Relocate channels-last cat regression test

.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

@@ -113,7 +113,6 @@ case "$tag" in
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
-    INSTALL_MINGW=yes
     ;;
   pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
     CUDA_VERSION=13.0.0
@@ -182,7 +181,7 @@ case "$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 [[ $tag =~ "benchmarks" ]]; then
       INDUCTOR_BENCHMARKS=yes
     fi
@@ -345,7 +344,7 @@ docker build \
        --build-arg "NINJA_VERSION=${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=${UCC_COMMIT}" \
@@ -362,7 +361,6 @@ docker build \
        --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
        --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
-       --build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \
        -t "$tmp_tag" \
        "$@" \

.ci/docker/common/install_cpython.sh

@@ -83,6 +83,10 @@ function build_cpython {
         py_suffix=${py_ver::-1}
         py_folder=$py_suffix
     fi
+    # Update to rc2 due to https://github.com/python/cpython/commit/c72699086fe4
+    if [ "$py_suffix" == "3.14.0" ]; then
+        py_suffix="3.14.0rc2"
+    fi
     wget -q $PYTHON_DOWNLOAD_URL/$py_folder/Python-$py_suffix.tgz -O Python-$py_ver.tgz
     do_cpython_build $py_ver Python-$py_suffix
 

.ci/docker/common/install_mingw.sh

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

.ci/docker/common/install_onnx.sh

@@ -20,7 +20,7 @@ pip_install \
 
 pip_install coloredlogs packaging
 pip_install onnxruntime==1.23.0
-pip_install onnxscript==0.5.4
+pip_install onnxscript==0.5.3
 
 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/

.ci/docker/libtorch/build.sh

@@ -39,13 +39,9 @@ case ${DOCKER_TAG_PREFIX} in
         DOCKER_GPU_BUILD_ARG=""
         ;;
     rocm*)
-        # we want the patch version of 7.0 instead
-        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
-        fi
         # we want the patch version of 6.4 instead
         if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

.ci/docker/manywheel/build.sh

@@ -75,13 +75,9 @@ case ${image} in
         DOCKERFILE_SUFFIX="_cuda_aarch64"
         ;;
     manylinux2_28-builder:rocm*)
-        # we want the patch version of 7.0 instead
-        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
-        fi
         # we want the patch version of 6.4 instead
         if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
         fi
         TARGET=rocm_final
         MANY_LINUX_VERSION="2_28"

.ci/docker/ubuntu/Dockerfile

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

.ci/lumen_cli/cli/lib/common/git_helper.py

@@ -57,8 +57,8 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
         logger.info("Successfully cloned %s", target)
         return r, commit
 
-    except GitCommandError:
-        logger.exception("Git operation failed")
+    except GitCommandError as e:
+        logger.error("Git operation failed: %s", e)
         raise
 
 

.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/build.sh

@@ -233,9 +233,7 @@ if [[ "${BUILD_ENVIRONMENT}" != *cuda* ]]; then
   export BUILD_STATIC_RUNTIME_BENCHMARK=ON
 fi
 
-if [[ "$BUILD_ENVIRONMENT" == *-full-debug* ]]; then
-  export CMAKE_BUILD_TYPE=Debug
-elif [[ "$BUILD_ENVIRONMENT" == *-debug* ]]; then
+if [[ "$BUILD_ENVIRONMENT" == *-debug* ]]; then
   export CMAKE_BUILD_TYPE=RelWithAssert
 fi
 
@@ -301,11 +299,6 @@ else
       python -m build --wheel --no-isolation
     fi
     pip_install_whl "$(echo dist/*.whl)"
-    if [[ "$BUILD_ENVIRONMENT" == *full-debug* ]]; then
-      # Regression test for https://github.com/pytorch/pytorch/issues/164297
-      # Torch should be importable and that's about it
-      pushd /; python -c "import torch;print(torch.__config__.show(), torch.randn(5) + 1.7)"; popd
-    fi
 
     if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *vision* ]]; then
       install_torchvision

.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
 }
 
@@ -485,22 +485,6 @@ test_inductor_aoti() {
   /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
-test_inductor_aoti_cross_compile_for_windows() {
-
-  TEST_REPORTS_DIR=$(pwd)/test/test-reports
-  mkdir -p "$TEST_REPORTS_DIR"
-
-  # Set WINDOWS_CUDA_HOME environment variable
-  WINDOWS_CUDA_HOME="$(pwd)/win-torch-wheel-extracted"
-  export WINDOWS_CUDA_HOME
-
-  echo "WINDOWS_CUDA_HOME is set to: $WINDOWS_CUDA_HOME"
-  echo "Contents:"
-  ls -lah "$(pwd)/win-torch-wheel-extracted/lib/x64/" || true
-
-  python test/inductor/test_aoti_cross_compile_windows.py -k compile --package-dir "$TEST_REPORTS_DIR" --win-torch-lib-dir "$(pwd)/win-torch-wheel-extracted/torch/lib"
-}
-
 test_inductor_cpp_wrapper_shard() {
   if [[ -z "$NUM_TEST_SHARDS" ]]; then
     echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@@ -916,7 +900,7 @@ test_inductor_set_cpu_affinity(){
   export LD_PRELOAD="$JEMALLOC_LIB":"$LD_PRELOAD"
   export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1"
 
