Add an option to put store large mmap weights on disk (#164526)

Summary:
As title

In windows, we cannot modify the .dll to append weights at the end, the windows .dll loader will complain it's not a valid .dll file. So we store the weight blob as a separete file.

1. We add the following API which allows passing in a pointer to the weight blob and get the size of the weight blob.

```cpp
AOTI_API AOTIRuntimeError AOTInductorModelContainerGetConstantsBlobSize(
    AOTInductorModelContainerHandle container_handle,
    uint64_t* ret_size);

// Load weights from a single blob in weight_blob_ptr
AOTI_API AOTIRuntimeError AOTInductorModelUpdateConstantsFromBlob(
    AOTInductorModelContainerHandle container_handle,
    const uint8_t* weight_blob_ptr);
```

2. We also add a method in ModelContainerRunner to load the weight:

If the runner see that there is a `.blob` file in the package, if will mmap the .blob file and use the content to load the constants.

3. We also add the `USE_MMAP_EXTERNAL` macro. When this macro is defined, the model expects to load the weights from external mmap'd weights.


Test Plan:
```
buck run mode/dev-nosan caffe2/test/inductor:test_aot_inductor -- -r test_large_mmaped_weights_on_disk
```

Also tested for windows-cross compilation with 6542566585/demo/main_voxtral.cpp

```
Loaded model.dll
audio_encoder loaded
C:\Users\shangdiy\source\repos\torchnative\demo\token_embedding\data\aotinductor\model\model.wrapper.so
Loaded model.dll
token_embedding loaded
C:\Users\shangdiy\source\repos\torchnative\demo\text_decoder\data\aotinductor\model\model.wrapper.so
Loaded model.dll
Loading weights from C:\Users\shangdiy\source\repos\torchnative\demo\text_decoder\data\aotinductor\model\model.wrapper_weights.blob
text_decoder loaded
Load latency (ms):
  audio_encoder: 1011.234
    archive extraction: 0.000
    .so loading: 1011.197
  token_embedding: 525.773
    archive extraction: 0.000
    .so loading: 525.704
  text_decoder: 3324.130
    archive extraction: 0.000
    .so loading: 3323.979
Run latency (ms):
  audio_encoder: 285.958
    audio_encoder output: dtype=bfloat16, shape=[1, 1125, 3072], numel=3456000
  token_embedding: 6.676
    token_embedding output: dtype=bfloat16, shape=[1, 1138, 3072], numel=3495936
  text_decoder: 576.519
    text_decoder output: dtype=bfloat16, shape=[1, 1138, 131072], numel=149159936
```


cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben

Differential Revision: D84093310

Pulled By: yushangdi

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -8,8 +8,6 @@ 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

@@ -181,7 +181,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950;gfx1100"
+    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
     if [[ $tag =~ "benchmarks" ]]; then
       INDUCTOR_BENCHMARKS=yes
     fi
@@ -344,7 +344,7 @@ docker build \
        --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
-       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}" \
+       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH:-gfx90a;gfx942;gfx1100}" \
        --build-arg "IMAGE_NAME=${IMAGE_NAME}" \
        --build-arg "UCX_COMMIT=${UCX_COMMIT}" \
        --build-arg "UCC_COMMIT=${UCC_COMMIT}" \

.ci/docker/common/install_cache.sh

@@ -9,7 +9,7 @@ install_ubuntu() {
   # Instead use lib and headers from OpenSSL1.1 installed in `install_openssl.sh``
   apt-get install -y cargo
   echo "Checking out sccache repo"
-  git clone https://github.com/mozilla/sccache -b v0.11.0
+  git clone https://github.com/mozilla/sccache -b v0.10.0
   cd sccache
   echo "Building sccache"
   cargo build --release

.ci/docker/manywheel/build_scripts/ssl-check.py

@@ -10,6 +10,11 @@ BAD_SSL = "https://self-signed.badssl.com"
 
 print("Testing SSL certificate checking for Python:", sys.version)
 
+if sys.version_info[:2] < (2, 7) or sys.version_info[:2] < (3, 4):
+    print("This version never checks SSL certs; skipping tests")
+    sys.exit(0)
+
+
 EXC = OSError
 
 print(f"Connecting to {GOOD_SSL} should work")

.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_language_model_test_extended_generation_28_failure_test": {
+        "vllm_languagde_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"
+        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
     )
 
     # the cleaning script to remove torch dependencies from pip

.ci/manywheel/build_cuda.sh

@@ -187,22 +187,19 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             export USE_CUFILE=0
         else
             DEPS_LIST+=(
+                "/usr/local/cuda/lib64/libnvToolsExt.so.1"
                 "/usr/local/cuda/lib64/libcublas.so.12"
                 "/usr/local/cuda/lib64/libcublasLt.so.12"
                 "/usr/local/cuda/lib64/libcudart.so.12"
                 "/usr/local/cuda/lib64/libnvrtc.so.12"
                 "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
             DEPS_SONAME+=(
+                "libnvToolsExt.so.1"
                 "libcublas.so.12"
                 "libcublasLt.so.12"
                 "libcudart.so.12"
                 "libnvrtc.so.12"
                 "libcupti.so.12")
-
-            if [[ $CUDA_VERSION != 12.9* ]]; then
-                DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
-                DEPS_SONAME+=("libnvToolsExt.so.1")
-            fi
         fi
     else
         echo "Using nvidia libs from pypi."

.ci/pytorch/test.sh

@@ -337,13 +337,13 @@ test_python() {
 
 test_python_smoke() {
   # Smoke tests for H100/B200
-  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 
 test_python_smoke_b200() {
   # Targeted smoke tests for B200 - staged approach to avoid too many failures
-  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 

.flake8

@@ -12,7 +12,7 @@ ignore =
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
-    B007,B008,B017,B019,B023,B028,B903,B905,B906,B907,B908,B910
+    B007,B008,B017,B019,B023,B028,B903,B904,B905,B906,B907,B908,B910
     # these ignores are from flake8-comprehensions; please fix!
     C407,
     # these ignores are from flake8-logging-format; please fix!

.github/ISSUE_TEMPLATE/ci-sev.md

@@ -8,7 +8,6 @@ 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: "D❌​\U0001F519​ ISABLE AUTOREVERT"
+name: DISABLE AUTOREVERT
 about: Disables autorevert when open
-title: "[DISABLE AUTOREVERT]"
+title: "❌​\U0001F519​ [DISABLE AUTOREVERT]"
 labels: 'ci: disable-autorevert'
 assignees: ''
 

.github/actions/setup-rocm/action.yml

@@ -111,16 +111,3 @@ runs:
         # This video group ID maps to subgid 1 inside the docker image due to the /etc/subgid entries.
         # The group name corresponding to group ID 1 can change depending on the OS, so both are necessary.
         echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd $DEVICE_FLAG --group-add video --group-add $render_gid --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host" >> "${GITHUB_ENV}"
-
-    - name: configure aws credentials
-      id: aws_creds
-      uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-      with:
-        role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-        aws-region: us-east-1
-        role-duration-seconds: 18000
-
-    - name: Login to Amazon ECR
-      id: login-ecr
-      continue-on-error: true
-      uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-1b013f5b5a87a1882eb143c26d79d091150d6a37
+87ff22e49ed0e92576c4935ccb8c143daac4a3cd

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-faffd5cf673615583da6517275e361cb3dbc77e6
+966da7e46f65d6d49df3e31214470a4fe5cc8e66

.github/ci_commit_pins/vllm.txt

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

.github/ci_commit_pins/xla.txt

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

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -1,41 +1,59 @@
+# 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 system dependencies and uv, then create Python virtual environment
+# Install Python and other dependencies
 RUN apt-get update -y \
-    && 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 ${PYTHON_VERSION} \
-    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
-    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
-    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
-    && ln -s /opt/venv/bin/pip /usr/bin/pip \
+    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
+    && add-apt-repository -y ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${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} \
     && 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
-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
+# 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
 
-# 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
 
@@ -43,32 +61,36 @@ 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 ####################
+# 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
 
-# Only work with PyTorch manylinux builder
+# 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 wget sudo vim; \
+        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
     else \
-        dnf install -y git wget sudo; \
+        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 \
-    python3 -m pip install uv==0.8.4
-
+    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
@@ -76,15 +98,15 @@ 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
@@ -116,36 +138,56 @@ 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 --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
+RUN echo ${TORCH_CUDA_ARCH_LIST}
+RUN echo ${MAX_JOBS}
+RUN pip freeze | grep -E 'ninja'
 
-    pushd xformers
-    git checkout v0.0.32.post2
-    git submodule update --init --recursive
-    python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose
-    popd
+# 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
 
-    rm -rf xformers
-BASH
+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
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system xformers-dist/*.whl
+    uv pip install --system xformers-dist/*.whl --verbose
 
+# 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 \
@@ -155,17 +197,20 @@ 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=${max_jobs}
-ARG nvcc_threads=8
+ARG nvcc_threads=4
 ENV NVCC_THREADS=$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
 
-# Use sccache to speed up compilation
+# if USE_SCCACHE is set, 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 \
@@ -190,9 +235,6 @@ 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
@@ -206,10 +248,17 @@ 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
 
@@ -217,7 +266,7 @@ ARG CUDA_VERSION
 ARG PYTHON_VERSION
 ARG GET_PIP_URL
 
-# Only work with PyTorch manylinux builder
+# TODO (huydhn): Only work with PyTorch manylinux builder
 ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
 
 # prepare for environment starts
@@ -226,19 +275,20 @@ 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 sudo vim python3-pip; \
+        && 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}; \
     else \
-        dnf install -y git wget sudo; \
+        dnf install -y git curl 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 stage
+# Get the torch versions, and whls used in previous stagtes for consistency
 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
@@ -247,29 +297,33 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
     echo "[INFO] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
-# 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
-
-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
 
+# 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
+
+
 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
-# to ./requirements, it will pull the nightly versions using pip. Otherwise,
-# it will use the local wheels from TORCH_WHEELS_PATH
+# 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
 RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
@@ -290,14 +344,18 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Install xformers wheel from previous stage
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
-
-# Build FlashInfer from source
+# Build flashinfer from source.
 ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
+# install package for build flashinfer
+# see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
+
+RUN pip freeze | grep -E 'setuptools|packaging|build'
+
 ENV TORCH_CUDA_ARCH_LIST=${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} \
@@ -309,7 +367,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     && cd .. \
     && rm -rf flashinfer
 
-# Install FlashInfer
+# install flashinfer python
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system wheels/flashinfer/*.whl --verbose
 
@@ -319,6 +377,49 @@ 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

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

.github/scripts/filter_test_configs.py

@@ -512,8 +512,6 @@ 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

@@ -16,18 +16,16 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "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",
 }
 
@@ -40,7 +38,7 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "12.9-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
@@ -78,23 +76,6 @@ 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' | "
@@ -341,7 +322,7 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.9", "12.8", "12.6"]
+                arch_version in ["13.0", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -405,6 +386,5 @@ 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

@@ -2042,6 +2042,10 @@ 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/templates/linux_binary_build_workflow.yml.j2

@@ -177,9 +177,6 @@ jobs:
     runs-on: linux.rocm.gpu.mi250
     timeout-minutes: !{{ common.timeout_minutes }}
     !{{ upload.binary_env(config) }}
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

.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.c7i.2xlarge
+            runner: ${{ inputs.runner_prefix }}linux.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

@@ -107,457 +107,457 @@ jobs:
             All testing is done inside the container, to start an interactive session run:
               docker exec -it $(docker container ps --format '{{.ID}}') bash
 
-      # - name: Checkout PyTorch
-      #   uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
-      #   with:
-      #     no-sudo: true
+      - name: Checkout PyTorch
+        uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
+        with:
+          no-sudo: true
 
-      # - name: Setup Python
-      #   if: contains(matrix.runner, 'b200')
-      #   uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
-      #   with:
-      #     python-version: '3.12'
-      #     cache: pip
+      - name: Setup Python
+        if: contains(matrix.runner, 'b200')
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+        with:
+          python-version: '3.12'
+          cache: pip
 
-      # - name: Setup Linux
-      #   uses: ./.github/actions/setup-linux
-      #   if: inputs.build-environment != 'linux-s390x-binary-manywheel' && !contains(matrix.runner, 'b200')
+      - name: Setup Linux
+        uses: ./.github/actions/setup-linux
+        if: inputs.build-environment != 'linux-s390x-binary-manywheel' && !contains(matrix.runner, 'b200')
 
-      # - name: configure aws credentials
-      #   if: ${{ inputs.aws-role-to-assume != '' && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
-      #   uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-      #   with:
-      #     role-to-assume: ${{ inputs.aws-role-to-assume }}
-      #     role-session-name: gha-linux-test
-      #     aws-region: us-east-1
+      - name: configure aws credentials
+        if: ${{ inputs.aws-role-to-assume != '' && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
+        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+        with:
+          role-to-assume: ${{ inputs.aws-role-to-assume }}
+          role-session-name: gha-linux-test
+          aws-region: us-east-1
 
-      # - name: Login to Amazon ECR
-      #   if: ${{ inputs.aws-role-to-assume != '' && contains(matrix.runner, 'b200') }}
-      #   id: login-ecr
-      #   continue-on-error: true
-      #   uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
+      - name: Login to Amazon ECR
+        if: ${{ inputs.aws-role-to-assume != '' && contains(matrix.runner, 'b200') }}
+        id: login-ecr
+        continue-on-error: true
+        uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
 
-      # - name: Calculate docker image
-      #   id: calculate-docker-image
-      #   uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
-      #   if: inputs.build-environment != 'linux-s390x-binary-manywheel'
-      #   with:
-      #     docker-image-name: ${{ inputs.docker-image }}
+      - name: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        if: inputs.build-environment != 'linux-s390x-binary-manywheel'
+        with:
+          docker-image-name: ${{ inputs.docker-image }}
 
-      # - name: Use following to pull public copy of the image
-      #   id: print-ghcr-mirror
-      #   if: inputs.build-environment != 'linux-s390x-binary-manywheel'
-      #   env:
-      #     ECR_DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      #   shell: bash
-      #   run: |
-      #     tag=${ECR_DOCKER_IMAGE##*:}
-      #     echo "docker pull ghcr.io/pytorch/ci-image:${tag/:/-}"
+      - name: Use following to pull public copy of the image
+        id: print-ghcr-mirror
+        if: inputs.build-environment != 'linux-s390x-binary-manywheel'
+        env:
+          ECR_DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+        shell: bash
+        run: |
+          tag=${ECR_DOCKER_IMAGE##*:}
+          echo "docker pull ghcr.io/pytorch/ci-image:${tag/:/-}"
 
-      # - name: Pull docker image
-      #   uses: pytorch/test-infra/.github/actions/pull-docker-image@main
-      #   if: inputs.build-environment != 'linux-s390x-binary-manywheel'
-      #   with:
-      #     docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Pull docker image
+        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
+        if: inputs.build-environment != 'linux-s390x-binary-manywheel'
+        with:
+          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
 
-      # - name: Check if in a container runner
-      #   shell: bash
-      #   id: check_container_runner
-      #   run: echo "IN_CONTAINER_RUNNER=$(if [ -f /.inarc ] || [ -f /.incontainer ]; then echo true ; else echo false; fi)" >> "$GITHUB_OUTPUT"
+      - name: Check if in a container runner
+        shell: bash
+        id: check_container_runner
+        run: echo "IN_CONTAINER_RUNNER=$(if [ -f /.inarc ] || [ -f /.incontainer ]; then echo true ; else echo false; fi)" >> "$GITHUB_OUTPUT"
 
-      # - name: Install nvidia driver, nvidia-docker runtime, set GPU_FLAG
-      #   id: install-nvidia-driver
-      #   uses: pytorch/test-infra/.github/actions/setup-nvidia@main
-      #   with:
-      #     driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '580.82.07' }}
-      #   if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && !contains(matrix.runner, 'b200') }}
+      - name: Install nvidia driver, nvidia-docker runtime, set GPU_FLAG
+        id: install-nvidia-driver
+        uses: pytorch/test-infra/.github/actions/setup-nvidia@main
+        with:
+          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '580.82.07' }}
+        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && !contains(matrix.runner, 'b200') }}
 
-      # - name: Setup GPU_FLAG for docker run
-      #   id: setup-gpu-flag
-      #   run: echo "GPU_FLAG=--gpus all -e NVIDIA_DRIVER_CAPABILITIES=all" >> "${GITHUB_ENV}"
-      #   if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && (steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' || contains(matrix.runner, 'b200')) }}
+      - name: Setup GPU_FLAG for docker run
+        id: setup-gpu-flag
+        run: echo "GPU_FLAG=--gpus all -e NVIDIA_DRIVER_CAPABILITIES=all" >> "${GITHUB_ENV}"
+        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && (steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' || contains(matrix.runner, 'b200')) }}
 
