor_iterator_impl(
   }
 }
 
-static void and_kernel_impl(TensorIterator& iter) {
+void and_kernel_impl(TensorIterator& iter) {
   if (iter.dtype() == ScalarType::Byte) {
     // Refer [all, any : uint8 compatibility]
     binary_kernel_reduce_vec(
@@ -312,7 +312,7 @@ static void and_kernel_impl(TensorIterator& iter) {
   }
 }
 
-static void or_kernel_impl(TensorIterator& iter) {
+void or_kernel_impl(TensorIterator& iter) {
   if (iter.dtype() == ScalarType::Byte) {
     // Refer [all, any : uint8 compatibility]
     binary_kernel_reduce_vec(
@@ -346,7 +346,7 @@ struct MinValuesOps: public at::native::MinOps<scalar_t> {
   }
 };
 
-static void min_values_kernel_impl(TensorIterator& iter) {
+void min_values_kernel_impl(TensorIterator& iter) {
   if (iter.dtype() == kLong) {
     // This case is special because of Vectorized<int64_t> does not
     // handle upper_bound<int64_t>().
@@ -367,7 +367,7 @@ static void min_values_kernel_impl(TensorIterator& iter) {
   });
 }
 
-static void max_values_kernel_impl(TensorIterator& iter) {
+void max_values_kernel_impl(TensorIterator& iter) {
   AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, iter.dtype(), "max_values_cpu", [&iter] {
     binary_kernel_reduce_vec(
       iter,
@@ -377,7 +377,7 @@ static void max_values_kernel_impl(TensorIterator& iter) {
   });
 }
 
-static void argmax_kernel_impl(TensorIterator &iter) {
+void argmax_kernel_impl(TensorIterator &iter) {
   AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, iter.dtype(1), "argmax_cpu", [&] {
     if (is_reduce_lastdim(iter)) {
       using arg_t = std::pair<scalar_t, int64_t>;
@@ -401,7 +401,7 @@ static void argmax_kernel_impl(TensorIterator &iter) {
   });
 }
 
-static void argmin_kernel_impl(TensorIterator &iter) {
+void argmin_kernel_impl(TensorIterator &iter) {
   AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, iter.dtype(1), "argmin_cpu", [&] {
     if (is_reduce_lastdim(iter)) {
       using arg_t = std::pair<scalar_t, int64_t>;
@@ -459,7 +459,7 @@ struct XorSumOps {
   }
 };
 
-static void xor_sum_kernel_impl(TensorIterator& iter) {
+void xor_sum_kernel_impl(TensorIterator& iter) {
   // Use iter.dtype(1) to dispatch based on the type of the input tensor
   AT_DISPATCH_ALL_TYPES_AND3(
       kBFloat16, kHalf, kBool, iter.dtype(1), "xor_sum_cpu", [&] {

aten/src/ATen/native/cpu/ScatterGatherKernel.cpp

@@ -41,7 +41,7 @@ public:
     *self_data = c10::load(self_data) && c10::load(src_data);
   }
 };
-static ReduceMultiply reduce_multiply;
+ReduceMultiply reduce_multiply;
 
 class ReduceAdd {
 public:
@@ -51,7 +51,7 @@ public:
     *self_data += opmath_t(c10::load(src_data));
   }
 };
-static ReduceAdd reduce_add;
+ReduceAdd reduce_add;
 
 class ReduceMean {
 public:
@@ -61,7 +61,7 @@ public:
     *self_data += opmath_t(c10::load(src_data));
   }
 };
-static ReduceMean reduce_mean;
+ReduceMean reduce_mean;
 
 class ReduceMaximum {
 public:
@@ -73,7 +73,7 @@ public:
     *self_data = at::_isnan<scalar_t>(src_value) ? opmath_t(src_value) : std::max(self_value, opmath_t(src_value));
   }
 };
-static ReduceMaximum reduce_maximum;
+ReduceMaximum reduce_maximum;
 
 class ReduceMinimum {
 public:
@@ -85,7 +85,7 @@ public:
     *self_data = at::_isnan<scalar_t>(src_value) ? opmath_t(src_value) : std::min(self_value, opmath_t(src_value));
   }
 };
-static ReduceMinimum reduce_minimum;
+ReduceMinimum reduce_minimum;
 
 class TensorAssign {
 public:
@@ -95,7 +95,7 @@ public:
     *self_data = opmath_t(c10::load(src_data));
   }
 };
-static TensorAssign tensor_assign;
+TensorAssign tensor_assign;
 
 template <bool is_scatter_like = true>
 struct _cpu_scatter_gather_dim_loop {

aten/src/ATen/native/cpu/SoftMaxKernel.cpp

@@ -968,7 +968,7 @@ struct vec_host_softmax_backward {
   }
 };
 