-  if [[ "$(uname -m)" != "aarch64" ]]; then
+  if [[ "${TEST_CONFIG}" != *aarch64* ]]; then
     # Use Intel OpenMP for x86
     IOMP_LIB="$(dirname "$(which python)")/../lib/libiomp5.so"
     export LD_PRELOAD="$IOMP_LIB":"$LD_PRELOAD"
@@ -930,7 +914,7 @@ test_inductor_set_cpu_affinity(){
   cores=$((cpus / thread_per_core))
 
   # Set number of cores to 16 on aarch64 for performance runs
-  if [[ "$(uname -m)" == "aarch64" && $cores -gt 16 ]]; then
+  if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
     cores=16
   fi
   export OMP_NUM_THREADS=$cores
@@ -1631,7 +1615,6 @@ test_operator_benchmark() {
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
   TEST_DIR=$(pwd)
-  ARCH=$(uname -m)
 
   test_inductor_set_cpu_affinity
 
@@ -1646,7 +1629,7 @@ test_operator_benchmark() {
   pip_install pandas
   python check_perf_csv.py \
       --actual "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
-      --expected "${ARCH}_expected_ci_operator_benchmark_eager_float32_cpu.csv"
+      --expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
 }
 
 test_operator_microbenchmark() {
@@ -1683,7 +1666,7 @@ if [[ "${TEST_CONFIG}" == *numpy_2* ]]; then
     python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
   fi
   python test/run_test.py --include dynamo/test_functions.py dynamo/test_unspec.py test_binary_ufuncs.py test_fake_tensor.py test_linalg.py test_numpy_interop.py test_tensor_creation_ops.py test_torch.py torch_np/test_basic.py
-elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]]; then
+elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
   test_linux_aarch64
 elif [[ "${TEST_CONFIG}" == *backward* ]]; then
   test_forward_backward_compatibility
@@ -1734,8 +1717,6 @@ elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
   test_inductor_micro_benchmark
-elif [[ "${TEST_CONFIG}" == *aoti_cross_compile_for_windows* ]]; then
-  test_inductor_aoti_cross_compile_for_windows
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
   install_torchvision
   id=$((SHARD_NUMBER-1))

.flake8

@@ -7,12 +7,16 @@ max-line-length = 120
 # C408 ignored because we like the dict keyword argument syntax
 # E501 is not flexible enough, we're using B950 instead
 ignore =
-    E203,E305,E402,E501,E704,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
     # shebang has extra meaning in fbcode lints, so I think it's not worth trying
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
-    B007,B008,B017,B019,B023,B028,B903,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!
+    G100,G101,G200
     # these ignores are from flake8-simplify. please fix or ignore with commented reason
     SIM105,SIM108,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
     # SIM104 is already covered by pyupgrade ruff

.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/build-external-packages/action.yml

@@ -65,7 +65,7 @@ runs:
           cd .ci/lumen_cli
           python3 -m pip install -e .
         )
-        MAX_JOBS="$(nproc --ignore=10)"
+        MAX_JOBS="$(nproc --ignore=6)"
         export MAX_JOBS
 
         # Split the comma-separated list and build each target

.github/actions/linux-test/action.yml

@@ -274,6 +274,8 @@ runs:
           -w /var/lib/jenkins/workspace \
           "${DOCKER_IMAGE}"
         )
+        # Propagate download.pytorch.org IP to container
+        grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" sudo bash -c "/bin/cat >> /etc/hosts"
         echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
         docker exec -t "${container_name}" sh -c "pip install $(echo dist/*.whl)[opt-einsum] && ${TEST_COMMAND}"
 

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

@@ -33,6 +33,10 @@ runs:
         )
 
         echo "CONTAINER_NAME=${container_name}" >> "$GITHUB_ENV"
+        if [[ "${GPU_ARCH_TYPE}" != "rocm" && "${BUILD_ENVIRONMENT}" != "linux-aarch64-binary-manywheel" && "${BUILD_ENVIRONMENT}" != "linux-s390x-binary-manywheel" && "${GPU_ARCH_TYPE}" != "xpu" ]]; then
+          # Propagate download.pytorch.org IP to container. This is only needed on Linux non aarch64 runner
+          grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" bash -c "/bin/cat >> /etc/hosts"
+        fi
 
         docker exec -t -w "${PYTORCH_ROOT}" "${container_name}" bash -c "bash .circleci/scripts/binary_populate_env.sh"
         # Generate test script