-      # - name: Setup SCCACHE_SERVER_PORT environment for docker run when on container
-      #   id: setup-sscache-port-flag
-      #   run: echo "SCCACHE_SERVER_PORT_DOCKER_FLAG=-e SCCACHE_SERVER_PORT=$((RUNNER_UID + 4226))" >> "${GITHUB_ENV}"
-      #   if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' && !contains(matrix.runner, 'b200') }}
+      - name: Setup SCCACHE_SERVER_PORT environment for docker run when on container
+        id: setup-sscache-port-flag
+        run: echo "SCCACHE_SERVER_PORT_DOCKER_FLAG=-e SCCACHE_SERVER_PORT=$((RUNNER_UID + 4226))" >> "${GITHUB_ENV}"
+        if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' && !contains(matrix.runner, 'b200') }}
 
-      # - name: Lock NVIDIA A100 40GB Frequency
-      #   run: |
-      #     sudo nvidia-smi -pm 1
-      #     sudo nvidia-smi -ac 1215,1410
-      #     nvidia-smi
-      #   if: ${{ contains(matrix.runner, 'a100') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' }}
+      - name: Lock NVIDIA A100 40GB Frequency
+        run: |
+          sudo nvidia-smi -pm 1
+          sudo nvidia-smi -ac 1215,1410
+          nvidia-smi
+        if: ${{ contains(matrix.runner, 'a100') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' }}
 
-      # - name: Get workflow job id
-      #   id: get-job-id
-      #   uses: ./.github/actions/get-workflow-job-id
-      #   if: always()
-      #   with:
-      #     github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Get workflow job id
+        id: get-job-id
+        uses: ./.github/actions/get-workflow-job-id
+        if: always()
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
-      # - name: Start monitoring script
-      #   id: monitor-script
-      #   if: ${{ !inputs.disable-monitor }}
-      #   shell: bash
-      #   continue-on-error: true
-      #   env:
-      #     JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
-      #     JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
-      #     WORKFLOW_NAME: ${{ github.workflow }}
-      #     WORKFLOW_RUN_ID: ${{github.run_id}}
-      #     MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
-      #     MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
-      #   run: |
-      #     python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
-      #     python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
-      #     echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
+      - name: Start monitoring script
+        id: monitor-script
+        if: ${{ !inputs.disable-monitor }}
+        shell: bash
+        continue-on-error: true
+        env:
+          JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
+          JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
+          WORKFLOW_NAME: ${{ github.workflow }}
+          WORKFLOW_RUN_ID: ${{github.run_id}}
+          MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
+          MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
+        run: |
+          python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7 nvidia-ml-py==11.525.84
+          python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
+          echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 
-      # - name: Download build artifacts
-      #   uses: ./.github/actions/download-build-artifacts
-      #   with:
-      #     name: ${{ inputs.build-environment }}
-      #     s3-bucket: ${{ inputs.s3-bucket }}
-      #     use-gha: ${{ inputs.use-gha }}
+      - name: Download build artifacts
+        uses: ./.github/actions/download-build-artifacts
+        with:
+          name: ${{ inputs.build-environment }}
+          s3-bucket: ${{ inputs.s3-bucket }}
+          use-gha: ${{ inputs.use-gha }}
 
-      # - name: Download TD artifacts
-      #   continue-on-error: true
-      #   uses: ./.github/actions/download-td-artifacts
+      - name: Download TD artifacts
+        continue-on-error: true
+        uses: ./.github/actions/download-td-artifacts
 
-      # - name: Parse ref
-      #   id: parse-ref
-      #   run: .github/scripts/parse_ref.py
+      - name: Parse ref
+        id: parse-ref
+        run: .github/scripts/parse_ref.py
 
-      # - name: Check for keep-going label and re-enabled test issues
-      #   # This uses the filter-test-configs action because it conveniently
-      #   # checks for labels and re-enabled test issues.  It does not actually do
-      #   # any filtering.  All filtering is done in the build step.
-      #   id: keep-going
-      #   uses: ./.github/actions/filter-test-configs
-      #   with:
-      #     github-token: ${{ secrets.GITHUB_TOKEN }}
-      #     test-matrix: ${{ inputs.test-matrix }}
-      #     job-name: ${{ steps.get-job-id.outputs.job-name }}
+      - name: Check for keep-going label and re-enabled test issues
+        # This uses the filter-test-configs action because it conveniently
+        # checks for labels and re-enabled test issues.  It does not actually do
+        # any filtering.  All filtering is done in the build step.
+        id: keep-going
+        uses: ./.github/actions/filter-test-configs
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          test-matrix: ${{ inputs.test-matrix }}
+          job-name: ${{ steps.get-job-id.outputs.job-name }}
 
-      # - name: Set Test step time
-      #   id: test-timeout
-      #   shell: bash
-      #   env:
-      #     JOB_TIMEOUT: ${{ matrix.mem_leak_check == 'mem_leak_check' && 600 || inputs.timeout-minutes }}
-      #   run: |
-      #     echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
+      - name: Set Test step time
+        id: test-timeout
+        shell: bash
+        env:
+          JOB_TIMEOUT: ${{ matrix.mem_leak_check == 'mem_leak_check' && 600 || inputs.timeout-minutes }}
+        run: |
+          echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
 
-      # - name: Preserve github env variables for use in docker
-      #   shell: bash
-      #   run: |
-      #     env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-      #     env | grep '^CI' >> "/tmp/github_env_${GITHUB_RUN_ID}"
+      - name: Preserve github env variables for use in docker
+        shell: bash
+        run: |
+          env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
+          env | grep '^CI' >> "/tmp/github_env_${GITHUB_RUN_ID}"
 
-      # - name: Test
-      #   id: test
-      #   timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
-      #   env:
-      #     BUILD_ENVIRONMENT: ${{ inputs.build-environment }}
-      #     PR_NUMBER: ${{ github.event.pull_request.number }}
-      #     GITHUB_REPOSITORY: ${{ github.repository }}
-      #     GITHUB_WORKFLOW: ${{ github.workflow }}
-      #     GITHUB_JOB: ${{ github.job }}
-      #     GITHUB_RUN_ID: ${{ github.run_id }}
-      #     GITHUB_RUN_NUMBER: ${{ github.run_number }}
-      #     GITHUB_RUN_ATTEMPT: ${{ github.run_attempt }}
-      #     JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
-      #     JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
-      #     BRANCH: ${{ steps.parse-ref.outputs.branch }}
-      #     SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
-      #     BASE_SHA: ${{ github.event.pull_request.base.sha || github.sha }}
-      #     TEST_CONFIG: ${{ matrix.config }}
-      #     SHARD_NUMBER: ${{ matrix.shard }}
-      #     NUM_TEST_SHARDS: ${{ matrix.num_shards }}
-      #     EXTRA_FLAGS: ${{ matrix.extra_flags || '' }}
-      #     OP_BENCHMARK_TESTS: ${{ matrix.op_benchmark_tests }}
-      #     REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
-      #     CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
-      #     VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
-      #     TEST_SHOWLOCALS: ${{ steps.keep-going.outputs.ci-test-showlocals }}
-      #     NO_TEST_TIMEOUT: ${{ steps.keep-going.outputs.ci-no-test-timeout }}
-      #     NO_TD: ${{ steps.keep-going.outputs.ci-no-td }}
-      #     TD_DISTRIBUTED: ${{ steps.keep-going.outputs.ci-td-distributed }}
-      #     # Do not set SCCACHE_S3_KEY_PREFIX to share the cache between all build jobs
-      #     SCCACHE_BUCKET: ${{ !contains(matrix.runner, 'b200') && 'ossci-compiler-cache-circleci-v2' || '' }}
-      #     SCCACHE_REGION: ${{ !contains(matrix.runner, 'b200') && 'us-east-1' || '' }}
-      #     SHM_SIZE: ${{ contains(inputs.build-environment, 'cuda') && '2g' || '1g' }}
-      #     DOCKER_IMAGE: ${{ inputs.docker-image }}
-      #     XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
-      #     XLA_CLANG_CACHE_S3_BUCKET_NAME: ossci-compiler-clang-cache-circleci-xla
-      #     PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
-      #     PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
-      #     DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-      #     VLLM_TEST_HUGGING_FACE_TOKEN: ${{ secrets.VLLM_TEST_HUGGING_FACE_TOKEN }}
-      #     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 }}
-      #   run: |
-      #     set -x
+      - name: Test
+        id: test
+        timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
+        env:
+          BUILD_ENVIRONMENT: ${{ inputs.build-environment }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          GITHUB_REPOSITORY: ${{ github.repository }}
+          GITHUB_WORKFLOW: ${{ github.workflow }}
+          GITHUB_JOB: ${{ github.job }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+          GITHUB_RUN_NUMBER: ${{ github.run_number }}
+          GITHUB_RUN_ATTEMPT: ${{ github.run_attempt }}
+          JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
+          JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
+          BRANCH: ${{ steps.parse-ref.outputs.branch }}
+          SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+          BASE_SHA: ${{ github.event.pull_request.base.sha || github.sha }}
+          TEST_CONFIG: ${{ matrix.config }}
+          SHARD_NUMBER: ${{ matrix.shard }}
+          NUM_TEST_SHARDS: ${{ matrix.num_shards }}
+          EXTRA_FLAGS: ${{ matrix.extra_flags || '' }}
+          OP_BENCHMARK_TESTS: ${{ matrix.op_benchmark_tests }}
+          REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
+          CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
+          VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
+          TEST_SHOWLOCALS: ${{ steps.keep-going.outputs.ci-test-showlocals }}
+          NO_TEST_TIMEOUT: ${{ steps.keep-going.outputs.ci-no-test-timeout }}
+          NO_TD: ${{ steps.keep-going.outputs.ci-no-td }}
+          TD_DISTRIBUTED: ${{ steps.keep-going.outputs.ci-td-distributed }}
+          # Do not set SCCACHE_S3_KEY_PREFIX to share the cache between all build jobs
+          SCCACHE_BUCKET: ${{ !contains(matrix.runner, 'b200') && 'ossci-compiler-cache-circleci-v2' || '' }}
+          SCCACHE_REGION: ${{ !contains(matrix.runner, 'b200') && 'us-east-1' || '' }}
+          SHM_SIZE: ${{ contains(inputs.build-environment, 'cuda') && '2g' || '1g' }}
+          DOCKER_IMAGE: ${{ inputs.docker-image }}
+          XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
+          XLA_CLANG_CACHE_S3_BUCKET_NAME: ossci-compiler-clang-cache-circleci-xla
+          PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
+          PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
+          DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
+          VLLM_TEST_HUGGING_FACE_TOKEN: ${{ secrets.VLLM_TEST_HUGGING_FACE_TOKEN }}
+          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 }}
+        run: |
+          set -x
 
-      #     if [[ $TEST_CONFIG == 'multigpu' ]]; then
-      #       TEST_COMMAND=.ci/pytorch/multigpu-test.sh
-      #     elif [[ $BUILD_ENVIRONMENT == *onnx* ]]; then
-      #       TEST_COMMAND=.ci/onnx/test.sh
-      #     else
-      #       TEST_COMMAND=.ci/pytorch/test.sh
-      #     fi
+          if [[ $TEST_CONFIG == 'multigpu' ]]; then
+            TEST_COMMAND=.ci/pytorch/multigpu-test.sh
+          elif [[ $BUILD_ENVIRONMENT == *onnx* ]]; then
+            TEST_COMMAND=.ci/onnx/test.sh
+          else
+            TEST_COMMAND=.ci/pytorch/test.sh
+          fi
 
-      #     # Leaving 1GB for the runner and other things
-      #     TOTAL_AVAILABLE_MEMORY_IN_GB=$(awk '/MemTotal/ { printf "%.3f \n", $2/1024/1024 - 1 }' /proc/meminfo)
-      #     # https://docs.docker.com/engine/containers/resource_constraints/#--memory-swap-details, the 3GB swap
-      #     # comes from https://github.com/pytorch/test-infra/pull/6058
-      #     TOTAL_MEMORY_WITH_SWAP=$(("${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}" + 3))
+          # Leaving 1GB for the runner and other things
+          TOTAL_AVAILABLE_MEMORY_IN_GB=$(awk '/MemTotal/ { printf "%.3f \n", $2/1024/1024 - 1 }' /proc/meminfo)
+          # https://docs.docker.com/engine/containers/resource_constraints/#--memory-swap-details, the 3GB swap
+          # comes from https://github.com/pytorch/test-infra/pull/6058
+          TOTAL_MEMORY_WITH_SWAP=$(("${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}" + 3))
 
-      #     if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
-      #       SHM_OPTS=
-      #       JENKINS_USER=
-      #       # ensure that docker container cleanly exits in 12 hours
-      #       # if for some reason cleanup action doesn't stop container
-      #       # when job is cancelled
-      #       DOCKER_SHELL_CMD="sleep 12h"
-      #     else
-      #       SHM_OPTS="--shm-size=${SHM_SIZE}"
-      #       JENKINS_USER="--user jenkins"
-      #       DOCKER_SHELL_CMD=
-      #     fi
+          if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
+            SHM_OPTS=
+            JENKINS_USER=
+            # ensure that docker container cleanly exits in 12 hours
+            # if for some reason cleanup action doesn't stop container
+            # when job is cancelled
+            DOCKER_SHELL_CMD="sleep 12h"
+          else
+            SHM_OPTS="--shm-size=${SHM_SIZE}"
+            JENKINS_USER="--user jenkins"
+            DOCKER_SHELL_CMD=
+          fi
 
-      #     # detached container should get cleaned up by teardown_ec2_linux
-      #     # TODO: Stop building test binaries as part of the build phase
-      #     # Used for GPU_FLAG, SHM_OPTS, JENKINS_USER and DOCKER_SHELL_CMD since that doesn't play nice
-      #     # shellcheck disable=SC2086,SC2090
-      #     container_name=$(docker run \
-      #       ${GPU_FLAG:-} \
-      #       ${SCCACHE_SERVER_PORT_DOCKER_FLAG:-} \
-      #       -e BUILD_ENVIRONMENT \
-      #       -e PR_NUMBER \
-      #       -e GITHUB_ACTIONS \
-      #       -e GITHUB_REPOSITORY \
-      #       -e GITHUB_WORKFLOW \
-      #       -e GITHUB_JOB \
-      #       -e GITHUB_RUN_ID \
-      #       -e GITHUB_RUN_NUMBER \
-      #       -e GITHUB_RUN_ATTEMPT \
-      #       -e JOB_ID \
-      #       -e JOB_NAME \
-      #       -e BASE_SHA \
-      #       -e BRANCH \
-      #       -e SHA1 \
-      #       -e AWS_DEFAULT_REGION \
-      #       -e IN_WHEEL_TEST \
-      #       -e SHARD_NUMBER \
-      #       -e TEST_CONFIG \
-      #       -e NUM_TEST_SHARDS \
-      #       -e REENABLED_ISSUES \
-      #       -e CONTINUE_THROUGH_ERROR \
-      #       -e VERBOSE_TEST_LOGS \
-      #       -e TEST_SHOWLOCALS \
-      #       -e NO_TEST_TIMEOUT \
-      #       -e NO_TD \
-      #       -e TD_DISTRIBUTED \
-      #       -e PR_LABELS \
-      #       -e MAX_JOBS="$(nproc --ignore=2)" \
-      #       -e SCCACHE_BUCKET \
-      #       -e SCCACHE_REGION \
-      #       -e XLA_CUDA \
-      #       -e XLA_CLANG_CACHE_S3_BUCKET_NAME \
-      #       -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
-      #       -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
-      #       -e SKIP_SCCACHE_INITIALIZATION=1 \
-      #       -e HUGGING_FACE_HUB_TOKEN \
-      #       -e VLLM_TEST_HUGGING_FACE_TOKEN \
-      #       -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
-      #       -e DASHBOARD_TAG \
-      #       -e ARTIFACTS_FILE_SUFFIX \
-      #       --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
-      #       --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \
-      #       --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
-      #       --security-opt seccomp=unconfined \
-      #       --cap-add=SYS_PTRACE \
-      #       --ipc=host \
-      #       ${SHM_OPTS} \
-      #       --tty \
-      #       --detach \
-      #       --name="${container_name}" \
-      #       ${JENKINS_USER} \
-      #       -v "${GITHUB_WORKSPACE}:/var/lib/jenkins/workspace" \
-      #       -w /var/lib/jenkins/workspace \
-      #       "${DOCKER_IMAGE}" \
-      #       ${DOCKER_SHELL_CMD}
-      #     )
-      #     echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
+          # detached container should get cleaned up by teardown_ec2_linux
+          # TODO: Stop building test binaries as part of the build phase
+          # Used for GPU_FLAG, SHM_OPTS, JENKINS_USER and DOCKER_SHELL_CMD since that doesn't play nice
+          # shellcheck disable=SC2086,SC2090
+          container_name=$(docker run \
+            ${GPU_FLAG:-} \
+            ${SCCACHE_SERVER_PORT_DOCKER_FLAG:-} \
+            -e BUILD_ENVIRONMENT \
+            -e PR_NUMBER \
+            -e GITHUB_ACTIONS \
+            -e GITHUB_REPOSITORY \
+            -e GITHUB_WORKFLOW \
+            -e GITHUB_JOB \
+            -e GITHUB_RUN_ID \
+            -e GITHUB_RUN_NUMBER \
+            -e GITHUB_RUN_ATTEMPT \
+            -e JOB_ID \
+            -e JOB_NAME \
+            -e BASE_SHA \
+            -e BRANCH \
+            -e SHA1 \
+            -e AWS_DEFAULT_REGION \
+            -e IN_WHEEL_TEST \
+            -e SHARD_NUMBER \
+            -e TEST_CONFIG \
+            -e NUM_TEST_SHARDS \
+            -e REENABLED_ISSUES \
+            -e CONTINUE_THROUGH_ERROR \
+            -e VERBOSE_TEST_LOGS \
+            -e TEST_SHOWLOCALS \
+            -e NO_TEST_TIMEOUT \
+            -e NO_TD \
+            -e TD_DISTRIBUTED \
+            -e PR_LABELS \
+            -e MAX_JOBS="$(nproc --ignore=2)" \
+            -e SCCACHE_BUCKET \
+            -e SCCACHE_REGION \
+            -e XLA_CUDA \
+            -e XLA_CLANG_CACHE_S3_BUCKET_NAME \
+            -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
+            -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
+            -e SKIP_SCCACHE_INITIALIZATION=1 \
+            -e HUGGING_FACE_HUB_TOKEN \
+            -e VLLM_TEST_HUGGING_FACE_TOKEN \
+            -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
+            -e DASHBOARD_TAG \
+            -e ARTIFACTS_FILE_SUFFIX \
+            --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
+            --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \
+            --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
+            --security-opt seccomp=unconfined \
+            --cap-add=SYS_PTRACE \
+            --ipc=host \
+            ${SHM_OPTS} \
+            --tty \
+            --detach \
+            --name="${container_name}" \
+            ${JENKINS_USER} \
+            -v "${GITHUB_WORKSPACE}:/var/lib/jenkins/workspace" \
+            -w /var/lib/jenkins/workspace \
+            "${DOCKER_IMAGE}" \
+            ${DOCKER_SHELL_CMD}
+          )
+          echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
 