-static void softmax_lastdim_kernel_impl(
+void softmax_lastdim_kernel_impl(
     const Tensor& result,
     const Tensor& self) {
   AT_DISPATCH_FLOATING_TYPES_AND2(
@@ -977,13 +977,13 @@ static void softmax_lastdim_kernel_impl(
       [&] { vec_host_softmax_lastdim<scalar_t, false>::apply(result, self); });
 }
 
-static void softmax_kernel_impl(const Tensor& result, const Tensor& self, int64_t dim) {
+void softmax_kernel_impl(const Tensor& result, const Tensor& self, int64_t dim) {
   AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::BFloat16, at::ScalarType::Half, self.scalar_type(),
     "softmax_kernel_impl",
     [&] { vec_softmax<scalar_t, false>::apply(result, self, dim); });
 }
 
-static void log_softmax_lastdim_kernel_impl(
+void log_softmax_lastdim_kernel_impl(
     const Tensor& result,
     const Tensor& self) {
   AT_DISPATCH_FLOATING_TYPES_AND2(
@@ -992,13 +992,13 @@ static void log_softmax_lastdim_kernel_impl(
       [&] { vec_host_softmax_lastdim<scalar_t, true>::apply(result, self); });
 }
 
-static void log_softmax_kernel_impl(const Tensor& result, const Tensor& self, int64_t dim) {
+void log_softmax_kernel_impl(const Tensor& result, const Tensor& self, int64_t dim) {
   AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::BFloat16, at::ScalarType::Half, self.scalar_type(),
     "softmax_kernel_impl",
     [&] { vec_softmax<scalar_t, true>::apply(result, self, dim); });
 }
 
-static void softmax_backward_lastdim_kernel_impl(
+void softmax_backward_lastdim_kernel_impl(
     const Tensor& grad_input,
     const Tensor& grad,
     const Tensor& output) {
@@ -1010,7 +1010,7 @@ static void softmax_backward_lastdim_kernel_impl(
       });
 }
 
-static void log_softmax_backward_lastdim_kernel_impl(
+void log_softmax_backward_lastdim_kernel_impl(
     const Tensor& grad_input,
     const Tensor& grad,
     const Tensor& output) {
@@ -1022,7 +1022,7 @@ static void log_softmax_backward_lastdim_kernel_impl(
       });
 }
 
-static void softmax_backward_kernel_impl(
+void softmax_backward_kernel_impl(
     const Tensor& grad_input,
     const Tensor& grad,
     const Tensor& output,
@@ -1038,7 +1038,7 @@ static void softmax_backward_kernel_impl(
       });
 }
 
-static void log_softmax_backward_kernel_impl(
+void log_softmax_backward_kernel_impl(
     const Tensor& grad_input,
     const Tensor& grad,
     const Tensor& output,

aten/src/ATen/native/cpu/SortingKernel.cpp

@@ -90,7 +90,7 @@ struct KeyValueCompDesc {
 };
 
 #ifdef USE_FBGEMM
-static bool can_use_radix_sort(const TensorBase& values, const bool descending) {
+bool can_use_radix_sort(const TensorBase& values, const bool descending) {
   // radix_sort can be used only for 1D data
   if (values.dim() != 1) return false;
   // radix_sort sorts in ascending order
@@ -106,7 +106,7 @@ static bool can_use_radix_sort(const TensorBase& values, const bool descending)
   return true;
 }
 
-static void parallel_sort1d_kernel(
+void parallel_sort1d_kernel(
     const TensorBase& values,
     const TensorBase& indices) {
   AT_DISPATCH_INTEGRAL_TYPES(values.scalar_type(), "parallel_sort1d_kernel", [&] {
@@ -140,7 +140,7 @@ static void parallel_sort1d_kernel(
 #endif
 
 template <typename scalar_t, typename value_accessor_t, typename indices_accessor_t>
-static inline void sort_kernel_impl(const value_accessor_t& value_accessor,
+inline void sort_kernel_impl(const value_accessor_t& value_accessor,
             const indices_accessor_t& indices_accessor,
             int64_t dim_size, bool descending, bool stable) {
   auto composite_accessor = CompositeRandomAccessorCPU<
@@ -165,7 +165,7 @@ static inline void sort_kernel_impl(const value_accessor_t& value_accessor,
   }
 }
 
-static void sort_kernel(
+void sort_kernel(
     const TensorBase& self,
     const TensorBase& values,
     const TensorBase& indices,
@@ -222,7 +222,7 @@ static void sort_kernel(
   );
 }
 