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-69bbe7363897764f9e758d851cd0340147d27f94
+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/labeler.yml

@@ -133,32 +133,3 @@
 
 "ciflow/vllm":
 - .github/ci_commit_pins/vllm.txt
-
-"ciflow/b200":
-- test/test_matmul_cuda.py
-- test/test_scaled_matmul_cuda.py
-- test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/Blas.cpp
-- torch/**/*cublas*
-- torch/_inductor/kernel/mm.py
-- test/inductor/test_max_autotune.py
-- third_party/fbgemm
-
-"ciflow/h100":
-- test/test_matmul_cuda.py
-- test/test_scaled_matmul_cuda.py
-- test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/Blas.cpp
-- torch/**/*cublas*
-- torch/_inductor/kernel/mm.py
-- test/inductor/test_max_autotune.py
-- third_party/fbgemm
-
-"ciflow/rocm":
-- test/test_matmul_cuda.py
-- test/test_scaled_matmul_cuda.py
-- test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/Blas.cpp
-- torch/_inductor/kernel/mm.py
-- test/inductor/test_max_autotune.py
-- third_party/fbgemm

.github/pytorch-probot.yml

@@ -3,7 +3,6 @@ ciflow_tracking_issue: 64124
 ciflow_push_tags:
 - ciflow/b200
 - ciflow/b200-symm-mem
-- ciflow/b200-distributed
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
@@ -16,8 +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
@@ -32,8 +30,6 @@ ciflow_push_tags:
 - ciflow/riscv64
 - ciflow/rocm
 - ciflow/rocm-mi300
-- ciflow/rocm-mi355
-- ciflow/rocm-navi31
 - ciflow/s390
 - ciflow/slow
 - ciflow/torchbench

.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' | "
-        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
-    ),
     "13.0": (
         "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
         "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
@@ -241,11 +222,7 @@ def generate_libtorch_matrix(
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
         elif os == "windows":
-            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
-            # in 2.10
-            windows_cuda_arches = CUDA_ARCHES.copy()
-            windows_cuda_arches.remove("12.9")
-            arches += windows_cuda_arches
+            arches += CUDA_ARCHES
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -309,11 +286,7 @@ def generate_wheels_matrix(
         if os == "linux":
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
-            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
-            # in 2.10
-            windows_cuda_arches = CUDA_ARCHES.copy()
-            windows_cuda_arches.remove("12.9")
-            arches += windows_cuda_arches + XPU_ARCHES
+            arches += CUDA_ARCHES + XPU_ARCHES
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -349,7 +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
             ):
@@ -413,6 +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

@@ -1092,7 +1092,7 @@ class GitHubPR:
         editor = node["editor"]
         return GitHubComment(
             body_text=node["bodyText"],
-            created_at=node.get("createdAt", ""),
+            created_at=node["createdAt"] if "createdAt" in node else "",
             author_login=node["author"]["login"],
             author_url=node["author"].get("url", None),
             author_association=node["authorAssociation"],
@@ -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/templates/macos_binary_build_workflow.yml.j2

@@ -26,8 +26,9 @@ name: !{{ build_environment }}
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "!{{ py_ver.strip('t') + ('.4' if '3.14' not in py_ver else '.0') }}"
+          python-version: "!{{ (py_ver.strip('t') + '.4') if '3.14' not in py_ver else '3.14.0-rc.2' }}"
           freethreaded: !{{ "true" if py_ver.endswith('t') else "false" }}
 {%- endmacro %}
 

.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-build.yml

@@ -37,7 +37,7 @@ on:
       runner:
         required: false
         type: string
-        default: "linux.c7i.2xlarge"
+        default: "linux.2xlarge"
         description: |
           Label of the runner this job should run on.
       test-matrix:

.github/workflows/_linux-test.yml

@@ -224,46 +224,6 @@ jobs:
         continue-on-error: true
         uses: ./.github/actions/download-td-artifacts
 