-      #     if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
-      #       docker exec -t "${container_name}" sh -c "python3 -m pip install -r .ci/docker/requirements-ci.txt"
-      #     fi
+          if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
+            docker exec -t "${container_name}" sh -c "python3 -m pip install -r .ci/docker/requirements-ci.txt"
+          fi
 
-      #     docker exec -t "${container_name}" sh -c "python3 -m pip install $(echo dist/*.whl)[opt-einsum] && ${TEST_COMMAND}"
+          docker exec -t "${container_name}" sh -c "python3 -m pip install $(echo dist/*.whl)[opt-einsum] && ${TEST_COMMAND}"
 
-      # - name: Upload pytest cache if tests failed
-      #   uses: ./.github/actions/pytest-cache-upload
-      #   continue-on-error: true
-      #   if: failure() && steps.test.conclusion && steps.test.conclusion == 'failure' && inputs.build-environment != 'linux-s390x-binary-manywheel'
-      #   with:
-      #     cache_dir: .pytest_cache
-      #     shard: ${{ matrix.shard }}
-      #     sha: ${{ github.event.pull_request.head.sha || github.sha }}
-      #     test_config: ${{ matrix.config }}
-      #     job_identifier: ${{ github.workflow }}_${{ inputs.build-environment }}
+      - name: Upload pytest cache if tests failed
+        uses: ./.github/actions/pytest-cache-upload
+        continue-on-error: true
+        if: failure() && steps.test.conclusion && steps.test.conclusion == 'failure' && inputs.build-environment != 'linux-s390x-binary-manywheel'
+        with:
+          cache_dir: .pytest_cache
+          shard: ${{ matrix.shard }}
+          sha: ${{ github.event.pull_request.head.sha || github.sha }}
+          test_config: ${{ matrix.config }}
+          job_identifier: ${{ github.workflow }}_${{ inputs.build-environment }}
 
-      # - name: Authenticate with AWS
-      #   if: ${{ always() && contains(matrix.runner, 'b200') }}
-      #   uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-      #   with:
-      #     role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results
-      #     # The max duration enforced by the server side
-      #     role-duration-seconds: 18000
-      #     aws-region: us-east-1
+      - name: Authenticate with AWS
+        if: ${{ always() && contains(matrix.runner, 'b200') }}
+        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+        with:
+          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results
+          # The max duration enforced by the server side
+          role-duration-seconds: 18000
+          aws-region: us-east-1
 
-      # - name: Upload the benchmark results
-      #   uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
-      #   if: inputs.build-environment != 'linux-s390x-binary-manywheel'
-      #   with:
-      #     benchmark-results-dir: test/test-reports
-      #     dry-run: false
-      #     schema-version: v3
-      #     github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Upload the benchmark results
+        uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
+        if: inputs.build-environment != 'linux-s390x-binary-manywheel'
+        with:
+          benchmark-results-dir: test/test-reports
+          dry-run: false
+          schema-version: v3
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
-      # - name: Print remaining test logs
-      #   shell: bash
-      #   if: always() && steps.test.conclusion
-      #   run: |
-      #     cat test/**/*_toprint.log || true
+      - name: Print remaining test logs
+        shell: bash
+        if: always() && steps.test.conclusion
+        run: |
+          cat test/**/*_toprint.log || true
 
-      # - name: Stop monitoring script
-      #   if: ${{ always() && steps.monitor-script.outputs.monitor-script-pid }}
-      #   shell: bash
-      #   continue-on-error: true
-      #   env:
-      #     MONITOR_SCRIPT_PID: ${{ steps.monitor-script.outputs.monitor-script-pid }}
-      #   run: |
-      #     kill "$MONITOR_SCRIPT_PID"
+      - name: Stop monitoring script
+        if: ${{ always() && steps.monitor-script.outputs.monitor-script-pid }}
+        shell: bash
+        continue-on-error: true
+        env:
+          MONITOR_SCRIPT_PID: ${{ steps.monitor-script.outputs.monitor-script-pid }}
+        run: |
+          kill "$MONITOR_SCRIPT_PID"
 
-      # - name: Upload test artifacts
-      #   uses: ./.github/actions/upload-test-artifacts
-      #   if: always() && steps.test.conclusion && steps.test.conclusion != 'skipped'
-      #   with:
-      #     file-suffix: ${{ github.job }}-${{ matrix.config }}-${{ matrix.shard }}-${{ matrix.num_shards }}-${{ matrix.runner }}_${{ steps.get-job-id.outputs.job-id }}
-      #     use-gha: ${{ inputs.use-gha }}
-      #     s3-bucket: ${{ inputs.s3-bucket }}
+      - name: Upload test artifacts
+        uses: ./.github/actions/upload-test-artifacts
+        if: always() && steps.test.conclusion && steps.test.conclusion != 'skipped'
+        with:
+          file-suffix: ${{ github.job }}-${{ matrix.config }}-${{ matrix.shard }}-${{ matrix.num_shards }}-${{ matrix.runner }}_${{ steps.get-job-id.outputs.job-id }}
+          use-gha: ${{ inputs.use-gha }}
+          s3-bucket: ${{ inputs.s3-bucket }}
 
-      # - name: Collect backtraces from coredumps (if any)
-      #   if: always()
-      #   run: |
-      #     # shellcheck disable=SC2156
-      #     find . -iname "core.[1-9]*" -exec docker exec "${DOCKER_CONTAINER_ID}" sh -c "gdb python {} -ex 'bt' -ex 'q'" \;
+      - name: Collect backtraces from coredumps (if any)
+        if: always()
+        run: |
+          # shellcheck disable=SC2156
+          find . -iname "core.[1-9]*" -exec docker exec "${DOCKER_CONTAINER_ID}" sh -c "gdb python {} -ex 'bt' -ex 'q'" \;
 
-      # - name: Store Core dumps on S3
-      #   uses: seemethere/upload-artifact-s3@baba72d0712b404f646cebe0730933554ebce96a # v5.1.0
-      #   if: failure()
-      #   with:
-      #     name: coredumps-${{ matrix.config }}-${{ matrix.shard }}-${{ matrix.num_shards }}-${{ matrix.runner }}
-      #     retention-days: 14
-      #     if-no-files-found: ignore
-      #     path: ./**/core.[1-9]*
+      - name: Store Core dumps on S3
+        uses: seemethere/upload-artifact-s3@baba72d0712b404f646cebe0730933554ebce96a # v5.1.0
+        if: failure()
+        with:
+          name: coredumps-${{ matrix.config }}-${{ matrix.shard }}-${{ matrix.num_shards }}-${{ matrix.runner }}
+          retention-days: 14
+          if-no-files-found: ignore
+          path: ./**/core.[1-9]*
 
-      # - name: Upload utilization stats
-      #   if: ${{ always() && steps.test.conclusion && steps.test.conclusion != 'skipped' && !inputs.disable-monitor && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
-      #   continue-on-error: true
-      #   uses: ./.github/actions/upload-utilization-stats
-      #   with:
-      #     job_id: ${{ steps.get-job-id.outputs.job-id }}
-      #     job_name: ${{ steps.get-job-id.outputs.job-name }}
-      #     workflow_name: ${{ github.workflow }}
-      #     workflow_run_id: ${{github.run_id}}
-      #     workflow_attempt: ${{github.run_attempt}}
+      - name: Upload utilization stats
+        if: ${{ always() && steps.test.conclusion && steps.test.conclusion != 'skipped' && !inputs.disable-monitor && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
+        continue-on-error: true
+        uses: ./.github/actions/upload-utilization-stats
+        with:
+          job_id: ${{ steps.get-job-id.outputs.job-id }}
+          job_name: ${{ steps.get-job-id.outputs.job-name }}
+          workflow_name: ${{ github.workflow }}
+          workflow_run_id: ${{github.run_id}}
+          workflow_attempt: ${{github.run_attempt}}
 
-      # - name: Teardown Linux
-      #   uses: pytorch/test-infra/.github/actions/teardown-linux@main
-      #   if: always() && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false'
+      - name: Teardown Linux
+        uses: pytorch/test-infra/.github/actions/teardown-linux@main
+        if: always() && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false'
 
-      # # NB: We are currently having an intermittent GPU-related issue on G5 runners with
-      # # A10G GPU. Once this happens, trying to reset the GPU as done in setup-nvidia does
-      # # not seem to help. Here are some symptoms:
-      # #   * Calling nvidia-smi timeouts after 60 second
-      # #   * Fail to run nvidia-smi with an unable to determine the device handle for GPU
-      # #     unknown error
-      # #   * Test fails with a missing CUDA GPU error when initializing CUDA in PyTorch
-      # #   * Run docker --gpus all fails with error response from daemon
-      # #
-      # # As both the root cause and recovery path are unclear, let's take the runner out of
-      # # service so that it doesn't get any more jobs
-      # - name: Check NVIDIA driver installation step
-      #   if: failure() && steps.install-nvidia-driver.outcome && steps.install-nvidia-driver.outcome != 'skipped'
-      #   shell: bash
-      #   run: |
-      #     set +e
-      #     set -x
+      # NB: We are currently having an intermittent GPU-related issue on G5 runners with
+      # A10G GPU. Once this happens, trying to reset the GPU as done in setup-nvidia does
+      # not seem to help. Here are some symptoms:
+      #   * Calling nvidia-smi timeouts after 60 second
+      #   * Fail to run nvidia-smi with an unable to determine the device handle for GPU
+      #     unknown error
+      #   * Test fails with a missing CUDA GPU error when initializing CUDA in PyTorch
+      #   * Run docker --gpus all fails with error response from daemon
+      #
+      # As both the root cause and recovery path are unclear, let's take the runner out of
+      # service so that it doesn't get any more jobs
+      - name: Check NVIDIA driver installation step
+        if: failure() && steps.install-nvidia-driver.outcome && steps.install-nvidia-driver.outcome != 'skipped'
+        shell: bash
+        run: |
+          set +e
+          set -x
 
-      #     nvidia-smi
-      #     # NB: Surprisingly, nvidia-smi command returns successfully with return code 0 even in
-      #     # the case where the driver has already crashed as it still can get the driver version
-      #     # and some basic information like the bus ID.  However, the rest of the information
-      #     # would be missing (ERR!), for example:
-      #     #
-      #     # +-----------------------------------------------------------------------------+
-      #     # | NVIDIA-SMI 525.89.02    Driver Version: 525.89.02    CUDA Version: 12.0     |
-      #     # |-------------------------------+----------------------+----------------------+
-      #     # | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
-      #     # | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
-      #     # |                               |                      |               MIG M. |
-      #     # |===============================+======================+======================|
-      #     # |   0  ERR!                Off  | 00000000:00:1E.0 Off |                 ERR! |
-      #     # |ERR!  ERR! ERR!    ERR! / ERR! |   4184MiB / 23028MiB |    ERR!      Default |
-      #     # |                               |                      |                 ERR! |
-      #     # +-------------------------------+----------------------+----------------------+
-      #     #
-      #     # +-----------------------------------------------------------------------------+
-      #     # | Processes:                                                                  |
-      #     # |  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
-      #     # |        ID   ID                                                   Usage      |
-      #     # |=============================================================================|
-      #     # +-----------------------------------------------------------------------------+
-      #     #
-      #     # This should be reported as a failure instead as it will guarantee to fail when
-      #     # Docker tries to run with --gpus all
-      #     #
-      #     # So, the correct check here is to query one of the missing piece of info like
-      #     # GPU name, so that the command can fail accordingly
-      #     nvidia-smi --query-gpu=gpu_name --format=csv,noheader --id=0
-      #     NVIDIA_SMI_STATUS=$?
+          nvidia-smi
+          # NB: Surprisingly, nvidia-smi command returns successfully with return code 0 even in
+          # the case where the driver has already crashed as it still can get the driver version
+          # and some basic information like the bus ID.  However, the rest of the information
+          # would be missing (ERR!), for example:
+          #
+          # +-----------------------------------------------------------------------------+
+          # | NVIDIA-SMI 525.89.02    Driver Version: 525.89.02    CUDA Version: 12.0     |
+          # |-------------------------------+----------------------+----------------------+
+          # | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
+          # | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
+          # |                               |                      |               MIG M. |
+          # |===============================+======================+======================|
+          # |   0  ERR!                Off  | 00000000:00:1E.0 Off |                 ERR! |
+          # |ERR!  ERR! ERR!    ERR! / ERR! |   4184MiB / 23028MiB |    ERR!      Default |
+          # |                               |                      |                 ERR! |
+          # +-------------------------------+----------------------+----------------------+
+          #
+          # +-----------------------------------------------------------------------------+
+          # | Processes:                                                                  |
+          # |  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
+          # |        ID   ID                                                   Usage      |
+          # |=============================================================================|
+          # +-----------------------------------------------------------------------------+
+          #
+          # This should be reported as a failure instead as it will guarantee to fail when
+          # Docker tries to run with --gpus all
+          #
+          # So, the correct check here is to query one of the missing piece of info like
+          # GPU name, so that the command can fail accordingly
+          nvidia-smi --query-gpu=gpu_name --format=csv,noheader --id=0
+          NVIDIA_SMI_STATUS=$?
 
-      #     # These are acceptable return code from nvidia-smi as copied from setup-nvidia GitHub action
-      #     if [ "$NVIDIA_SMI_STATUS" -ne 0 ] && [ "$NVIDIA_SMI_STATUS" -ne 14 ]; then
-      #       echo "NVIDIA driver installation has failed, shutting down the runner..."
-      #       .github/scripts/stop_runner_service.sh
-      #     fi
+          # These are acceptable return code from nvidia-smi as copied from setup-nvidia GitHub action
+          if [ "$NVIDIA_SMI_STATUS" -ne 0 ] && [ "$NVIDIA_SMI_STATUS" -ne 14 ]; then
+            echo "NVIDIA driver installation has failed, shutting down the runner..."
+            .github/scripts/stop_runner_service.sh
+          fi
 
-      #     # For runner with multiple GPUs, we also want to confirm that the number of GPUs are the
-      #     # power of 2, i.e. 1, 2, 4, or 8. This is to avoid flaky test issue when one GPU fails
-      #     # https://github.com/pytorch/test-infra/issues/4000
-      #     GPU_COUNT=$(nvidia-smi --list-gpus | wc -l)
-      #     NVIDIA_SMI_STATUS=$?
+          # For runner with multiple GPUs, we also want to confirm that the number of GPUs are the
+          # power of 2, i.e. 1, 2, 4, or 8. This is to avoid flaky test issue when one GPU fails
+          # https://github.com/pytorch/test-infra/issues/4000
+          GPU_COUNT=$(nvidia-smi --list-gpus | wc -l)
+          NVIDIA_SMI_STATUS=$?
 