-static void topk_kernel(
+void topk_kernel(
     const TensorBase &values,
     const TensorBase &indices,
     const TensorBase &self,

aten/src/ATen/native/cpu/SumKernel.cpp

@@ -286,12 +286,12 @@ struct CastStoreAccumulate {
 };
 
 template <typename StorePolicy, typename scalar_t>
-static void store(char * C10_RESTRICT data, int64_t stride, int64_t index, scalar_t value) {
+void store(char * C10_RESTRICT data, int64_t stride, int64_t index, scalar_t value) {
   StorePolicy::store(data, stride, index, value);
 }
 
 template <typename StorePolicy, typename scalar_t, size_t numel>
-static void store(char * C10_RESTRICT data, int64_t stride, int64_t index,
+void store(char * C10_RESTRICT data, int64_t stride, int64_t index,
                   const std::array<scalar_t, numel> &values) {
   auto *base_ptr = data + stride * index;
   for (const auto k : c10::irange(numel)) {
@@ -301,7 +301,7 @@ static void store(char * C10_RESTRICT data, int64_t stride, int64_t index,
 }
 
 template <typename StorePolicy, typename scalar_t>
-static void store(char * C10_RESTRICT data, int64_t stride, int64_t index,
+void store(char * C10_RESTRICT data, int64_t stride, int64_t index,
                   const Vectorized<scalar_t> &values) {
   using vec_t = Vectorized<scalar_t>;
   alignas(64) std::array<scalar_t, vec_t::size()> array_values{};

aten/src/ATen/native/cpu/TensorCompareKernel.cpp

@@ -29,7 +29,7 @@
 namespace at::native { namespace {
 
 template <typename scalar_t, typename scalar_t_2 = int64_t, typename loop1d_t>
-static inline void compare_base_kernel_core(
+inline void compare_base_kernel_core(
     const Tensor& result1,
     const Tensor& result2,
     const Tensor& self,
@@ -71,7 +71,7 @@ static inline void compare_base_kernel_core(
 }
 
 template <typename scalar_t, typename scalar_t_2=int64_t, typename func_t>
-static inline void compare_base_kernel(const Tensor& result1, const Tensor& result2,
+inline void compare_base_kernel(const Tensor& result1, const Tensor& result2,
     const Tensor& self,
     int64_t dim,
     bool keepdim,
@@ -98,7 +98,7 @@ static inline void compare_base_kernel(const Tensor& result1, const Tensor& resu
       result1, result2, self, dim, keepdim, loop);
 }
 
-static void min_kernel_impl(
+void min_kernel_impl(
     const Tensor& result,
     const Tensor& indice,
     const Tensor& self,
@@ -131,7 +131,7 @@ static void min_kernel_impl(
   });
 }
 
-static void max_kernel_impl(
+void max_kernel_impl(
     const Tensor& result,
     const Tensor& indice,
     const Tensor& self,
@@ -164,7 +164,7 @@ static void max_kernel_impl(
   });
 }
 
-static void aminmax_kernel(
+void aminmax_kernel(
     const Tensor& self,
     int64_t dim,
     bool keepdim,
@@ -212,7 +212,7 @@ static void aminmax_kernel(
   });
 }
 
-static void where_kernel_impl(TensorIterator &iter) {
+void where_kernel_impl(TensorIterator &iter) {
   AT_DISPATCH_V2(
     iter.dtype(), "where_cpu", [&] {
       cpu_kernel(
@@ -224,19 +224,19 @@ static void where_kernel_impl(TensorIterator &iter) {
   kComplexHalf, kHalf, kBFloat16, kBool, AT_EXPAND(AT_ALL_TYPES_AND_COMPLEX), AT_EXPAND(AT_FLOAT8_TYPES));
 }
 
-static void isposinf_kernel_impl(TensorIteratorBase& iter) {
+void isposinf_kernel_impl(TensorIteratorBase& iter) {
   AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, iter.input_dtype(), "isposinf_cpu", [&]() {
     cpu_kernel(iter, [](scalar_t a) -> bool { return a == std::numeric_limits<scalar_t>::infinity(); });
   });
 }
 
-static void isneginf_kernel_impl(TensorIteratorBase& iter) {
+void isneginf_kernel_impl(TensorIteratorBase& iter) {
   AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, iter.input_dtype(), "isneginf_cpu", [&]() {
     cpu_kernel(iter, [](scalar_t a) -> bool { return a == -std::numeric_limits<scalar_t>::infinity(); });
   });
 }
 
-static void mode_kernel_impl(
+void mode_kernel_impl(
     Tensor& values,
     Tensor& indices,
     const Tensor& self,
@@ -308,7 +308,7 @@ static void mode_kernel_impl(
 