-      - name: Download Windows torch wheel for cross-compilation
-        if: matrix.win_torch_wheel_artifact != ''
-        uses: seemethere/download-artifact-s3@1da556a7aa0a088e3153970611f6c432d58e80e6 # v4.2.0
-        with:
-          name: ${{ matrix.win_torch_wheel_artifact }}
-          path: win-torch-wheel
-
-      - name: Extract Windows wheel and setup CUDA libraries
-        if: matrix.win_torch_wheel_artifact != ''
-        shell: bash
-        run: |
-          set -x
-
-          # Find the wheel file
-          WHEEL_FILE=$(find win-torch-wheel -name "*.whl" -type f | head -n 1)
-          if [ -z "$WHEEL_FILE" ]; then
-            echo "Error: No wheel file found in win-torch-wheel directory"
-            exit 1
-          fi
-          echo "Found wheel file: $WHEEL_FILE"
-
-          # Unzip the wheel file
-          unzip -q "$WHEEL_FILE" -d win-torch-wheel-extracted
-          echo "Extracted wheel contents"
-
-          # Setup CUDA libraries (cuda.lib and cudart.lib) directory
-          mkdir -p win-torch-wheel-extracted/lib/x64
-          if [ -f "win-torch-wheel/cuda.lib" ]; then
-            mv win-torch-wheel/cuda.lib win-torch-wheel-extracted/lib/x64/
-            echo "Moved cuda.lib to win-torch-wheel-extracted/lib/x64/"
-          fi
-          if [ -f "win-torch-wheel/cudart.lib" ]; then
-            mv win-torch-wheel/cudart.lib win-torch-wheel-extracted/lib/x64/
-            echo "Moved cudart.lib to win-torch-wheel-extracted/lib/x64/"
-          fi
-
-          # Verify CUDA libraries are present
-          echo "CUDA libraries:"
-          ls -la win-torch-wheel-extracted/lib/x64/ || echo "No CUDA libraries found"
-
       - name: Parse ref
         id: parse-ref
         run: .github/scripts/parse_ref.py
@@ -429,6 +389,8 @@ jobs:
             "${DOCKER_IMAGE}" \
             ${DOCKER_SHELL_CMD}
           )
+          # Propagate download.pytorch.org IP to container
+          grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" sudo bash -c "/bin/cat >> /etc/hosts"
           echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
 
           if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then

.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/_win-build.yml

@@ -168,31 +168,6 @@ jobs:
         run: |
           .ci/pytorch/win-build.sh
 
-      # Collect Windows torch libs and CUDA libs for cross-compilation
-      - name: Collect Windows CUDA libs for cross-compilation
-        if: steps.build.outcome != 'skipped' && inputs.cuda-version != 'cpu'
-        shell: bash
-        run: |
-          set -ex
-
-          # Create directory structure if does not exist
-          mkdir -p /c/${{ github.run_id }}/build-results
-
-          # Copy CUDA libs
-          CUDA_PATH="/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v${{ inputs.cuda-version }}"
-
-          if [ -f "${CUDA_PATH}/lib/x64/cuda.lib" ]; then
-            cp "${CUDA_PATH}/lib/x64/cuda.lib" /c/${{ github.run_id }}/build-results/
-          fi
-
-          if [ -f "${CUDA_PATH}/lib/x64/cudart.lib" ]; then
-            cp "${CUDA_PATH}/lib/x64/cudart.lib" /c/${{ github.run_id }}/build-results/
-          fi
-
-          # List collected files
-          echo "Collected CUDA libs:"
-          ls -lah /c/${{ github.run_id }}/build-results/*.lib
-
       # Upload to github so that people can click and download artifacts
       - name: Upload artifacts to s3
         if: steps.build.outcome != 'skipped'

.github/workflows/b200-distributed.yml

@@ -1,62 +0,0 @@
-name: CI for distributed tests on B200
-
-on:
-  pull_request:
-    paths:
-      - .github/workflows/b200-distributed.yml
-  workflow_dispatch:
-  push:
-    tags:
-      - ciflow/b200-distributed/*
-  schedule:
-    - cron: 46 8 * * *  # about 1:46am PDT
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-
-  get-label-type:
-    if: github.repository_owner == 'pytorch'
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-build-distributed-b200
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '10.0'
-      test-matrix: |
-        { include: [
-          { config: "distributed", shard: 1, num_shards: 2, runner: "linux.dgx.b200.8" },
-          { config: "distributed", shard: 2, num_shards: 2, runner: "linux.dgx.b200.8" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc11-test-distributed-b200:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-test-b200
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200
-    with:
-      timeout-minutes: 1200
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-    secrets: inherit

.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/build-vllm-wheel.yml

@@ -27,8 +27,9 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [ '3.12' ]
+        # TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
         platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129', 'cu130' ]
+        device: [ 'cu128', 'cu129' ]
         include:
           - platform: manylinux_2_28_x86_64
             device: cu128
@@ -38,10 +39,6 @@ jobs:
             device: cu129
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
             runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_x86_64
-            device: cu130
-            manylinux-image: 'pytorch/manylinux2_28-builder:cuda13.0'
-            runner: linux.12xlarge.memory
           - platform: manylinux_2_28_aarch64
             device: cu128
             manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
@@ -50,11 +47,6 @@ jobs:
             device: cu129
             manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
             runner: linux.arm64.r7g.12xlarge.memory
-        exclude:
-          # TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
-          # xformers is update to support 13.0
-          - platform: manylinux_2_28_aarch64
-            device: cu130
     name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 480
@@ -177,12 +169,7 @@ jobs:
       fail-fast: false
       matrix:
         platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129', 'cu130' ]
-        exclude:
-          # TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
-          # xformers is update to support 13.0
-          - platform: manylinux_2_28_aarch64
-            device: cu130
+        device: [ 'cu128', 'cu129' ]
     env:
       PLATFORM: ${{ matrix.platform }}
       BUILD_DEVICE: ${{ matrix.device }}