-      #     # These are acceptable return code from nvidia-smi as copied from setup-nvidia GitHub action
-      #     if [ "$NVIDIA_SMI_STATUS" -ne 0 ] && [ "$NVIDIA_SMI_STATUS" -ne 14 ]; then
-      #       echo "NVIDIA driver installation has failed, shutting down the runner..."
-      #       .github/scripts/stop_runner_service.sh
-      #     fi
+          # These are acceptable return code from nvidia-smi as copied from setup-nvidia GitHub action
+          if [ "$NVIDIA_SMI_STATUS" -ne 0 ] && [ "$NVIDIA_SMI_STATUS" -ne 14 ]; then
+            echo "NVIDIA driver installation has failed, shutting down the runner..."
+            .github/scripts/stop_runner_service.sh
+          fi
 
-      #     # Check the GPU count to be a power of 2
-      #     if [ "$GPU_COUNT" -le 8 ] && [ "$GPU_COUNT" -ne 1 ] && [ "$GPU_COUNT" -ne 2 ] && [ "$GPU_COUNT" -ne 4 ] && [ "$GPU_COUNT" -ne 8 ]; then
-      #       echo "NVIDIA driver detects $GPU_COUNT GPUs. The runner has a broken GPU, shutting it down..."
-      #       .github/scripts/stop_runner_service.sh
-      #     fi
+          # Check the GPU count to be a power of 2
+          if [ "$GPU_COUNT" -le 8 ] && [ "$GPU_COUNT" -ne 1 ] && [ "$GPU_COUNT" -ne 2 ] && [ "$GPU_COUNT" -ne 4 ] && [ "$GPU_COUNT" -ne 8 ]; then
+            echo "NVIDIA driver detects $GPU_COUNT GPUs. The runner has a broken GPU, shutting it down..."
+            .github/scripts/stop_runner_service.sh
+          fi
 
-      # - name: Cleanup docker
-      #   if: always() && inputs.build-environment == 'linux-s390x-binary-manywheel'
-      #   shell: bash
-      #   run: |
-      #     # on s390x stop the container for clean worker stop
-      #     docker stop -a || true
-      #     docker kill -a || true
+      - name: Cleanup docker
+        if: always() && inputs.build-environment == 'linux-s390x-binary-manywheel'
+        shell: bash
+        run: |
+          # on s390x stop the container for clean worker stop
+          docker stop -a || true
+          docker kill -a || true

.github/workflows/_rocm-test.yml

@@ -102,6 +102,19 @@ jobs:
             exit 1
           fi
 
+      - name: configure aws credentials
+        id: aws_creds
+        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+        with:
+          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+          aws-region: us-east-1
+          role-duration-seconds: 18000
+
+      - name: Login to Amazon ECR
+        id: login-ecr
+        continue-on-error: true
+        uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
+
       - name: Calculate docker image
         id: calculate-docker-image
         uses: pytorch/test-infra/.github/actions/calculate-docker-image@main

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

@@ -46,12 +46,10 @@ 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

@@ -204,52 +204,6 @@ 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
@@ -453,52 +407,6 @@ 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
@@ -702,52 +610,6 @@ 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
@@ -951,52 +813,6 @@ 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
@@ -1200,52 +1016,6 @@ 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
@@ -1449,52 +1219,6 @@ 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
@@ -1698,52 +1422,6 @@ 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

@@ -248,74 +248,6 @@ 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
@@ -426,9 +358,6 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -544,9 +473,6 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

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

@@ -241,72 +241,6 @@ 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
@@ -413,9 +347,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.10"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -528,9 +459,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.10"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -907,72 +835,6 @@ 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
@@ -1079,9 +941,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.11"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1194,9 +1053,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.11"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1573,72 +1429,6 @@ 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
@@ -1745,9 +1535,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.12"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1860,9 +1647,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.12"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2239,72 +2023,6 @@ 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
@@ -2411,9 +2129,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.13"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2526,9 +2241,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.13"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2905,72 +2617,6 @@ 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
@@ -3077,9 +2723,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.13t"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3192,9 +2835,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.13t"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3571,72 +3211,6 @@ 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
@@ -3743,9 +3317,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.14"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3858,9 +3429,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.14"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -4237,72 +3805,6 @@ 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
@@ -4409,9 +3911,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.14t"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -4524,9 +4023,6 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.14t"
-    permissions:
-      id-token: write
-      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

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

@@ -788,256 +788,6 @@ 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

@@ -788,256 +788,6 @@ 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-h100.yml

@@ -130,7 +130,7 @@ jobs:
     name: test-periodically
     uses: ./.github/workflows/_linux-test.yml
     needs: build
-    if: github.event.schedule == '15 0 * * 1-6'
+    if: github.event.schedule == '15 0,12 * * 1-6'
     with:
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true

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

@@ -1,132 +0,0 @@
-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.yml

@@ -1,11 +1,11 @@
-name: inductor-perf-nightly-rocm-mi355
+name: inductor-perf-nightly-rocm
 
 on:
   push:
     tags:
-      - ciflow/inductor-perf-test-nightly-rocm-mi355/*
+      - ciflow/inductor-perf-test-nightly-rocm/*
   schedule:
-    - cron: 15 0 * * *
+    - cron: 0 7 * * 0,3
   # 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_mi355,inductor_timm_perf_rocm_mi355,inductor_torchbench_perf_rocm_mi355
+        default: inductor_huggingface_perf_rocm,inductor_timm_perf_rocm,inductor_torchbench_perf_rocm
 
 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' }}
@@ -88,27 +88,23 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
       test-matrix: |
         { include: [
-          { 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" },
+          { 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" },
         ]}
     secrets: inherit
 

.github/workflows/lint.yml

@@ -12,7 +12,6 @@ on:
       - landchecks/*
     tags:
       - ciflow/pull/*
-      - ciflow/trunk/*
   workflow_dispatch:
 
 permissions: read-all
@@ -33,12 +32,10 @@ jobs:
     name: Get changed files
     uses: ./.github/workflows/_get-changed-files.yml
     with:
-      all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') || github.event_name == 'push' }}
+      all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') }}
 
   lintrunner-clang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    # Needed to prevent deduping on HUD
-    name: lintrunner-clang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     # Only run if there are changed files relevant to clangtidy / clangformat
     if: |
@@ -78,7 +75,6 @@ jobs:
   #       fails to find types when it should
   lintrunner-mypy:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    name: lintrunner-mypy-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     # Only run if there are changed files relevant to mypy
     if: |
@@ -103,7 +99,6 @@ jobs:
 
   lintrunner-noclang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    name: lintrunner-noclang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     with:
       timeout: 120

.github/workflows/periodic.yml

@@ -182,11 +182,11 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11
       test-matrix: |
         { include: [
-          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
           { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
         ]}
     secrets: inherit

.github/workflows/pull.yml

@@ -127,7 +127,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

.github/workflows/slow.yml

@@ -140,7 +140,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

.github/workflows/trunk.yml

@@ -56,7 +56,7 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       build-generates-artifacts: false
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: "linux.c7i.4xlarge"
+      runner: "linux.4xlarge"
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 1 },
@@ -180,13 +180,13 @@ jobs:
       disable-monitor: false
     secrets: inherit
 
-  win-vs2022-cuda12_8-py3-build:
-    name: win-vs2022-cuda12.8-py3
+  win-vs2022-cuda12_6-py3-build:
+    name: win-vs2022-cuda12.6-py3
     uses: ./.github/workflows/_win-build.yml
     needs: get-label-type
     with:
-      build-environment: win-vs2022-cuda12.8-py3
-      cuda-version: "12.8"
+      build-environment: win-vs2022-cuda12.6-py3
+      cuda-version: "12.6"
       runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
     secrets: inherit
 

.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_language_model_test_extended_generation_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_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

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

aten/src/ATen/CMakeLists.txt

@@ -256,7 +256,6 @@ endif()
 IF(USE_FBGEMM_GENAI)
   set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
   set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
-
   if(USE_CUDA)
     # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
     # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
@@ -293,64 +292,58 @@ IF(USE_FBGEMM_GENAI)
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
     )
 
-    target_include_directories(fbgemm_genai PRIVATE
+    target_include_directories(fbgemm_genai PUBLIC
       ${FBGEMM_THIRD_PARTY}/cutlass/include
       ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
       ${fbgemm_genai_mx8mx8bf16_grouped}
       ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
       ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
     )
+  else()
+    if(USE_ROCM)
+      # Only include the kernels we want to build to avoid increasing binary size.
+      file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
+        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
+        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
+      set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-    # Add FBGEMM_GENAI include directories for torch_ops.h
-    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
-  elseif(USE_ROCM)
-    # Only include the kernels we want to build to avoid increasing binary size.
-    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
+      # Add additional HIPCC compiler flags for performance
+      set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+        -mllvm
+        -amdgpu-coerce-illegal-types=1
+        -mllvm
+        -enable-post-misched=0
+        -mllvm
+        -greedy-reverse-local-assignment=1
+        -fhip-new-launch-api)
 
-    # Add additional HIPCC compiler flags for performance
-    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-      -mllvm
-      -amdgpu-coerce-illegal-types=1
-      -mllvm
-      -enable-post-misched=0
-      -mllvm
-      -greedy-reverse-local-assignment=1
-      -fhip-new-launch-api)
+      # Only compile for gfx942 for now.
+      # This is rather hacky, I could not figure out a clean solution :(
+      set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
+      string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
+      if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
+        list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
+      endif()
+      set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
 
-    # Only compile for gfx942 for now.
-    # This is rather hacky, I could not figure out a clean solution :(
-    set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
-    string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
-    if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-      list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
+      hip_add_library(
+        fbgemm_genai STATIC
+        ${fbgemm_genai_native_rocm_hip}
+        HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
+      set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
+      set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
+      target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
+
+      target_include_directories(fbgemm_genai PUBLIC
+        # FBGEMM version of Composable Kernel is used due to some customizations
+        ${FBGEMM_THIRD_PARTY}/composable_kernel/include
+        ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
+        ${FBGEMM_THIRD_PARTY}/cutlass/include
+        ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
+        ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
+        ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
+      )
     endif()
-    set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
-
-    hip_add_library(
-      fbgemm_genai STATIC
-      ${fbgemm_genai_native_rocm_hip}
-      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
-    set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
-    set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
-
-    target_include_directories(fbgemm_genai PRIVATE
-      # FBGEMM version of Composable Kernel is used due to some customizations
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-      ${FBGEMM_THIRD_PARTY}/cutlass/include
-      ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
-      ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
-    )
-
-    # Add FBGEMM_GENAI include directories for torch_ops.h
-    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
   endif()
 endif()
 
@@ -699,6 +692,12 @@ if(USE_CUDA AND NOT USE_ROCM)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
 
+  # Add FBGEMM_GENAI include directories for torch_ops.h
+  if(USE_FBGEMM_GENAI)
+    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
+  endif()
+
   if($ENV{ATEN_STATIC_CUDA})
     if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
       list(APPEND ATen_CUDA_DEPENDENCY_LIBS

aten/src/ATen/DLConvertor.cpp

@@ -389,16 +389,37 @@ 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 = src;
+  atDLMTensor->handle = view;
   atDLMTensor->tensor.manager_ctx = atDLMTensor;
   atDLMTensor->tensor.deleter = &deleter<T>;
-  atDLMTensor->tensor.dl_tensor.data = src.data_ptr();
+  atDLMTensor->tensor.dl_tensor.data = view.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*>(src.sizes().data());
-  atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(src.strides().data());
+  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.byte_offset = 0;
   fillVersion(&atDLMTensor->tensor);
 

aten/src/ATen/DLConvertor.h

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

aten/src/ATen/EmptyTensor.h

@@ -16,8 +16,8 @@ inline void check_size_nonnegative(ArrayRef<int64_t> size) {
 
 inline void check_size_nonnegative(ArrayRef<c10::SymInt> size) {
   for (const auto& x : size) {
-    TORCH_SYM_CHECK(
-        x.sym_ge(0),
+    TORCH_CHECK(
+        x.expect_size(__FILE__, __LINE__),
         "Trying to create tensor with negative dimension ",
         x,
         ": ",

aten/src/ATen/InferSize.h

@@ -4,7 +4,6 @@
 #include <c10/core/ScalarType.h>
 #include <c10/core/SymIntArrayRef.h>
 #include <c10/util/DimVector.h>
-#include <c10/util/Exception.h>
 #include <optional>
 #include <sstream>
 #include <vector>
@@ -27,7 +26,9 @@ inline void infer_size_impl(
   std::optional<int64_t> infer_dim;
   for (int64_t dim = 0, ndim = shape.size(); dim != ndim; dim++) {
     if (TORCH_GUARD_OR_FALSE(sym_eq(shape[dim], -1))) {
-      TORCH_CHECK(!infer_dim, "only one dimension can be inferred");
+      if (infer_dim) {
+        throw std::runtime_error("only one dimension can be inferred");
+      }
       infer_dim = dim;
     } else {
       // in case of unbacked shape[dim] we assume it's not -1 and add a runtime

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.
-bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
+static 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);
 }
 
-int64_t getGradInputPhysicalDim(int64_t dim, IntArrayRef input_sizes, int64_t num_batch_dims) {
+static 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.
-void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_batch_dims) {
+static 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 @@ void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_bat
 // 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).
-std::optional<int64_t> maximum_indexable_location(
+static 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 @@ 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.
-void checkBasicAsStridedValidForSlice(
+static void checkBasicAsStridedValidForSlice(
     const Tensor& physical_tensor,
     int64_t num_batch_dims,
     IntArrayRef sizes,

aten/src/ATen/PythonTorchFunctionTLS.cpp

@@ -42,14 +42,8 @@ const PythonTorchFunctionTLS& PythonTorchFunctionTLS::get_state() {
 }
 
 bool torch_function_mode_enabled() {
-  // 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();
+  return PythonTorchFunctionTLS::get_disabled_state() != TorchFunctionDisabledState::ALL_DISABLED &&
+         PythonTorchFunctionTLS::stack_len() > 0;
 }
 
 // This is needed to disambiguate the ternary torch function disabled states

aten/src/ATen/PythonTorchFunctionTLS.h

@@ -27,7 +27,6 @@ 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_.
-  bool is_initialized(false);
+  static 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_
   }
 }
 
-OptionalTensorRef make_otr(const TensorBase &tensor) {
+static OptionalTensorRef make_otr(const TensorBase &tensor) {
   if (tensor.defined()) {
     return OptionalTensorRef(tensor);
   } else {

aten/src/ATen/Version.cpp

@@ -103,7 +103,9 @@ std::string get_cpu_capability() {
 #elif defined(HAVE_ZVECTOR_CPU_DEFINITION)
     case native::CPUCapability::ZVECTOR:
       return "Z VECTOR";
-#elif defined(HAVE_SVE256_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
+#elif defined(HAVE_SVE_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
+    case native::CPUCapability::SVE128:
+      return "SVE128";
     case native::CPUCapability::SVE256:
       return "SVE256";
 #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>;
 
-ska::flat_hash_map<TensorImpl*, val_type>& get_cached_casts() {
+static 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 {
 
-std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
+static std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
     allocator_array{};
-std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
+static std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
     allocator_priority{};
 
 } // anonymous namespace

aten/src/ATen/core/ivalue.h

@@ -624,14 +624,7 @@ 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/cpu/vec/functional_base.h

@@ -102,8 +102,31 @@ struct VecReduceAllSIMD<float, Op> {
 #endif // defined(__GNUC__) && (__GNUC__ > 5) && !defined(_MSC_VER) &&
        // !defined(C10_MOBILE)
 
-#if defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
-    !defined(CPU_CAPABILITY_SVE)
+#if defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__)
+#if defined(CPU_CAPABILITY_SVE256)
+template <typename Op>
+struct VecReduceAllSIMD<float, Op> {
+  static inline float apply(
+      const Op& vec_fun,
+      const Vectorized<float>& acc_vec) {
+    using Vec = Vectorized<float>;
+    Vec v = acc_vec;
+    // 128-bit shuffle
+    svuint32_t ind = svdupq_n_u32(4, 5, 6, 7);
+    Vec v1 = svtbl_f32(v, ind);
+    v = vec_fun(v, v1);
+    // 64-bit shuffle
+    ind = svdupq_n_u32(2, 3, 0, 1);
+    v1 = svtbl_f32(v, ind);
+    v = vec_fun(v, v1);
+    // 32-bit shuffle
+    ind = svdupq_n_u32(1, 0, 2, 3);
+    v1 = svtbl_f32(v, ind);
+    v = vec_fun(v, v1);
+    return svlasta(svpfalse(), v);
+  }
+};
+#else
 template <typename Op>
 struct VecReduceAllSIMD<float, Op> {
   static inline float apply(
@@ -140,35 +163,8 @@ struct VecReduceAllSIMD<float, std::plus<Vectorized<float>>> {
     return vaddvq_f32(acc_vec);
   }
 };
+#endif // defined(CPU_CAPABILITY_SVE256)
 #endif // defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__)
-       // && !defined(CPU_CAPABILITY_SVE)
-
-#if defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
-    defined(CPU_CAPABILITY_SVE256)
-template <typename Op>
-struct VecReduceAllSIMD<float, Op> {
-  static inline float apply(
-      const Op& vec_fun,
-      const Vectorized<float>& acc_vec) {
-    using Vec = Vectorized<float>;
-    Vec v = acc_vec;
-    // 128-bit shuffle
-    svuint32_t ind = svdupq_n_u32(4, 5, 6, 7);
-    Vec v1 = svtbl_f32(v, ind);
-    v = vec_fun(v, v1);
-    // 64-bit shuffle
-    ind = svdupq_n_u32(2, 3, 0, 1);
-    v1 = svtbl_f32(v, ind);
-    v = vec_fun(v, v1);
-    // 32-bit shuffle
-    ind = svdupq_n_u32(1, 0, 2, 3);
-    v1 = svtbl_f32(v, ind);
-    v = vec_fun(v, v1);
-    return svlasta(svpfalse(), v);
-  }
-};
-#endif // defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__)
-       // && defined(CPU_CAPABILITY_SVE256)
 
 template <typename scalar_t, typename Op>
 inline scalar_t vec_reduce_all(

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

@@ -1,9 +1,21 @@
 #pragma once
 
 #include <ATen/cpu/vec/intrinsics.h>
+#include <c10/macros/Macros.h>
+#include <cstdint>
 
 #include <ATen/cpu/vec/vec_base.h>
 
+#if defined(__aarch64__) &&                     \
+    (defined(AT_BUILD_ARM_VEC256_WITH_SLEEF) || \
+     defined(AT_BUILD_ARM_VECSVE_WITH_SLEEF))
+#define SLEEF_STATIC_LIBS
+#include <sleef.h>
+#define USE_SLEEF(sleef_code, non_sleef_code) sleef_code
+#else
+#define USE_SLEEF(sleef_code, non_sleef_code) non_sleef_code
+#endif
+
 #if defined(CPU_CAPABILITY_SVE)
 
 // Define the data type of VLS(vector-length specific).