 // Default brute force implementation of isin(). Used when the number of test elements is small.
 // Iterates through each element and checks it against each test element.
-static void isin_default_kernel_cpu(
+void isin_default_kernel_cpu(
     const Tensor& elements,
     const Tensor& test_elements,
     bool invert,
@@ -339,7 +339,7 @@ static void isin_default_kernel_cpu(
   });
 }
 
-static void clamp_kernel_impl(TensorIteratorBase& iter) {
+void clamp_kernel_impl(TensorIteratorBase& iter) {
   AT_DISPATCH_ALL_TYPES_AND2(kBFloat16, kHalf, iter.common_dtype(), "clamp_cpu", [&]() {
     cpu_kernel_vec(iter,
       [](scalar_t a, scalar_t min, scalar_t max) -> scalar_t {
@@ -355,7 +355,7 @@ static void clamp_kernel_impl(TensorIteratorBase& iter) {
   });
 }
 
-static void clamp_scalar_kernel_impl(TensorIteratorBase& iter, const Scalar& min_, const Scalar& max_) {
+void clamp_scalar_kernel_impl(TensorIteratorBase& iter, const Scalar& min_, const Scalar& max_) {
   AT_DISPATCH_ALL_TYPES_AND2(kBFloat16, kHalf, iter.common_dtype(), "clamp_scalar_cpu", [&]() {
     const auto min = min_.to<scalar_t>();
     const auto max = max_.to<scalar_t>();
@@ -371,7 +371,7 @@ static void clamp_scalar_kernel_impl(TensorIteratorBase& iter, const Scalar& min
   });
 }
 
-static void clamp_max_scalar_kernel_impl(TensorIteratorBase& iter, Scalar max_) {
+void clamp_max_scalar_kernel_impl(TensorIteratorBase& iter, Scalar max_) {
   AT_DISPATCH_ALL_TYPES_AND2(kBFloat16, kHalf, iter.common_dtype(), "clamp_max_scalar_cpu", [&]() {
     const auto max = max_.to<scalar_t>();
     const Vectorized<scalar_t> max_vec(max);
@@ -385,7 +385,7 @@ static void clamp_max_scalar_kernel_impl(TensorIteratorBase& iter, Scalar max_)
   });
 }
 
-static void clamp_min_scalar_kernel_impl(TensorIteratorBase& iter, Scalar min_) {
+void clamp_min_scalar_kernel_impl(TensorIteratorBase& iter, Scalar min_) {
   AT_DISPATCH_ALL_TYPES_AND2(kBFloat16, kHalf, iter.common_dtype(), "clamp_min_scalar_cpu", [&]() {
     const auto min = min_.to<scalar_t>();
     const Vectorized<scalar_t> min_vec(min);

aten/src/ATen/native/cpu/Unfold2d.cpp

@@ -13,7 +13,7 @@ namespace at::native {
 namespace {
 
 template <typename scalar_t>
-static inline void cadd(
+inline void cadd(
     scalar_t* z,
     const scalar_t* x,
     const scalar_t* y,
@@ -34,7 +34,7 @@ static inline void cadd(
 }
 
 template <typename scalar_t>
-static void unfolded2d_acc(
+void unfolded2d_acc(
     scalar_t* finput_data,
     scalar_t* input_data,
     int64_t kH,
@@ -113,7 +113,7 @@ static void unfolded2d_acc(
 }
 
 template <typename scalar_t>
-static void unfolded2d_acc_channels_last(
+void unfolded2d_acc_channels_last(
     scalar_t* finput_data,
     scalar_t* input_data,
     int64_t kH,
@@ -225,7 +225,7 @@ void unfolded2d_acc_kernel(
 }
 
 template <typename scalar_t>
-static void unfolded2d_copy(
+void unfolded2d_copy(
     const scalar_t* input_data,
     scalar_t* finput_data,
     int64_t kH,
@@ -326,7 +326,7 @@ static void unfolded2d_copy(
 }
 
 template <typename scalar_t>
-static void unfolded2d_copy_channels_last(
+void unfolded2d_copy_channels_last(
     const scalar_t* input_data,
     scalar_t* finput_data,
     int64_t kH,

aten/src/ATen/native/cpu/UpSampleKernel.cpp

@@ -157,13 +157,13 @@ struct Interpolate<1, scalar_t, opmath_t, index_t, 2> {
 };
 