.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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-      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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-      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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-      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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-      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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-      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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-      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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-      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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-    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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-    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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-    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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-    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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-    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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-    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' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
-    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-macos-arm64-binary-libtorch-release-nightly.yml

@@ -63,6 +63,7 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.10.4"
           freethreaded: false

.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml

@@ -59,6 +59,7 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.10.4"
           freethreaded: false
@@ -168,6 +169,7 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.11.4"
           freethreaded: false
@@ -277,6 +279,7 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.12.4"
           freethreaded: false
@@ -386,6 +389,7 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.13.4"
           freethreaded: false
@@ -495,6 +499,7 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.13.4"
           freethreaded: true
@@ -604,8 +609,9 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0"
+          python-version: "3.14.0-rc.2"
           freethreaded: false
       - name: Checkout PyTorch
         uses: actions/checkout@v4
@@ -713,8 +719,9 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
+          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0"
+          python-version: "3.14.0-rc.2"
           freethreaded: true
       - name: Checkout PyTorch
         uses: actions/checkout@v4

.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.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { 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-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     with:
       timeout: 120
@@ -118,9 +113,9 @@ jobs:
         CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
         echo "Running all other linters"
         if [ "$CHANGED_FILES" = '*' ]; then
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY --all-files" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT --all-files" .github/scripts/lintrunner.sh
         else
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT ${CHANGED_FILES}" .github/scripts/lintrunner.sh
         fi
 
   quick-checks:

.github/workflows/operator_benchmark.yml

@@ -7,11 +7,9 @@ on:
   workflow_dispatch:
     inputs:
       test_mode:
-        type: choice
-        options:
-          - 'short'
-          - 'long'
-          - 'all'
+        required: false
+        type: string
+        default: 'short'
         description: tag filter for operator benchmarks, options from long, short, all
   schedule:
     # Run at 07:00 UTC every Sunday
@@ -30,49 +28,38 @@ permissions:
   contents: read
 
 jobs:
-  x86-opbenchmark-build:
+  opbenchmark-build:
     if: github.repository_owner == 'pytorch'
-    name: x86-opbenchmark-build
+    name: opbenchmark-build
     uses: ./.github/workflows/_linux-build.yml
     with:
       build-environment: linux-jammy-py3.10-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
-          { config: "cpu_operator_benchmark_${{ inputs.test_mode || 'short' }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
+          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
         ]}
     secrets: inherit
 
-  x86-opbenchmark-test:
-    name: x86-opbenchmark-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs: x86-opbenchmark-build
-    with:
-      build-environment: linux-jammy-py3.10-gcc11-build
-      docker-image: ${{ needs.x86-opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.x86-opbenchmark-build.outputs.test-matrix }}
-    secrets: inherit
-
-  aarch64-opbenchmark-build:
-    if: github.repository_owner == 'pytorch'
-    name: aarch64-opbenchmark-build
+  opbenchmark-on-demand-build:
+    if: ${{ github.event_name == 'workflow_dispatch' && github.repository_owner == 'pytorch' }}
+    name: opbenchmark-on-demand-build
     uses: ./.github/workflows/_linux-build.yml
     with:
-      build-environment: linux-jammy-aarch64-py3.10
-      runner: linux.arm64.m7g.4xlarge
-      docker-image-name: ci-image:pytorch-linux-jammy-aarch64-py3.10-gcc11
+      build-environment: linux-jammy-py3.10-gcc11-build
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
-          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.arm64.m8g.4xlarge" },
+          { config: "cpu_operator_benchmark_${{ inputs.test_mode }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
         ]}
     secrets: inherit
 
-  aarch64-opbenchmark-test:
-    name: aarch64-opbenchmark-test
+  opbenchmark-test:
+    name: opbenchmark-test
     uses: ./.github/workflows/_linux-test.yml
-    needs: aarch64-opbenchmark-build
+    needs: opbenchmark-build
     with:
-      build-environment: linux-jammy-aarch64-py3.10
-      docker-image: ${{ needs.aarch64-opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.aarch64-opbenchmark-build.outputs.test-matrix }}
+      build-environment: linux-jammy-py3.10-gcc11-build
+      docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
     secrets: inherit