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

@@ -2,7 +2,6 @@
 
 #include <ATen/cpu/vec/intrinsics.h>
 #include <ATen/cpu/vec/sve/sve_helper.h>
-#include <ATen/cpu/vec/sve/vec_common_sve.h>
 #include <ATen/cpu/vec/sve/vec_float.h>
 #include <ATen/cpu/vec/vec_base.h>
 #include <c10/util/bit_cast.h>

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

@@ -1,6 +1,8 @@
 #pragma once
 
-#if defined(CPU_CAPABILITY_AVX512)
+#if defined(__aarch64__)
+#include <ATen/cpu/vec/vec_common_aarch64.h>
+#elif defined(CPU_CAPABILITY_AVX512)
 #include <ATen/cpu/vec/vec512/vec512.h>
 #else
 #include <ATen/cpu/vec/vec128/vec128.h>
@@ -11,6 +13,34 @@ namespace at::vec {
 // See Note [CPU_CAPABILITY namespace]
 inline namespace CPU_CAPABILITY {
 
+inline std::ostream& operator<<(std::ostream& stream, const c10::qint32& val) {
+  stream << val.val_;
+  return stream;
+}
+inline std::ostream& operator<<(std::ostream& stream, const c10::qint8& val) {
+  stream << static_cast<int>(val.val_);
+  return stream;
+}
+inline std::ostream& operator<<(std::ostream& stream, const c10::quint8& val) {
+  stream << static_cast<unsigned int>(val.val_);
+  return stream;
+}
+
+template <typename T>
+std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
+  T buf[Vectorized<T>::size()];
+  vec.store(buf);
+  stream << "vec[";
+  for (int i = 0; i != Vectorized<T>::size(); i++) {
+    if (i != 0) {
+      stream << ", ";
+    }
+    stream << buf[i];
+  }
+  stream << "]";
+  return stream;
+}
+
 inline Vectorized<bool> convert_to_bool(Vectorized<int8_t> x) {
   __at_align__ bool buffer[x.size()];
   x.ne(Vectorized<int8_t>(0)).store(buffer);

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

@@ -2,6 +2,7 @@
 
 // DO NOT DEFINE STATIC DATA IN THIS HEADER!
 // See Note [Do not compile initializers with AVX]
+#include <ATen/cpu/vec/sve/sve_helper.h>
 #include <ATen/cpu/vec/vec128/vec128_float_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_reduced_precision_common_neon.h>
 #include <ATen/cpu/vec/vec_base.h>
@@ -262,6 +263,13 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
             c10::bit_cast<at_bfloat16_t>(val6.x),
             c10::bit_cast<at_bfloat16_t>(val7.x)}) {}
 
+#ifdef CPU_CAPABILITY_SVE128
+  Vectorized(svbfloat16_t v) : Vectorized16(svget_neonq(v)) {}
+  operator svbfloat16_t() const {
+    return svset_neonq(svundef_bf16(), values);
+  }
+#endif
+
   static Vectorized<c10::BFloat16> blendv(
       const Vectorized<c10::BFloat16>& a,
       const Vectorized<c10::BFloat16>& b,
@@ -374,6 +382,23 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
   Vectorized ge(const Vectorized& other) const;
   Vectorized lt(const Vectorized& other) const;
   Vectorized le(const Vectorized& other) const;
+
+#ifdef CPU_CAPABILITY_SVE128
+
+  template <typename step_t>
+  static Vectorized<BFloat16> arange(
+      BFloat16 base = 0.f,
+      step_t step = static_cast<step_t>(1)) {
+    __at_align__ BFloat16 buffer[size()];
+    for (int64_t i = 0; i < size(); i++) {
+      buffer[i] = base + i * step;
+    }
+    return svget_neonq(
+        svld1_bf16(ptrue, reinterpret_cast<bfloat16_t*>(buffer)));
+  }
+
+#endif // CPU_CAPABILITY_SVE128
+
 }; // Vectorized<c10::BFloat16>
 
 inline std::tuple<Vectorized<float>, Vectorized<float>> convert_bfloat16_float(
@@ -397,6 +422,24 @@ inline Vectorized<c10::BFloat16> convert_float_bfloat16(
   return Vectorized<c10::BFloat16>(at_vcombine_bf16(x1, x2));
 }
 
+inline void load_fp32_from_bf16(const BFloat16* data, Vectorized<float>& out) {
+  __at_align__ float values[Vectorized<float>::size()];
+  for (const auto k : c10::irange(Vectorized<float>::size())) {
+    values[k] = data[k];
+  }
+  out = Vectorized<float>::loadu(values);
+}
+
+inline void load_fp32_from_bf16(
+    const BFloat16* data,
+    Vectorized<float>& out1,
+    Vectorized<float>& out2) {
+  Vectorized<BFloat16> bf16_vec = Vectorized<BFloat16>::loadu(data);
+  auto floats = convert_bfloat16_float(bf16_vec);
+  out1 = std::get<0>(floats);
+  out2 = std::get<1>(floats);
+}
+
 template <typename Op>
 Vectorized<c10::BFloat16> binary_operator_via_float(
     Op op,
@@ -579,6 +622,12 @@ Vectorized<c10::BFloat16> inline fnmsub(
   return -a * b - c;
 }
 
+#else //
+
+CONVERT_NON_VECTORIZED_INIT(BFloat16, bfloat16)
+
+LOAD_FP32_NON_VECTORIZED_INIT(BFloat16, bf16)
+
 #endif // !defined(C10_MOBILE) && defined(__aarch64__)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/vec128/vec128_convert.h

@@ -4,7 +4,7 @@
 
 namespace at::vec {
 inline namespace CPU_CAPABILITY {
-#if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
+#if defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256)
 template <typename src_t>
 struct VecConvert<
     float,
@@ -60,6 +60,7 @@ struct VecConvert<float, 1, BFloat16, 1> {
   }
 };
 
-#endif // defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256)
+#endif // defined(__aarch64__) && (!defined(CPU_CAPABILITY_SVE) ||
+       // defined(CPU_CAPABILITY_SVE128))
 } // namespace CPU_CAPABILITY
 } // namespace at::vec

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

@@ -4,13 +4,10 @@
 // See Note [Do not compile initializers with AVX]
 
 #include <ATen/cpu/vec/intrinsics.h>
+#include <ATen/cpu/vec/sve/sve_helper.h>
 #include <ATen/cpu/vec/vec_base.h>
 #include <c10/util/irange.h>
 
-#if defined(__aarch64__) && defined(AT_BUILD_ARM_VEC256_WITH_SLEEF)
-#include <sleef.h>
-#endif
-
 // Sleef offers vectorized versions of some transcedentals
 // such as sin, cos, tan etc..
 // However for now opting for STL, since we are not building
@@ -35,12 +32,6 @@ inline namespace CPU_CAPABILITY {
 #error "Big endian is not supported."
 #endif
 
-#if defined(AT_BUILD_ARM_VEC256_WITH_SLEEF)
-#define USE_SLEEF(sleef_code, non_sleef_code) sleef_code
-#else
-#define USE_SLEEF(sleef_code, non_sleef_code) non_sleef_code
-#endif
-
 template <int index, bool mask_val>
 struct BlendRegs {
   static float32x4_t impl(
@@ -94,6 +85,12 @@ class Vectorized<float> {
   operator float32x4_t() const {
     return values;
   }
+#ifdef CPU_CAPABILITY_SVE128
+  Vectorized(svfloat32_t v) : values(svget_neonq(v)) {}
+  operator svfloat32_t() const {
+    return svset_neonq(svundef_f32(), values);
+  }
+#endif
   template <int64_t mask>
   static Vectorized<float> blend(
       const Vectorized<float>& a,

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

@@ -4,7 +4,6 @@
 // See Note [Do not compile initializers with AVX]
 
 #include <ATen/cpu/vec/intrinsics.h>
-#include <ATen/cpu/vec/vec128/vec128_convert.h>
 #include <ATen/cpu/vec/vec128/vec128_float_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_reduced_precision_common_neon.h>
 #include <ATen/cpu/vec/vec_base.h>
@@ -25,7 +24,6 @@ inline namespace CPU_CAPABILITY {
 //    https://bugs.llvm.org/show_bug.cgi?id=45824
 // Most likely we will do aarch32 support with inline asm.
 #if !defined(C10_MOBILE) && defined(__aarch64__)
-
 #ifdef __BIG_ENDIAN__
 #error "Big endian is not supported."
 #endif
@@ -421,6 +419,24 @@ Vectorized<c10::Half> inline operator+(
 #endif
 }
 
+inline void load_fp32_from_fp16(const c10::Half* data, Vectorized<float>& out) {
+  __at_align__ float values[Vectorized<float>::size()];
+  for (const auto k : c10::irange(Vectorized<float>::size())) {
+    values[k] = data[k];
+  }
+  out = Vectorized<float>::loadu(values);
+}
+
+inline void load_fp32_from_fp16(
+    const c10::Half* data,
+    Vectorized<float>& out1,
+    Vectorized<float>& out2) {
+  Vectorized<c10::Half> f16_vec = Vectorized<c10::Half>::loadu(data);
+  auto floats = convert_half_float(f16_vec);
+  out1 = std::get<0>(floats);
+  out2 = std::get<1>(floats);
+}
+
 template <>
 Vectorized<c10::Half> inline operator-(
     const Vectorized<c10::Half>& a,
@@ -656,6 +672,53 @@ Vectorized<c10::Half> inline fnmsub(
   return -a * b - c;
 #endif
 }
+
+#else
+
+#define CONVERT_NON_VECTORIZED_INIT(type, name)                     \
+  inline std::tuple<Vectorized<float>, Vectorized<float>>           \
+      convert_##name##_float(const Vectorized<type>& a) {           \
+    constexpr int64_t K = Vectorized<type>::size();                 \
+    __at_align__ float arr[K];                                      \
+    __at_align__ type arr2[K];                                      \
+    a.store(arr2);                                                  \
+    convert(arr2, arr, K);                                          \
+    return std::make_tuple(                                         \
+        Vectorized<float>::loadu(arr),                              \
+        Vectorized<float>::loadu(arr + Vectorized<float>::size())); \
+  }                                                                 \
+  inline Vectorized<type> convert_float_##name(                     \
+      const Vectorized<float>& a, const Vectorized<float>& b) {     \
+    constexpr int64_t K = Vectorized<type>::size();                 \
+    __at_align__ float arr[K];                                      \
+    __at_align__ type arr2[K];                                      \
+    a.store(arr);                                                   \
+    b.store(arr + Vectorized<float>::size());                       \
+    convert(arr, arr2, K);                                          \
+    return Vectorized<type>::loadu(arr2);                           \
+  }
+
+#define LOAD_FP32_NON_VECTORIZED_INIT(type, name)                           \
+  inline void load_fp32_from_##name(                                        \
+      const type* data, Vectorized<float>& out) {                           \
+    __at_align__ float values[Vectorized<float>::size()];                   \
+    for (const auto k : c10::irange(Vectorized<float>::size())) {           \
+      values[k] = data[k];                                                  \
+    }                                                                       \
+    out = Vectorized<float>::loadu(values);                                 \
+  }                                                                         \
+                                                                            \
+  inline void load_fp32_from_##name(                                        \
+      const type* data, Vectorized<float>& out1, Vectorized<float>& out2) { \
+    load_fp32_from_##name(data, out1);                                      \
+    data += Vectorized<float>::size();                                      \
+    load_fp32_from_##name(data, out2);                                      \
+  }
+
+CONVERT_NON_VECTORIZED_INIT(Half, half)
+
+LOAD_FP32_NON_VECTORIZED_INIT(Half, fp16)
+
 #endif // !defined(C10_MOBILE) && defined(__aarch64__)
 
 } // namespace CPU_CAPABILITY

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

@@ -9,21 +9,16 @@
 #if !(                                                 \
     defined(__VSX__) || defined(CPU_CAPABILITY_VSX) || \
     defined(CPU_CAPABILITY_ZVECTOR))
-#if defined(CPU_CAPABILITY_SVE256)
-#include <ATen/cpu/vec/sve/vec_common_sve.h>
-#else
-// clang-format off
-#include <ATen/cpu/vec/vec256/vec256_float.h>
 #include <ATen/cpu/vec/vec256/vec256_double.h>
+#include <ATen/cpu/vec/vec256/vec256_float.h>
 #include <ATen/cpu/vec/vec256/vec256_int.h>
 #include <ATen/cpu/vec/vec256/vec256_qint.h>
-#endif
 #if !defined(CPU_CAPABILITY_SVE256) || !defined(__ARM_FEATURE_BF16)
 #include <ATen/cpu/vec/vec256/vec256_bfloat16.h>
 #endif
-#include <ATen/cpu/vec/vec256/vec256_half.h>
-#include <ATen/cpu/vec/vec256/vec256_complex_float.h>
 #include <ATen/cpu/vec/vec256/vec256_complex_double.h>
+#include <ATen/cpu/vec/vec256/vec256_complex_float.h>
+#include <ATen/cpu/vec/vec256/vec256_half.h>
 // clang-format on
 #elif defined(__VSX__) || defined(CPU_CAPABILITY_VSX)
 #include <ATen/cpu/vec/vec256/vsx/vec256_common_vsx.h>
@@ -56,34 +51,6 @@ namespace at::vec {
 // accessed as `at::vec`.
 inline namespace CPU_CAPABILITY {
 
-inline std::ostream& operator<<(std::ostream& stream, const c10::qint32& val) {
-  stream << val.val_;
-  return stream;
-}
-inline std::ostream& operator<<(std::ostream& stream, const c10::qint8& val) {
-  stream << static_cast<int>(val.val_);
-  return stream;
-}
-inline std::ostream& operator<<(std::ostream& stream, const c10::quint8& val) {
-  stream << static_cast<unsigned int>(val.val_);
-  return stream;
-}
-
-template <typename T>
-std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
-  T buf[Vectorized<T>::size()];
-  vec.store(buf);
-  stream << "vec[";
-  for (int i = 0; i != Vectorized<T>::size(); i++) {
-    if (i != 0) {
-      stream << ", ";
-    }
-    stream << buf[i];
-  }
-  stream << "]";
-  return stream;
-}
-
 #if defined(CPU_CAPABILITY_AVX2)
 
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ CAST (AVX2) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h

@@ -268,9 +268,7 @@ LOAD_FP32_VECTORIZED_INIT(BFloat16, bf16)
 
 #else // defined(CPU_CAPABILITY_AVX2)
 
-#if !(                                                                      \
-    defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
-    !defined(CPU_CAPABILITY_SVE256))
+#if !(defined(__aarch64__))
 CONVERT_NON_VECTORIZED_INIT(BFloat16, bfloat16)
 #endif
 

aten/src/ATen/cpu/vec/vec256/vec256_half.h

@@ -268,9 +268,7 @@ LOAD_FP32_VECTORIZED_INIT(Half, fp16)
 
 #else // defined(CPU_CAPABILITY_AVX2)
 
-#if !(                                                                      \
-    defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
-    !defined(CPU_CAPABILITY_SVE256))
+#if !defined(__aarch64__) || defined(CPU_CAPABILITY_SVE256)
 CONVERT_NON_VECTORIZED_INIT(Half, half)
 #endif
 

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

@@ -5,6 +5,13 @@
 
 #include <ATen/cpu/vec/intrinsics.h>
 #include <ATen/cpu/vec/vec_base.h>
+
+#ifdef __aarch64__
+#if defined(CPU_CAPABILITY_SVE128) || !defined(CPU_CAPABILITY_SVE)
+#include <ATen/cpu/vec/vec128/vec128_float_neon.h>
+#endif
+#endif
+
 #include <ATen/native/quantized/AffineQuantizerBase.h>
 