 template <int n, typename scalar_t, typename index_t, int interp_size>
-static inline scalar_t interpolate(char* src, char** data, const int64_t* strides, int64_t i) {
+inline scalar_t interpolate(char* src, char** data, const int64_t* strides, int64_t i) {
   using opmath_t = at::opmath_type<scalar_t>;
   return Interpolate<n, scalar_t, opmath_t, index_t, interp_size>::eval(src, data, strides, i);
 }
 
 template <typename scalar_t, typename index_t>
-static inline scalar_t interpolate_aa_single_dim_zero_strides(
+inline scalar_t interpolate_aa_single_dim_zero_strides(
     char* src,
     char** data,
     const index_t ids_stride) {
@@ -187,7 +187,7 @@ static inline scalar_t interpolate_aa_single_dim_zero_strides(
 }
 
 template <typename scalar_t, typename index_t>
-static inline scalar_t interpolate_aa_single_dim(
+inline scalar_t interpolate_aa_single_dim(
     char* src,
     char** data,
     const int64_t* strides,
@@ -213,7 +213,7 @@ static inline scalar_t interpolate_aa_single_dim(
 }
 
 template<int m>
-static inline bool is_zero_stride(const int64_t* strides) {
+inline bool is_zero_stride(const int64_t* strides) {
   bool output = strides[0] == 0;
   for (const auto i : c10::irange(1, m)) {
     output &= (strides[i] == 0);
@@ -222,7 +222,7 @@ static inline bool is_zero_stride(const int64_t* strides) {
 }
 
 template <typename scalar_t, typename index_t, int interp_size>
-static inline bool is_contiguous_stride(const int64_t* strides) {
+inline bool is_contiguous_stride(const int64_t* strides) {
   bool output = (strides[0] == sizeof(index_t)) && (strides[1] == sizeof(scalar_t));
   for (int i=2; i<2 * interp_size; i+=2) {
     output &= (strides[i] == sizeof(index_t)) && (strides[i + 1] == sizeof(scalar_t));
@@ -282,13 +282,13 @@ struct CheckAlmostAllZeroStrides<0, non_zero_stride_dim, scalar_t, index_t, inte
 };
 
 template <int n, int s, typename scalar_t, typename index_t, int interp_size>
-static inline bool check_almost_all_zero_stride(const int64_t* strides) {
+inline bool check_almost_all_zero_stride(const int64_t* strides) {
   return CheckAlmostAllZeroStrides<n, s, scalar_t, index_t, interp_size>::eval(strides);
 }
 
 // Helper method to compute interpolation for nearest, linear, cubic modes
 template <typename scalar_t, typename index_t, int out_ndims, int interp_size>
-static inline void basic_loop(char** data, const int64_t* strides, int64_t n) {
+inline void basic_loop(char** data, const int64_t* strides, int64_t n) {
   char* dst = data[0];
   char* src = data[1];
   for (const auto i : c10::irange(n)) {
@@ -298,7 +298,7 @@ static inline void basic_loop(char** data, const int64_t* strides, int64_t n) {
 }
 
 template <typename scalar_t>
-static inline void basic_loop_aa_vertical(
+inline void basic_loop_aa_vertical(
     char** data,
     const int64_t* strides,
     int64_t n,
@@ -354,7 +354,7 @@ inline void basic_loop_aa_vertical<uint8_t>(
 }
 
 template <typename scalar_t>
-static inline void basic_loop_aa_horizontal(
+inline void basic_loop_aa_horizontal(
     char** data,
     const int64_t* strides,
     int64_t n,

aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h

@@ -35,7 +35,7 @@ Like PIL, Pillow is licensed under the open source HPND License
 
 namespace {
 
-static inline __m128i mm_cvtsi32_si128(const uint8_t* C10_RESTRICT ptr, bool i32_aligned) {
+inline __m128i mm_cvtsi32_si128(const uint8_t* C10_RESTRICT ptr, bool i32_aligned) {
   int32_t v;
   if (i32_aligned) {
     v = *(const int32_t*)ptr;
@@ -45,11 +45,11 @@ static inline __m128i mm_cvtsi32_si128(const uint8_t* C10_RESTRICT ptr, bool i32
   return _mm_cvtsi32_si128(v);
 }
 
-static inline __m128i mm_cvtepu8_epi32(const uint8_t* C10_RESTRICT ptr, bool i32_aligned) {
+inline __m128i mm_cvtepu8_epi32(const uint8_t* C10_RESTRICT ptr, bool i32_aligned) {
   return _mm_cvtepu8_epi32(mm_cvtsi32_si128(ptr, i32_aligned));
 }
 
-static inline void _write_endline_rgb_as_uint32(
+inline void _write_endline_rgb_as_uint32(
     uint8_t* C10_RESTRICT output,
     uint32_t data
 ) {