.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/rocm-mi355.yml

@@ -1,9 +1,6 @@
 name: rocm-mi355
 
 on:
-  push:
-    tags:
-      - ciflow/rocm-mi355/*
   workflow_dispatch:
   schedule:
     - cron: 30 11,1 * * *  # about 4:30am PDT and 6:30pm PDT
@@ -45,12 +42,12 @@ jobs:
       sync-tag: rocm-build
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
         ]}
     secrets: inherit
 
@@ -67,7 +64,5 @@ jobs:
       build-environment: linux-noble-rocm-py3.12-mi355
       docker-image: ${{ needs.linux-noble-rocm-py3_12-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-noble-rocm-py3_12-build.outputs.test-matrix }}
-      tests-to-include: >-
-                        ${{ github.event_name == 'schedule' && 'test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor test_matmul_cuda test_scaled_matmul_cuda'
-                           || '' }}
+      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
     secrets: inherit

.github/workflows/rocm-navi31.yml

@@ -1,63 +0,0 @@
-name: rocm-navi31
-
-on:
-  push:
-    tags:
-      - ciflow/rocm-navi31/*
-  workflow_dispatch:
-  schedule:
-    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
-    # Also run less frequently on weekends.
-    - cron: 45 */2 * * 1-5
-    - cron: 45 4,12 * * 0,6
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  target-determination:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  linux-jammy-rocm-py3_10-build:
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    name: linux-jammy-rocm-py3.10
-    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
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3_10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-      tests-to-include: >-
-         ${{ github.event_name == 'schedule' && 'test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
-         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
-         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
-         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
-         inductor/test_flex_attention inductor/test_max_autotune' || '' }}
-    secrets: inherit

.github/workflows/rocm.yml

@@ -59,3 +59,29 @@ jobs:
       docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
     secrets: inherit
+
+  linux-jammy-rocm-py3_10-gfx1100-test:
+    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' }}
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3_10-gfx1100
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: |
+        { include: [
+          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+        ]}
+      tests-to-include: >
+         test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
+         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
+         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
+         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
+         inductor/test_flex_attention inductor/test_max_autotune
+    secrets: inherit

.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 },
@@ -180,50 +180,16 @@ 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
 
-  linux-jammy-rocm-py3_10-build:
-    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
-    secrets: inherit
-
   inductor-build:
     name: inductor-build
     uses: ./.github/workflows/_linux-build.yml
@@ -234,23 +200,6 @@ jobs:
       cuda-arch-list: '8.0'
     secrets: inherit
 
-  # Test cross-compiled models with Windows libs extracted from wheel
-  cross-compile-linux-test:
-    name: cross-compile-linux-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-build
-      - get-label-type
-      - win-vs2022-cuda12_8-py3-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
-      test-matrix: |
-        { include: [
-          { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
-        ]}
-    secrets: inherit
-
   verify-cachebench-cpu-build:
     name: verify-cachebench-cpu-build
     uses: ./.github/workflows/_linux-build.yml
@@ -300,14 +249,3 @@ jobs:
       docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
     secrets: inherit
-
-  linux-jammy-py3_10-gcc11-full-debug-build-only:
-    name: linux-jammy-py3.10-gcc11-full-debug-build-only
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: linux.2xlarge.memory
-      build-environment: linux-jammy-py3.10-gcc11-full-debug-build-only
-      docker-image-name: ci-image:pytorch-linux-jammy-py3.10-gcc11
-    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

@@ -374,7 +374,6 @@ third_party/ruy/
 third_party/glog/
 
 # Virtualenv
-.venv/
 venv/
 
 # Log files
@@ -396,4 +395,3 @@ android/pytorch_android_torchvision/.cxx
 CLAUDE.local.md
 /test_*.py
 /debug_*.py
-CLAUDE_CONTEXT/

.lintrunner.toml

@@ -209,46 +209,6 @@ command = [
     '@{{PATHSFILE}}'
 ]
 
-
-[[linter]]
-code = 'PYREFLY'
-include_patterns = [
-    'torch/**/*.py',
-    'torch/**/*.pyi',
-    'torchgen/**/*.py',
-    'torchgen/**/*.pyi',
-    'functorch/**/*.py',
-    'functorch/**/*.pyi',
-]
-exclude_patterns = []
-command = [
-    'python3',
-    'tools/linter/adapters/pyrefly_linter.py',
-    '--config=pyrefly.toml',
-]
-init_command = [
-    'python3',
-    'tools/linter/adapters/pip_init.py',
-    '--dry-run={{DRYRUN}}',
-    'numpy==2.1.0 ; python_version >= "3.12"',
-    'expecttest==0.3.0',
-    'pyrefly==0.36.2',
-    'sympy==1.13.3',
-    'types-requests==2.27.25',
-    'types-pyyaml==6.0.2',
-    'types-tabulate==0.8.8',
-    'types-protobuf==5.29.1.20250403',
-    'types-setuptools==79.0.0.20250422',
-    'types-jinja2==2.11.9',
-    'types-colorama==0.4.6',
-    'filelock==3.18.0',
-    'junitparser==2.1.1',
-    'rich==14.1.0',
-    'optree==0.17.0',
-    'types-openpyxl==3.1.5.20250919',
-    'types-python-dateutil==2.9.0.20251008'
-]
-
 [[linter]]
 code = 'CLANGTIDY'
 include_patterns = [