 #include <c10/util/irange.h>
@@ -915,7 +922,7 @@ Vectorized<c10::quint8> inline maximum(
   return a.maximum(b);
 }
 
-#elif !defined(CPU_CAPABILITY_SVE256)
+#else
 
 // NOTE: These are low-performance implementations that we fall back on
 // if we are not building with AVX2. This may not be an issue, because
@@ -1372,12 +1379,18 @@ Vectorized<c10::quint8> inline maximum(
   return a.maximum(b);
 }
 
-#endif // if defined(CPU_CAPABILITY_AVX2)
-
-#if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
+#if defined(__aarch64__) && \
+    (defined(CPU_CAPABILITY_SVE128) || !defined(CPU_CAPABILITY_SVE))
 std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
     at::vec::Vectorized<int8_t> src) {
+
+#ifdef CPU_CAPABILITY_SVE
+  svint8_t x = src;
+  auto s8x8 = vget_low_s8(svget_neonq(x));
+#else
   auto s8x8 = vld1_s8(src.operator const int8_t*());
+#endif
+
   auto s16x8 = vmovl_s8(s8x8);
 
   auto s32x4_hi = vmovl_s16(vget_high_s16(s16x8));
@@ -1402,7 +1415,14 @@ std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
 
 Vectorized<float> inline convert_int8_half_register_to_float(
     at::vec::Vectorized<int8_t> src) {
+
+#ifdef CPU_CAPABILITY_SVE
+  svint8_t x = src;
+  auto s8x8 = vget_low_s8(svget_neonq(x));
+#else
   auto s8x8 = vld1_s8(src.operator const int8_t*());
+#endif
+
   auto s16x8 = vmovl_s8(s8x8);
 
   auto s32x4_lo = vmovl_s16(vget_low_s16(s16x8));
@@ -1420,5 +1440,8 @@ Vectorized<float> inline convert_int8_half_register_to_float(
 }
 
 #endif
+
+#endif // if defined(CPU_CAPABILITY_AVX2)
+
 } // namespace CPU_CAPABILITY
 } // namespace at::vec

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

@@ -31,34 +31,6 @@ namespace vec {
 // See Note [CPU_CAPABILITY namespace]
 inline namespace CPU_CAPABILITY {
 
-inline std::ostream& operator<<(std::ostream& stream, const c10::qint32& val) {
-  stream << val.val_;
-  return stream;
-}
-inline std::ostream& operator<<(std::ostream& stream, const c10::qint8& val) {
-  stream << static_cast<int>(val.val_);
-  return stream;
-}
-inline std::ostream& operator<<(std::ostream& stream, const c10::quint8& val) {
-  stream << static_cast<unsigned int>(val.val_);
-  return stream;
-}
-
-template <typename T>
-std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
-  T buf[Vectorized<T>::size()];
-  vec.store(buf);
-  stream << "vec[";
-  for (int i = 0; i != Vectorized<T>::size(); i++) {
-    if (i != 0) {
-      stream << ", ";
-    }
-    stream << buf[i];
-  }
-  stream << "]";
-  return stream;
-}
-
 #if defined(CPU_CAPABILITY_AVX512)
 
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ CAST (AVX512)

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

@@ -67,18 +67,7 @@ Windows llvm will not have this definition.
 #endif
 #define VECTOR_WIDTH 64
 #define int_vector __m512i
-#elif defined(__aarch64__) && \
-    !defined(CPU_CAPABILITY_SVE) // CPU_CAPABILITY_AVX512
-// SVE code expects 256-vectors; leave that set for SVE?
-#if defined(__GNUC__)
-#define __at_align__ __attribute__((aligned(16)))
-#elif defined(_WIN32)
-#define __at_align__ __declspec(align(16))
-#else
-#define __at_align__
-#endif
-#define VECTOR_WIDTH 16
-#else // CPU_CAPABILITY_AVX512
+#elif defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_SVE256)
 #if defined(__GNUC__)
 #define __at_align__ __attribute__((aligned(32)))
 #elif defined(_WIN32)
@@ -88,7 +77,27 @@ Windows llvm will not have this definition.
 #endif
 #define VECTOR_WIDTH 32
 #define int_vector __m256i
-#endif // CPU_CAPABILITY_AVX512
+#elif defined(__aarch64__)
+// Define alignment and vector width for SVE128/Default (e.g., NEON)
+#if defined(__GNUC__)
+#define __at_align__ __attribute__((aligned(16)))
+#elif defined(_WIN32)
+#define __at_align__ __declspec(align(16))
+#else
+#define __at_align__
+#endif
+#define VECTOR_WIDTH 16
+#else
+// Fallback: define default alignment and vector width
+#if defined(__GNUC__)
+#define __at_align__ __attribute__((aligned(32)))
+#elif defined(_WIN32)
+#define __at_align__ __declspec(align(32))
+#else
+#define __at_align__
+#endif
+#define VECTOR_WIDTH 32
+#endif
 
 namespace at::vec {
 // See Note [CPU_CAPABILITY namespace]

aten/src/ATen/cpu/vec/vec_common_aarch64.h

@@ -8,13 +8,48 @@
 #include <ATen/cpu/vec/sve/sve_helper.h>
 #include <ATen/cpu/vec/vec_base.h>
 
-#if defined(CPU_CAPABILITY_SVE)
-#include <ATen/cpu/vec/sve/vec_bfloat16.h>
-#include <ATen/cpu/vec/sve/vec_double.h>
-#include <ATen/cpu/vec/sve/vec_float.h>
-#include <ATen/cpu/vec/sve/vec_int.h>
+#ifdef CPU_CAPABILITY_SVE128
+
+#include <ATen/cpu/vec/vec128/vec128_float_neon.h>
+
+#include <ATen/cpu/vec/vec128/vec128_bfloat16_neon.h>
+
+#include <ATen/cpu/vec/vec128/vec128_half_neon.h>
+
+#include <ATen/cpu/vec/vec128/vec128_convert.h>
+
 #include <ATen/cpu/vec/sve/vec_qint.h>
-#endif
+
+#elif defined(CPU_CAPABILITY_SVE)
+
+#include <ATen/cpu/vec/sve/vec_float.h>
+
+#include <ATen/cpu/vec/sve/vec_bfloat16.h>
+
+#include <ATen/cpu/vec/sve/vec_double.h>
+#include <ATen/cpu/vec/sve/vec_int.h>
+
+#include <ATen/cpu/vec/sve/vec_qint.h>
+
+#include <ATen/cpu/vec/vec256/vec256_half.h>
+
+#include <ATen/cpu/vec/vec256/vec256_convert.h>
+
+#else // NEON
+
+#include <ATen/cpu/vec/vec128/vec128_float_neon.h>
+
+#include <ATen/cpu/vec/vec128/vec128_half_neon.h>
+
+#include <ATen/cpu/vec/vec128/vec128_bfloat16_neon.h>
+
+#include <ATen/cpu/vec/vec128/vec128_convert.h>
+
+#include <ATen/cpu/vec/vec256/vec256_qint.h>
+
+#endif // defined(CPU_CAPABILITY_SVE128)
+
+#include <ATen/cpu/vec/functional.h>
 
 namespace at::vec {
 // Note [CPU_CAPABILITY namespace]
@@ -48,12 +83,6 @@ DEFINE_SVE_CAST(int32_t, s32, float, f32)
 DEFINE_SVE_CAST(int16_t, s16, float, f32)
 DEFINE_SVE_CAST(float, f32, double, f64)
 
-#ifdef __ARM_FEATURE_BF16
-DEFINE_SVE_CAST(int64_t, s64, c10::BFloat16, bf16)
-DEFINE_SVE_CAST(int32_t, s32, c10::BFloat16, bf16)
-DEFINE_SVE_CAST(int16_t, s16, c10::BFloat16, bf16)
-#endif // __ARM_FEATURE_BF16
-
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ GATHER ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 template <int64_t scale = 1>
@@ -173,9 +202,11 @@ std::pair<
   // group cols crossing lanes:
   //   return {a0, b0, a1, b1, a2, b2, a3, b3}
   //          {a4, b4, a5, b5, a6, b6, a7, b7}
-  return std::make_pair(
-      Vectorized<c10::BFloat16>(svzip1_bf16(a, b)),
-      Vectorized<c10::BFloat16>(svzip2_bf16(a, b)));
+  svbfloat16_t aReg = a;
+  svbfloat16_t bReg = b;
+  Vectorized<c10::BFloat16> c = svzip1_bf16(aReg, bReg);
+  Vectorized<c10::BFloat16> d = svzip2_bf16(aReg, bReg);
+  return std::make_pair(c, d);
 }
 #endif // __ARM_FEATURE_BF16
 
@@ -224,12 +255,27 @@ std::pair<
   // swap lanes:
   //   return {a0, a1, a2, a3, a4, a5, a6, a7}
   //          {b0, b1, b2, b3, b4, b5, b6, b7}
-  return std::make_pair(
-      Vectorized<c10::BFloat16>(svuzp1_bf16((svbfloat16_t)a, (svbfloat16_t)b)),
-      Vectorized<c10::BFloat16>(svuzp2_bf16((svbfloat16_t)a, (svbfloat16_t)b)));
+  svbfloat16_t aReg = a;
+  svbfloat16_t bReg = b;
+  Vectorized<c10::BFloat16> c = svuzp1_bf16(aReg, bReg);
+  Vectorized<c10::BFloat16> d = svuzp2_bf16(aReg, bReg);
+  return std::make_pair(c, d);
 }
 #endif // __ARM_FEATURE_BF16
 
+// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ FLIP ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+#define DEFINE_FLIP_FUNC(type, sve_func)                    \
+  inline Vectorized<type> flip(const Vectorized<type>& v) { \
+    return Vectorized<type>(sve_func(v));                   \
+  }
+// Use the macro to define the flip functions
+DEFINE_FLIP_FUNC(float, svrev_f32)
+DEFINE_FLIP_FUNC(double, svrev_f64)
+DEFINE_FLIP_FUNC(int64_t, svrev_s64)
+DEFINE_FLIP_FUNC(int32_t, svrev_s32)
+DEFINE_FLIP_FUNC(int16_t, svrev_s16)
+DEFINE_FLIP_FUNC(int8_t, svrev_s8)
+
 #endif // defined(CPU_CAPABILITY_SVE)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cuda/CUDABlas.cpp

@@ -108,7 +108,7 @@ static hipblasStatus_t rocBLASStatusToHIPStatus(rocblas_status error)
 
 namespace {
 
-cublasOperation_t _cublasOpFromChar(char op) {
+static cublasOperation_t _cublasOpFromChar(char op) {
   // NOLINTNEXTLINE(bugprone-switch-missing-default-case)
   switch (op) {
     case 'n':
@@ -128,7 +128,7 @@ cublasOperation_t _cublasOpFromChar(char op) {
       "_cublasOpFromChar input should be 't', 'n' or 'c' but got `", op, "`");
 }
 
-void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
+static 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 @@ void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
     *lda = std::max<int64_t>(m, 1);
 }
 
-void _cublasAdjustLdLevel3(
+static 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.
-int64_t num_gpus;
+static int64_t num_gpus;
 
 // Ensures default_gens_cuda is initialized once.
-std::deque<c10::once_flag> cuda_gens_init_flag;
+static std::deque<c10::once_flag> cuda_gens_init_flag;
 
 // Default, global CUDA generators, one per GPU.
-std::vector<Generator> default_gens_cuda;
+static std::vector<Generator> default_gens_cuda;
 
 /*
  * Populates the global variables related to CUDA generators
  * Warning: this function must only be called once!
  */
-void initCUDAGenVector() {
+static 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/cudnn/Types.cpp

@@ -2,8 +2,6 @@
 
 #include <ATen/ATen.h>
 
-#include <c10/util/Exception.h>
-
 namespace at::native {
 
 cudnnDataType_t getCudnnDataTypeFromScalarType(const at::ScalarType dtype) {
@@ -22,10 +20,9 @@ cudnnDataType_t getCudnnDataTypeFromScalarType(const at::ScalarType dtype) {
   } else if (dtype == at::kByte) {
     return CUDNN_DATA_UINT8;
   }
-  TORCH_CHECK(false,
-    "getCudnnDataTypeFromScalarType() not supported for ",
-    toString(dtype)
-  );
+  std::string msg("getCudnnDataTypeFromScalarType() not supported for ");
+  msg += toString(dtype);
+  throw std::runtime_error(msg);
 }
 
 cudnnDataType_t getCudnnDataType(const at::Tensor& tensor) {

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
-std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
+static 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 @@ std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
   TORCH_INTERNAL_ASSERT(false, "can't get here");
 }
 
-std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
+static 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 @@ std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
   return tv_batch_rule(self, self_bdim, other, other_bdim);
 }
 
-std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
+static 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 @@ std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
   return std::make_tuple( at::matmul(self_, other_), 0 );
 }
 
-std::tuple<Tensor, std::optional<int64_t>> bmm_batch_rule(
+static 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...>>
   }
 };
 
-void expect_at_least_rank(
+static 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);
 }
 
-std::tuple<Tensor, std::optional<int64_t>>
+static 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/BatchRulesModules.cpp

@@ -213,22 +213,40 @@ static cudnn_grid_sample_backward_batch_rule(
   return grid_sample_backward_helper_out(std::move(bw_out), 0, 0, bdim_size);
 }
 
-// uses functional formulation for one_hot under vmap to be compatible with
-// fakeTensor/dynamic shapes and compiled functorch transforms.
-// mirrors the meta path in aten/src/ATen/native/Onehot.cpp,
-// but requires explicit positive num_classes under vmap to avoid
-// data-dependent output shapes.
+// TODO: replace with targetable functionalization
 static Tensor one_hot_decomposition_hack(const Tensor &self, int64_t num_classes) {
     TORCH_CHECK(self.dtype() == kLong, "one_hot is only applicable to index tensor.");
+    auto shape = self.sym_sizes().vec();
+
+    // empty tensor could be converted to one hot representation,
+    // but shape inference is not possible.
+    if (self.sym_numel() == 0) {
+        if (num_classes <= 0) {
+            TORCH_CHECK(false, "Can not infer total number of classes from empty tensor.");
+        } else {
+            shape.emplace_back(num_classes);
+            return at::empty_symint(shape, self.options());
+        }
+    }
 
-    // disallow implicit inference under vmap; this would be data-dependent
-    // and is intentionally guarded by Dynamo in torch/_dynamo/variables/torch.py.
     TORCH_CHECK(num_classes > 0, "When vmap-ing torch.nn.functional.one_hot, please "
         "provide an explicit positive num_classes argument.");
 
-    const auto options = self.options();
-    at::Tensor index = at::arange(num_classes, options);
-    return at::eq(self.unsqueeze(-1), index).to(at::kLong);
+    // Disabling all of the following checks. This is OK because scatter has checks too.
+    // Maybe one_hot should be a primitive wrt autograd so we don't have to deal with this.
+    // // non-empty tensor
+    // if (self.device().type() != at::kCUDA) {
+    //   //for cuda, rely on device assert thrown by scatter
+    //   TORCH_CHECK(self.min().item().toLong() >= 0, "Class values must be non-negative.");
+    // }
+    // if (self.device().type() != at::kCUDA) {
+    //   //rely on device asserts from scatter to avoid sync here
+    //   TORCH_CHECK(num_classes > self.max().item().toLong(), "Class values must be smaller than num_classes.");
+    // }
+
+    shape.emplace_back(num_classes);
+    Tensor ret = at::zeros_symint(shape, self.options());
+    return ret.scatter(-1, self.unsqueeze(-1), 1);
 }
 
 template <typename A, A a, typename C>

aten/src/ATen/functorch/BatchRulesScatterOps.cpp

@@ -12,14 +12,13 @@
 #include <ATen/native/IndexKernel.h>
 #include <ATen/native/IndexingUtils.h>
 #include <torch/library.h>
-#include <c10/util/Exception.h>
 
 
 // NOLINTBEGIN(bugprone-unchecked-optional-access)
 namespace at::functorch {
 
 namespace {
-bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
+static bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
   for (const auto& bdim : bdims) {
     if (bdim.has_value()) {
       return true;
@@ -28,7 +27,7 @@ bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
   return false;
 }
 
-int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
+static 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 +39,7 @@ int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
   return result;
 }
 
-int64_t get_max_index_logical_dim(
+static 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 +56,7 @@ int64_t get_max_index_logical_dim(
   return max_logical_dim;
 }
 