CMakeLists.txt

@@ -390,7 +390,7 @@ option(USE_MIMALLOC "Use mimalloc" OFF)
 # Enable third party mimalloc library to improve memory allocation performance
 # on Windows and AArch64.
 option(USE_MIMALLOC_ON_MKL "Use mimalloc on MKL" OFF)
-if(WIN32 OR (CPU_AARCH64 AND NOT APPLE))
+if(WIN32 OR CPU_AARCH64)
   set(USE_MIMALLOC ON)
 
   # Not enable USE_MIMALLOC_ON_MKL due to it caused issue:

CODEOWNERS

@@ -201,17 +201,3 @@ torch/backends/cudnn/ @eqy @syed-ahmed @Aidyn-A
 /torch/csrc/stable/ @janeyx99 @mikaylagawarecki
 /torch/headeronly/ @janeyx99
 /torch/header_only_apis.txt @janeyx99
-
-# FlexAttention
-/torch/nn/attention/flex_attention.py @drisspg
-/torch/_higher_order_ops/flex_attention.py @drisspg
-/torch/_inductor/kernel/flex/ @drisspg
-/torch/_inductor/codegen/cpp_flex_attention_template.py @drisspg
-/test/inductor/test_flex_attention.py @drisspg
-/test/inductor/test_flex_decoding.py @drisspg
-
-# Low Precision GEMMs
-/aten/src/ATen/native/cuda/Blas.cpp @drisspg @slayton58
-/aten/src/ATen/cuda/CUDABlas.cpp @drisspg @slayton58
-/aten/src/ATen/cuda/CUDABlas.h @drisspg @slayton58
-/test/test_scaled_matmul_cuda.py @drisspg @slayton58

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 ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
-
   if(USE_CUDA)
     # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
     # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
@@ -293,65 +292,58 @@ IF(USE_FBGEMM_GENAI)
       "${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
-      -enable-post-misched=0
-      -mllvm
-      -greedy-reverse-local-assignment=1
-      -fhip-new-launch-api)
-    if(DEFINED ROCM_VERSION_DEV AND ROCM_VERSION_DEV VERSION_LESS "7.2.0")
-        list(PREPEND FBGEMM_GENAI_EXTRA_HIPCC_FLAGS -mllvm -amdgpu-coerce-illegal-types=1)
+      # 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()
 
@@ -700,6 +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/PhiloxRNGEngine.h

@@ -229,10 +229,10 @@ private:
   }
 
 
-  static constexpr uint32_t kPhilox10A = 0x9E3779B9;
-  static constexpr uint32_t kPhilox10B = 0xBB67AE85;
-  static constexpr uint32_t kPhiloxSA = 0xD2511F53;
-  static constexpr uint32_t kPhiloxSB = 0xCD9E8D57;
+  static const uint32_t kPhilox10A = 0x9E3779B9;
+  static const uint32_t kPhilox10B = 0xBB67AE85;
+  static const uint32_t kPhiloxSA = 0xD2511F53;
+  static const uint32_t kPhiloxSB = 0xCD9E8D57;
 };
 
 typedef philox_engine Philox4_32;

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.h

@@ -8,7 +8,6 @@
 #include <ATen/cpu/vec/vec128/vec128_bfloat16_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_float_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_half_neon.h>
-#include <ATen/cpu/vec/vec128/vec128_int_aarch64.h>
 #endif
 
 #include <ATen/cpu/vec/vec128/vec128_convert.h>

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/vec128/vec128_int_aarch64.h