-std::vector<std::optional<Tensor>> batchIndices(
+static std::vector<std::optional<Tensor>> batchIndices(
   at::TensorOptions options,
   ArrayRef<std::optional<Tensor>> indices,
   ArrayRef<std::optional<int64_t>> indices_bdims,
@@ -95,10 +94,9 @@ std::vector<std::optional<Tensor>> batchIndices(
     if (index.has_value() && index->sym_numel() != 0) {
       const auto idx_bdim = indices_bdims[i];
       indices_.emplace_back(maybePadToLogicalRank(moveBatchDimToFront(index.value(), idx_bdim), idx_bdim, maxLogicalRank));
-      TORCH_CHECK(
-        !(index.value().dtype() == kBool) || !indices_bdims[i].has_value(),
-        "vmap: We do not support batching operators that can support dynamic shape. Attempting to batch over indexing with a boolean mask."
-      );
+      if (index.value().dtype() == kBool && indices_bdims[i].has_value()) {
+        throw std::runtime_error("vmap: We do not support batching operators that can support dynamic shape. Attempting to batch over indexing with a boolean mask.");
+      }
     } else {
       indices_.push_back(index);
     }
@@ -126,7 +124,7 @@ 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 :).
-bool is_advanced_index(const std::optional<Tensor>& idx) {
+static bool is_advanced_index(const std::optional<Tensor>& idx) {
   if (!idx.has_value()) {
     return false;
   }
@@ -137,7 +135,7 @@ bool is_advanced_index(const std::optional<Tensor>& idx) {
 }
 
 // See NOTE: [advanced indices adjacent] for definition
-bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indices) {
+static 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 +163,7 @@ bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indices) {
 // - result: Tensor[B, 4, 5, 6, 2, 3, 7, 8]
 //                     -------  ----
 //                     region2  region1
-Tensor swap_regions(const Tensor& tensor, int64_t first_region_size, int64_t second_region_size) {
+static 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 +551,7 @@ Tensor &_index_put_impl__plumbing(Tensor &self, const List<std::optional<Tensor>
   return self;
 }
 
-Tensor maybe_permute_values(
+static Tensor maybe_permute_values(
     const Tensor& values,
     ArrayRef<std::optional<Tensor>> orig_indices,
     ArrayRef<std::optional<int64_t>> orig_indices_bdims) {
@@ -1052,7 +1050,7 @@ std::tuple<Tensor, std::optional<int64_t>> index_add_batch_rule(
                                    other, other_bdim, alpha, false);
 }
 
-std::tuple<Tensor,Tensor> binary_pointwise_align(
+static 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);
 }
 
-bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
+static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
   return dim == 0 || dim == -1;
 }
 

aten/src/ATen/functorch/Interpreter.h

@@ -3,7 +3,6 @@
 #include <ATen/functorch/Macros.h>
 #include <ATen/core/dispatch/Dispatcher.h>
 #include <c10/core/impl/LocalDispatchKeySet.h>
-#include <c10/util/Exception.h>
 #include <optional>
 #include <bitset>
 #include <utility>
@@ -107,10 +106,9 @@ struct VmapInterpreterMeta {
 
   template <typename T>
   friend void to_json(T& json_j, const VmapInterpreterMeta& json_t) {
-    TORCH_CHECK(
-      !json_t.batchSize_.is_heap_allocated(),
-      "Serialization for heap-allocated SymInt is not implemented yet"
-    );
+    if (json_t.batchSize_.is_heap_allocated()) {
+      throw std::runtime_error("Serialization for heap-allocated SymInt is not implemented yet");
+    }
     json_j["batchSize"] = json_t.batchSize_.as_int_unchecked();
     json_j["randomness"] = static_cast<int64_t>(json_t.randomness_);
   }
@@ -304,7 +302,7 @@ struct Interpreter {
     } else if (meta.contains("Functionalize")) {
       json_t.meta_.emplace<FunctionalizeInterpreterMeta>(meta["Functionalize"].template get<FunctionalizeInterpreterMeta>());
     } else {
-      TORCH_CHECK(false, "unknown interpreter metadata type");
+      throw std::runtime_error("unknown interpreter metadata type");
     }
   }
 

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.
-bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
+static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
   return dim == 0 || dim == -1;
 }
 
-int64_t get_current_level() {
+static 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...
-bool participatesInCurrentLevel(const Tensor& self) {
+static 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 @@ bool participatesInCurrentLevel(const Tensor& self) {
   return self_level == current_level;
 }
 
-bool participatesInCurrentLevel(ITensorListRef self) {
+static 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).
-std::optional<c10::SymInt> maximum_indexable_location(
+static 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 @@ 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.
-void checkBasicAsStridedValidForSlice(
+static void checkBasicAsStridedValidForSlice(
     const Tensor& physical_tensor,
     int64_t num_batch_dims,
     c10::SymIntArrayRef sizes,

aten/src/ATen/functorch/PyTorchOperatorHacks.cpp

@@ -6,7 +6,6 @@
 #include <ATen/functorch/BatchedTensorImpl.h>
 #include <ATen/Dispatch.h>
 #include <c10/util/irange.h>
-#include <c10/util/Exception.h>
 #include <ATen/NamedTensorUtils.h>
 #include <ATen/native/LinearAlgebraUtils.h>
 #include <ATen/native/xnnpack/Engine.h>
@@ -71,7 +70,7 @@ Tensor linear_hack(const Tensor& input, const Tensor& weight, const std::optiona
   return output;
 }
 
-inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
+static 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) {
@@ -109,7 +108,9 @@ Tensor binary_cross_entropy_with_logits_hack(
 }
 
 Tensor trace_backward_decomp(const Tensor& grad, IntArrayRef sizes) {
-  TORCH_CHECK(sizes.size() == 2, "expected matrix input");
+  if (sizes.size() != 2) {
+    throw std::runtime_error("expected matrix input");
+  }
   auto grad_input = at::zeros(sizes[0] * sizes[1], grad.options());
   auto indices = at::arange(0, grad_input.numel(), sizes[1] + 1, grad.options().dtype(at::kLong));
   // Workaround using index_put instead of yet unsupported index_fill_
@@ -127,7 +128,7 @@ namespace {
 template<bool inplace>
 using Ctype = std::conditional_t<inplace, Tensor&, Tensor>;
 
-Tensor make_feature_noise(const Tensor& input) {
+static 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 +142,7 @@ Tensor make_feature_noise(const Tensor& input) {
   return at::empty(sizes, input.options());
 }
 
-bool is_fused_kernel_acceptable(const Tensor& input, double p) {
+static 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 +211,7 @@ ALIAS_SPECIALIZATION(_feature_dropout,       true,  false)
 ALIAS_SPECIALIZATION(_alpha_dropout,         false, true )
 ALIAS_SPECIALIZATION(_feature_alpha_dropout, true,  true )
 
-Tensor dropout(const Tensor& input, double p, bool train) {
+static 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>
-void adaptive_avg_pool3d_out_frame(
+static 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>
-void adaptive_avg_pool3d_backward_out_frame(
+static 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>
-void adaptive_max_pool3d_single_out_frame(
+static void adaptive_max_pool3d_single_out_frame(
           const scalar_t *input_p,
           scalar_t *output_p,
           int64_t *ind_p,
@@ -170,7 +170,7 @@ void adaptive_max_pool3d_single_out_frame(
 }
 
 template <typename scalar_t>
-void adaptive_max_pool3d_out_frame(
+static void adaptive_max_pool3d_out_frame(
           const scalar_t *input_data,
           scalar_t *output_data,
           int64_t *indices_data,
@@ -202,7 +202,7 @@ void adaptive_max_pool3d_out_frame(
 }
 
 template <typename scalar_t>
-void adaptive_max_pool3d_backward_single_out_frame(
+static 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 @@ void adaptive_max_pool3d_backward_single_out_frame(
 }
 
 template <typename scalar_t>
-void adaptive_max_pool3d_backward_out_frame(
+static 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>
-void avg_pool3d_out_frame(
+static 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>
-void avg_pool3d_backward_out_frame(
+static 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>
-inline
+static 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>
-inline
+static 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>
-void apply_geqrf(const Tensor& input, const Tensor& tau) {
+static 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>
-void apply_svd(const Tensor& A,
+static void apply_svd(const Tensor& A,
                       const bool full_matrices,
                       const bool compute_uv,
                       const Tensor& U,
@@ -1157,103 +1157,103 @@ REGISTER_AVX512_DISPATCH(cholesky_stub, &cholesky_kernel)
 REGISTER_AVX2_DISPATCH(cholesky_stub, &cholesky_kernel)
 REGISTER_VSX_DISPATCH(cholesky_stub, &cholesky_kernel)
 REGISTER_ZVECTOR_DISPATCH(cholesky_stub, &cholesky_kernel)
-REGISTER_SVE256_DISPATCH(cholesky_stub, &cholesky_kernel)
+REGISTER_SVE_DISPATCH(cholesky_stub, &cholesky_kernel)
 
 REGISTER_ARCH_DISPATCH(cholesky_inverse_stub, DEFAULT, &cholesky_inverse_kernel_impl)
 REGISTER_AVX512_DISPATCH(cholesky_inverse_stub, &cholesky_inverse_kernel_impl)
 REGISTER_AVX2_DISPATCH(cholesky_inverse_stub, &cholesky_inverse_kernel_impl)
 REGISTER_VSX_DISPATCH(cholesky_inverse_stub, &cholesky_inverse_kernel_impl)
 REGISTER_ZVECTOR_DISPATCH(cholesky_inverse_stub, &cholesky_inverse_kernel_impl)
-REGISTER_SVE256_DISPATCH(cholesky_inverse_stub, &cholesky_inverse_kernel_impl)
+REGISTER_SVE_DISPATCH(cholesky_inverse_stub, &cholesky_inverse_kernel_impl)
 
 REGISTER_ARCH_DISPATCH(linalg_eig_stub, DEFAULT, &linalg_eig_kernel)
 REGISTER_AVX512_DISPATCH(linalg_eig_stub, &linalg_eig_kernel)
 REGISTER_AVX2_DISPATCH(linalg_eig_stub, &linalg_eig_kernel)
 REGISTER_VSX_DISPATCH(linalg_eig_stub, &linalg_eig_kernel)
 REGISTER_ZVECTOR_DISPATCH(linalg_eig_stub, &linalg_eig_kernel)
-REGISTER_SVE256_DISPATCH(linalg_eig_stub, &linalg_eig_kernel)
+REGISTER_SVE_DISPATCH(linalg_eig_stub, &linalg_eig_kernel)
 
 REGISTER_ARCH_DISPATCH(linalg_eigh_stub, DEFAULT, &linalg_eigh_kernel)
 REGISTER_AVX512_DISPATCH(linalg_eigh_stub, &linalg_eigh_kernel)
 REGISTER_AVX2_DISPATCH(linalg_eigh_stub, &linalg_eigh_kernel)
 REGISTER_VSX_DISPATCH(linalg_eigh_stub, &linalg_eigh_kernel)
 REGISTER_ZVECTOR_DISPATCH(linalg_eigh_stub, &linalg_eigh_kernel)
-REGISTER_SVE256_DISPATCH(linalg_eigh_stub, &linalg_eigh_kernel)
+REGISTER_SVE_DISPATCH(linalg_eigh_stub, &linalg_eigh_kernel)
 
 REGISTER_ARCH_DISPATCH(geqrf_stub, DEFAULT, &geqrf_kernel)
 REGISTER_AVX512_DISPATCH(geqrf_stub, &geqrf_kernel)
 REGISTER_AVX2_DISPATCH(geqrf_stub, &geqrf_kernel)
 REGISTER_VSX_DISPATCH(geqrf_stub, &geqrf_kernel)
 REGISTER_ZVECTOR_DISPATCH(geqrf_stub, &geqrf_kernel)
-REGISTER_SVE256_DISPATCH(geqrf_stub, &geqrf_kernel)
+REGISTER_SVE_DISPATCH(geqrf_stub, &geqrf_kernel)
 
 REGISTER_ARCH_DISPATCH(orgqr_stub, DEFAULT, &orgqr_kernel_impl)
 REGISTER_AVX512_DISPATCH(orgqr_stub, &orgqr_kernel_impl)
 REGISTER_AVX2_DISPATCH(orgqr_stub, &orgqr_kernel_impl)
 REGISTER_VSX_DISPATCH(orgqr_stub, &orgqr_kernel_impl)
 REGISTER_ZVECTOR_DISPATCH(orgqr_stub, &orgqr_kernel_impl)
-REGISTER_SVE256_DISPATCH(orgqr_stub, &orgqr_kernel_impl)
+REGISTER_SVE_DISPATCH(orgqr_stub, &orgqr_kernel_impl)
 
 REGISTER_ARCH_DISPATCH(ormqr_stub, DEFAULT, &ormqr_kernel)
 REGISTER_AVX512_DISPATCH(ormqr_stub, &ormqr_kernel)
 REGISTER_AVX2_DISPATCH(ormqr_stub, &ormqr_kernel)
 REGISTER_VSX_DISPATCH(ormqr_stub, &ormqr_kernel)
 REGISTER_ZVECTOR_DISPATCH(ormqr_stub, &ormqr_kernel)
-REGISTER_SVE256_DISPATCH(ormqr_stub, &ormqr_kernel)
+REGISTER_SVE_DISPATCH(ormqr_stub, &ormqr_kernel)
 
 REGISTER_ARCH_DISPATCH(lstsq_stub, DEFAULT, &lstsq_kernel)
 REGISTER_AVX512_DISPATCH(lstsq_stub, &lstsq_kernel)
 REGISTER_AVX2_DISPATCH(lstsq_stub, &lstsq_kernel)
 REGISTER_VSX_DISPATCH(lstsq_stub, &lstsq_kernel)
 REGISTER_ZVECTOR_DISPATCH(lstsq_stub, &lstsq_kernel)
-REGISTER_SVE256_DISPATCH(lstsq_stub, &lstsq_kernel)
+REGISTER_SVE_DISPATCH(lstsq_stub, &lstsq_kernel)
 
 REGISTER_ARCH_DISPATCH(triangular_solve_stub, DEFAULT, &triangular_solve_kernel)
 REGISTER_AVX512_DISPATCH(triangular_solve_stub, &triangular_solve_kernel)
 REGISTER_AVX2_DISPATCH(triangular_solve_stub, &triangular_solve_kernel)
 REGISTER_VSX_DISPATCH(triangular_solve_stub, &triangular_solve_kernel)
 REGISTER_ZVECTOR_DISPATCH(triangular_solve_stub, &triangular_solve_kernel)
-REGISTER_SVE256_DISPATCH(triangular_solve_stub, &triangular_solve_kernel)
+REGISTER_SVE_DISPATCH(triangular_solve_stub, &triangular_solve_kernel)
 
 REGISTER_ARCH_DISPATCH(lu_factor_stub, DEFAULT, &lu_factor_kernel)
 REGISTER_AVX512_DISPATCH(lu_factor_stub, &lu_factor_kernel)
 REGISTER_AVX2_DISPATCH(lu_factor_stub, &lu_factor_kernel)
 REGISTER_VSX_DISPATCH(lu_factor_stub, &lu_factor_kernel)
 REGISTER_ZVECTOR_DISPATCH(lu_factor_stub, &lu_factor_kernel)
-REGISTER_SVE256_DISPATCH(lu_factor_stub, &lu_factor_kernel)
+REGISTER_SVE_DISPATCH(lu_factor_stub, &lu_factor_kernel)
 
 REGISTER_ARCH_DISPATCH(ldl_factor_stub, DEFAULT, &ldl_factor_kernel)
 REGISTER_AVX512_DISPATCH(ldl_factor_stub, &ldl_factor_kernel)
 REGISTER_AVX2_DISPATCH(ldl_factor_stub, &ldl_factor_kernel)
 REGISTER_VSX_DISPATCH(ldl_factor_stub, &ldl_factor_kernel)
 REGISTER_ZVECTOR_DISPATCH(ldl_factor_stub, &ldl_factor_kernel)
-REGISTER_SVE256_DISPATCH(ldl_factor_stub, &ldl_factor_kernel)
+REGISTER_SVE_DISPATCH(ldl_factor_stub, &ldl_factor_kernel)
 
 REGISTER_ARCH_DISPATCH(ldl_solve_stub, DEFAULT, &ldl_solve_kernel)
 REGISTER_AVX512_DISPATCH(ldl_solve_stub, &ldl_solve_kernel)
 REGISTER_AVX2_DISPATCH(ldl_solve_stub, &ldl_solve_kernel)
 REGISTER_VSX_DISPATCH(ldl_solve_stub, &ldl_solve_kernel)
 REGISTER_ZVECTOR_DISPATCH(ldl_solve_stub, &ldl_solve_kernel)
-REGISTER_SVE256_DISPATCH(ldl_solve_stub, &ldl_solve_kernel)
+REGISTER_SVE_DISPATCH(ldl_solve_stub, &ldl_solve_kernel)
 
 REGISTER_ARCH_DISPATCH(lu_solve_stub, DEFAULT, &lu_solve_kernel)
 REGISTER_AVX512_DISPATCH(lu_solve_stub, &lu_solve_kernel)
 REGISTER_AVX2_DISPATCH(lu_solve_stub, &lu_solve_kernel)
 REGISTER_VSX_DISPATCH(lu_solve_stub, &lu_solve_kernel)
 REGISTER_ZVECTOR_DISPATCH(lu_solve_stub, &lu_solve_kernel)
-REGISTER_SVE256_DISPATCH(lu_solve_stub, &lu_solve_kernel)
+REGISTER_SVE_DISPATCH(lu_solve_stub, &lu_solve_kernel)
 
 REGISTER_ARCH_DISPATCH(svd_stub, DEFAULT, &svd_kernel)
 REGISTER_AVX512_DISPATCH(svd_stub, &svd_kernel)
 REGISTER_AVX2_DISPATCH(svd_stub, &svd_kernel)
 REGISTER_VSX_DISPATCH(svd_stub, &svd_kernel)
 REGISTER_ZVECTOR_DISPATCH(svd_stub, &svd_kernel)
-REGISTER_SVE256_DISPATCH(svd_stub, &svd_kernel)
+REGISTER_SVE_DISPATCH(svd_stub, &svd_kernel)
 
 REGISTER_ARCH_DISPATCH(unpack_pivots_stub, DEFAULT, &unpack_pivots_cpu_kernel)
 REGISTER_AVX512_DISPATCH(unpack_pivots_stub, &unpack_pivots_cpu_kernel)
 REGISTER_AVX2_DISPATCH(unpack_pivots_stub, &unpack_pivots_cpu_kernel)
 REGISTER_VSX_DISPATCH(unpack_pivots_stub, &unpack_pivots_cpu_kernel)
 REGISTER_ZVECTOR_DISPATCH(unpack_pivots_stub, &unpack_pivots_cpu_kernel)
-REGISTER_SVE256_DISPATCH(unpack_pivots_stub, &unpack_pivots_cpu_kernel)
+REGISTER_SVE_DISPATCH(unpack_pivots_stub, &unpack_pivots_cpu_kernel)
 } // namespace at::native

aten/src/ATen/native/Col2Im.cpp

@@ -71,7 +71,7 @@
 namespace at::native {
 namespace {
 
-void col2im_out_cpu_template(
+static 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 {
 
-Tensor compute_columns2d(
+static Tensor compute_columns2d(
     const Tensor& input,
     IntArrayRef padding,
     IntArrayRef stride,
@@ -93,7 +93,7 @@ Tensor compute_columns2d(
   return columns.contiguous();
 }
 
-inline void slow_conv2d_shape_check(
+static inline void slow_conv2d_shape_check(
     const Tensor& input,
     const Tensor& grad_output,
     const Tensor& weight,
@@ -205,7 +205,7 @@ inline void slow_conv2d_shape_check(
   }
 }
 
-inline Tensor view_weight_2d(const Tensor& weight_,
+static 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 @@ inline Tensor view_weight_2d(const Tensor& weight_,
 }
 
 template <typename scalar_t>
-void slow_conv2d_update_output_frame(
+static 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(
   }
 }
 
-void slow_conv2d_backward_weight_out_cpu_template(
+static 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 {
 
-Tensor compute_columns3d(
+static Tensor compute_columns3d(
     const Tensor& input_,
     IntArrayRef stride,
     IntArrayRef padding,
@@ -108,7 +108,7 @@ Tensor compute_columns3d(
   return columns;
 }
 
-inline void slow_conv3d_shape_check(
+static inline void slow_conv3d_shape_check(
     const Tensor& input,
     const Tensor& grad_output,
     const Tensor& weight,
@@ -273,7 +273,7 @@ inline void slow_conv3d_shape_check(
   }
 }
 
-Tensor view_weight_2d(const Tensor& weight_) {
+static 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 @@ Tensor view_weight_2d(const Tensor& weight_) {
 }
 
 template <typename scalar_t>
-void slow_conv3d_update_output_frame(
+static 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);
 }
 
-void slow_conv3d_backward_parameters_out_cpu_template(
+static void slow_conv3d_backward_parameters_out_cpu_template(
     Tensor& grad_weight,
     const Tensor& input,
     const Tensor& grad_output,

aten/src/ATen/native/DispatchStub.cpp

@@ -39,19 +39,21 @@ static CPUCapability compute_cpu_capability() {
     }
 #elif defined(HAVE_SVE_CPU_DEFINITION)
     int sve_vl = cpuinfo_get_max_arm_sve_length(); //Returns maximum SVE VL supported by your HW.
-#ifdef HAVE_SVE256_CPU_DEFINITION
-    if (envar == "sve256") {
+    if (envar == "sve") {
+      // Select SVE capability based on the maximum SVE VL supported by the HW.
       if (sve_vl == 256) {
-#ifdef HAVE_ARM_BF16_CPU_DEFINITION
         if (cpuinfo_has_arm_bf16()) {
           return CPUCapability::SVE256;
         }
-#endif
+      } else if (sve_vl == 128) {
+        if (cpuinfo_has_arm_bf16()) {
+          return CPUCapability::SVE128;
+        }
+      } else {
+        TORCH_WARN("SVE capability not available on hardware. Falling back to DEFAULT");
+        return CPUCapability::DEFAULT;
       }
-      TORCH_WARN("SVE256 capability not available on hardware. Falling back to DEFAULT");
-      return CPUCapability::DEFAULT;
     }
-#endif
 #else
 #ifdef HAVE_AVX512_CPU_DEFINITION
     if (envar == "avx512") {
@@ -113,6 +115,11 @@ static CPUCapability compute_cpu_capability() {
         #endif
         }
     #endif
+    #ifdef HAVE_SVE128_CPU_DEFINITION
+        if (sve_vl == 128) { // Check for SVE128
+            return CPUCapability::SVE128;
+        }
+    #endif
     // Return the default CPU capability.
     return CPUCapability::DEFAULT;
   }
@@ -147,6 +154,9 @@ DispatchResult DispatchStubImpl::try_get_call_ptr(
 #ifdef HAVE_SVE256_CPU_DEFINITION
   , void *SVE256
 #endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+  , void *SVE128
+#endif
 ) {
   constexpr auto supported_devices = c10::array_of<c10::DeviceType>(
         c10::DeviceType::CPU,
@@ -184,6 +194,9 @@ DispatchResult DispatchStubImpl::try_get_call_ptr(
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
           , SVE256
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+          , SVE128
 #endif
         );
         if (!std::holds_alternative<ErrorType>(result)) {
@@ -242,6 +255,9 @@ void* DispatchStubImpl::get_call_ptr(
 #ifdef HAVE_SVE256_CPU_DEFINITION
   , void *SVE256
 #endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+  , void *SVE128
+#endif
 ) {
 
   auto result = try_get_call_ptr(
@@ -266,6 +282,10 @@ void* DispatchStubImpl::get_call_ptr(
 #ifdef HAVE_SVE256_CPU_DEFINITION
       ,
       SVE256
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+      ,
+      SVE128
 #endif
   );
   if (std::holds_alternative<ErrorType>(result)) {
@@ -300,6 +320,9 @@ DispatchResult DispatchStubImpl::try_choose_cpu_impl(
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
     , void *SVE256
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+    , void *SVE128
 #endif
   ){
 
@@ -342,6 +365,16 @@ DispatchResult DispatchStubImpl::try_choose_cpu_impl(
       return DispatchResult(SVE256);
     }
   }
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+  if (capability >= static_cast<int>(CPUCapability::SVE128)) {
+    if (C10_UNLIKELY(!SVE128)) {
+      // dispatch to DEFAULT, since the SVE kernel is missing
+      return DEFAULT != nullptr ? DispatchResult(DEFAULT) : ErrorType::MissingDeviceKernel;
+    } else {
+      return DispatchResult(SVE128);
+    }
+  }
 #endif
   return DEFAULT != nullptr ? DispatchResult(DEFAULT) : ErrorType::MissingDeviceKernel;
 }
@@ -363,6 +396,9 @@ void* DispatchStubImpl::choose_cpu_impl(
 #ifdef HAVE_SVE256_CPU_DEFINITION
   , void *SVE256
 #endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+  , void *SVE128
+#endif
 ) {
   auto capability = static_cast<int>(get_cpu_capability());
   (void)capability;
@@ -408,6 +444,17 @@ void* DispatchStubImpl::choose_cpu_impl(
       return SVE256;
     }
   }
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+  if (capability >= static_cast<int>(CPUCapability::SVE128)) {
+    if (C10_UNLIKELY(!SVE128)) {
+      // dispatch to DEFAULT, since the SVE kernel is missing
+      TORCH_INTERNAL_ASSERT(DEFAULT, "DispatchStub: missing default kernel");
+      return DEFAULT;
+    } else {
+      return SVE128;
+    }
+  }
 #endif
   TORCH_INTERNAL_ASSERT(DEFAULT, "DispatchStub: missing default kernel");
   return DEFAULT;

aten/src/ATen/native/DispatchStub.h

@@ -64,8 +64,9 @@ enum class CPUCapability {
   VSX = 1,
 #elif defined(HAVE_ZVECTOR_CPU_DEFINITION)
   ZVECTOR = 1,
-#elif defined(HAVE_SVE256_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
+#elif defined(HAVE_SVE_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
   SVE256 = 1,
+  SVE128 = 2,
 #else
   AVX2 = 1,
   AVX512 = 2,
@@ -117,6 +118,9 @@ struct TORCH_API DispatchStubImpl {
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
       , void *SVE256
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+      , void *SVE128
 #endif
   );
 
@@ -138,6 +142,9 @@ struct TORCH_API DispatchStubImpl {
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
     , void *SVE256
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+    , void *SVE128
 #endif
   );
 
@@ -159,6 +166,9 @@ struct TORCH_API DispatchStubImpl {
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
       , void *SVE256
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+      , void *SVE128
 #endif
   );
 
@@ -183,6 +193,9 @@ struct TORCH_API DispatchStubImpl {
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
     , void *SVE256
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+    , void *SVE128
 #endif
   );
 
@@ -240,6 +253,9 @@ private:
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
       , reinterpret_cast<void*>(SVE256)
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+      , reinterpret_cast<void*>(SVE128)
 #endif
       )
     );
@@ -301,6 +317,9 @@ public:
 #endif
 #ifdef HAVE_SVE256_CPU_DEFINITION
       , reinterpret_cast<void*>(SVE256)
+#endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+      , reinterpret_cast<void*>(SVE128)
 #endif
       );
     if (std::holds_alternative<ErrorType>(result)){
@@ -325,6 +344,9 @@ public:
 #ifdef HAVE_SVE256_CPU_DEFINITION
   static TORCH_API FnPtr SVE256;
 #endif
+#ifdef HAVE_SVE128_CPU_DEFINITION
+  static TORCH_API FnPtr SVE128;
+#endif
 private:
   DispatchStubImpl impl;
 };
@@ -432,6 +454,12 @@ struct RegisterPRIVATEUSE1Dispatch {
 #define REGISTER_SVE256_DISPATCH(name, fn)
 #endif
 
+#ifdef HAVE_SVE128_CPU_DEFINITION
+#define REGISTER_SVE128_DISPATCH(name, fn) REGISTER_ARCH_DISPATCH(name, SVE128, fn)
+#else
+#define REGISTER_SVE128_DISPATCH(name, fn)
+#endif
+
 // Macro to register the same kernel for all CPU arch types. This is useful
 // if a kernel does not benefit from being recompiled across different arch types.
 #define REGISTER_ALL_CPU_DISPATCH(name, fn)                                    \
@@ -440,6 +468,11 @@ struct RegisterPRIVATEUSE1Dispatch {
   REGISTER_AVX2_DISPATCH(name, fn)                                             \
   REGISTER_VSX_DISPATCH(name, fn)                                              \
   REGISTER_ZVECTOR_DISPATCH(name, fn)                                          \
+  REGISTER_SVE256_DISPATCH(name, fn)                                           \
+  REGISTER_SVE128_DISPATCH(name, fn)
+
+#define REGISTER_SVE_DISPATCH(name, fn)                                        \
+  REGISTER_SVE128_DISPATCH(name, fn)                                           \
   REGISTER_SVE256_DISPATCH(name, fn)
 
 #define REGISTER_NO_CPU_DISPATCH(name)                                         \
@@ -482,6 +515,7 @@ struct RegisterPRIVATEUSE1Dispatch {
 // REGISTER_DISPATCH now dispatches an AVX512 kernel to nullptr but registers other dispatches.
 // ALSO_REGISTER_AVX512_DISPATCH should be used for ensuring AVX512 dispatch, among others.
 // ALSO_REGISTER_SVE256_DISPATCH should be used for ensuring SVE256 dispatch, among others.
+// ALSO_REGISTER_SVE128_DISPATCH should be used for ensuring SVE128 dispatch, among others.
 #ifdef CPU_CAPABILITY_AVX512
 #define REGISTER_DISPATCH(name, fn) REGISTER_ARCH_DISPATCH(name, CPU_CAPABILITY, ((void*)(fn) ? nullptr : nullptr))
 #else
@@ -489,6 +523,7 @@ struct RegisterPRIVATEUSE1Dispatch {
 #endif
 #define ALSO_REGISTER_AVX512_DISPATCH(name, fn) REGISTER_ARCH_DISPATCH(name, CPU_CAPABILITY, fn)
 #define ALSO_REGISTER_SVE256_DISPATCH(name, fn) REGISTER_ARCH_DISPATCH(name, CPU_CAPABILITY, fn)
+#define ALSO_REGISTER_SVE128_DISPATCH(name, fn) REGISTER_ARCH_DISPATCH(name, CPU_CAPABILITY, fn)
 #endif
 } // namespace at::native
 

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>
-std::enable_if_t<std::is_same_v<data_t, double>, void>
+static 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>
-std::enable_if_t<std::is_same_v<data_t, double>, void>
+static 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>
-void fractional_max_pool2d_out_single_batch_frame(
+static void fractional_max_pool2d_out_single_batch_frame(
   const scalar_t* input,
   scalar_t* output,
   int64_t* indices,
@@ -188,7 +188,7 @@ void fractional_max_pool2d_out_single_batch_frame(
 }
 
 template <typename scalar_t>
-void fractional_max_pool2d_out_frame(
+static void fractional_max_pool2d_out_frame(
   const scalar_t* input,
   scalar_t* output,
   int64_t* indices,
@@ -220,7 +220,7 @@ void fractional_max_pool2d_out_frame(
   }
 
 template <typename scalar_t>
-void fractional_max_pool2d_backward_out_single_batch_frame(
+static void fractional_max_pool2d_backward_out_single_batch_frame(
   scalar_t* gradInput,
   const scalar_t* gradOutput,
   const int64_t* indices,
@@ -247,7 +247,7 @@ void fractional_max_pool2d_backward_out_single_batch_frame(
 }
 
 template <typename scalar_t>
-void fractional_max_pool2d_backward_out_frame(
+static 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>
-void fractional_max_pool3d_out_single_batch_frame(
+static void fractional_max_pool3d_out_single_batch_frame(
   const scalar_t* input,
   scalar_t* output,
   int64_t* indices,
@@ -169,7 +169,7 @@ void fractional_max_pool3d_out_single_batch_frame(
 }
 
 template<typename scalar_t>
-void fractional_max_pool3d_out_frame(
+static 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>
-void fractional_max_pool3d_backward_out_single_batch_frame(
+static void fractional_max_pool3d_backward_out_single_batch_frame(
   scalar_t* gradInput,
   const scalar_t* gradOutput,
   const int64_t* indices,
@@ -287,7 +287,7 @@ void fractional_max_pool3d_backward_out_single_batch_frame(
 }
 
 template<typename scalar_t>
-void fractional_max_pool3d_backward_out_frame(
+static 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 {
 
-void im2col_out_cpu_template(
+static 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 {
-  inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
+  static 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>
-inline int64_t get_target_prime(target_t* target, int64_t offset, int64_t stride, int64_t idx, int64_t BLANK) {
+static 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>
-void multilabel_margin_loss_forward_out_frame(
+static void multilabel_margin_loss_forward_out_frame(
     const Tensor& input_contiguous,
     const Tensor& target_contiguous,
     Tensor& output,
@@ -108,7 +108,7 @@ void multilabel_margin_loss_forward_out_frame(
   }
 }
 
-void multilabel_margin_loss_forward_out_cpu_template(
+static void multilabel_margin_loss_forward_out_cpu_template(
     const Tensor& input,
     const Tensor& target,
     Tensor& output,
@@ -153,7 +153,7 @@ void multilabel_margin_loss_forward_out_cpu_template(
 }
 
 template <typename scalar_t>
-void multilabel_margin_loss_backward_out_frame(
+static void multilabel_margin_loss_backward_out_frame(
     Tensor& grad_input,
     const Tensor& grad_output,
     const Tensor& input_contiguous,
@@ -222,7 +222,7 @@ void multilabel_margin_loss_backward_out_frame(
   }
 }
 
-void multilabel_margin_loss_backward_out_cpu_template(
+static 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>
-inline void multi_margin_loss_cpu_kernel(
+static 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>
-void multi_margin_loss_backward_cpu_kernel(
+static 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>
-void nll_loss_out_frame(
+static 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>
-void nll_loss_backward_out_frame(
+static 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>
-void nll_loss2d_forward_out_frame(
+static 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>
-void nll_loss2d_backward_out_frame(
+static void nll_loss2d_backward_out_frame(
     Tensor& grad_input,
     const Tensor& grad_output,
     const Tensor& input,