@@ -1,794 +0,0 @@
-#pragma once
-
-#include <ATen/cpu/vec/intrinsics.h>
-#include <ATen/cpu/vec/vec_base.h>
-#include <c10/macros/Macros.h>
-#include <c10/util/irange.h>
-
-namespace at::vec {
-// Note [CPU_CAPABILITY namespace]
-// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-// This header, and all of its subheaders, will be compiled with
-// different architecture flags for each supported set of vector
-// intrinsics. So we need to make sure they aren't inadvertently
-// linked together. We do this by declaring objects in an `inline
-// namespace` which changes the name mangling, but can still be
-// accessed as `at::vec`.
-inline namespace CPU_CAPABILITY {
-
-#define VEC_INT_NEON_TEMPLATE(vl, bit)                                        \
-  template <>                                                                 \
-  struct is_vec_specialized_for<int##bit##_t> : std::bool_constant<true> {};  \
-                                                                              \
-  template <>                                                                 \
-  class Vectorized<int##bit##_t> {                                            \
-    using neon_type = int##bit##x##vl##_t;                                    \
-                                                                              \
-   private:                                                                   \
-    neon_type values;                                                         \
-                                                                              \
-   public:                                                                    \
-    using value_type = int##bit##_t;                                          \
-    using size_type = int;                                                    \
-    static constexpr size_type size() {                                       \
-      return vl;                                                              \
-    }                                                                         \
-    Vectorized() {                                                            \
-      values = vdupq_n_s##bit(0);                                             \
-    }                                                                         \
-    Vectorized(neon_type v) : values(v) {}                                    \
-    Vectorized(int##bit##_t val);                                             \
-    template <                                                                \
-        typename... Args,                                                     \
-        typename = std::enable_if_t<(sizeof...(Args) == size())>>             \
-    Vectorized(Args... vals) {                                                \
-      __at_align__ int##bit##_t buffer[size()] = {vals...};                   \
-      values = vld1q_s##bit(buffer);                                          \
-    }                                                                         \
-    operator neon_type() const {                                              \
-      return values;                                                          \
-    }                                                                         \
-    static Vectorized<int##bit##_t> loadu(                                    \
-        const void* ptr,                                                      \
-        int64_t count = size());                                              \
-    void store(void* ptr, int64_t count = size()) const;                      \
-    template <int64_t mask>                                                   \
-    static Vectorized<int##bit##_t> blend(                                    \
-        const Vectorized<int##bit##_t>& a,                                    \
-        const Vectorized<int##bit##_t>& b);                                   \
-    static Vectorized<int##bit##_t> blendv(                                   \
-        const Vectorized<int##bit##_t>& a,                                    \
-        const Vectorized<int##bit##_t>& b,                                    \
-        const Vectorized<int##bit##_t>& mask_) {                              \
-      return vbslq_s##bit(vreinterpretq_u##bit##_s##bit(mask_.values), b, a); \
-    }                                                                         \
-    template <typename step_t>                                                \
-    static Vectorized<int##bit##_t> arange(                                   \
-        value_type base = 0,                                                  \
-        step_t step = static_cast<step_t>(1));                                \
-    static Vectorized<int##bit##_t> set(                                      \
-        const Vectorized<int##bit##_t>& a,                                    \
-        const Vectorized<int##bit##_t>& b,                                    \
-        int64_t count = size());                                              \
-    const int##bit##_t& operator[](int idx) const = delete;                   \
-    int##bit##_t& operator[](int idx) = delete;                               \
-    Vectorized<int##bit##_t> abs() const {                                    \
-      return vabsq_s##bit(values);                                            \
-    }                                                                         \
-    Vectorized<int##bit##_t> real() const {                                   \
-      return values;                                                          \
-    }                                                                         \
-    Vectorized<int##bit##_t> imag() const {                                   \
-      return vdupq_n_s##bit(0);                                               \
-    }                                                                         \
-    Vectorized<int##bit##_t> conj() const {                                   \
-      return values;                                                          \
-    }                                                                         \
-    Vectorized<int##bit##_t> neg() const {                                    \
-      return vnegq_s##bit(values);                                            \
-    }                                                                         \
-    int##bit##_t reduce_add() const {                                         \
-      return vaddvq_s##bit(values);                                           \
-    }                                                                         \
-    int##bit##_t reduce_max() const;                                          \
-    Vectorized<int##bit##_t> operator==(                                      \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vceqq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator!=(                                      \
-        const Vectorized<int##bit##_t>& other) const;                         \
-    Vectorized<int##bit##_t> operator<(                                       \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcltq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator<=(                                      \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcleq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator>(                                       \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcgtq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator>=(                                      \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcgeq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> eq(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> ne(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> gt(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> ge(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> lt(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> le(const Vectorized<int##bit##_t>& other) const; \
-  };                                                                          \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator+(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vaddq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator-(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vsubq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator&(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vandq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator|(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vorrq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator^(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return veorq_s##bit(a, b);                                                \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::eq(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this == other) & Vectorized<int##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::ne(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this != other) & Vectorized<int##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::gt(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this > other) & Vectorized<int##bit##_t>(1);                     \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::ge(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this >= other) & Vectorized<int##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::lt(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this < other) & Vectorized<int##bit##_t>(1);                     \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::le(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this <= other) & Vectorized<int##bit##_t>(1);                    \
-  }
-
-VEC_INT_NEON_TEMPLATE(2, 64)
-VEC_INT_NEON_TEMPLATE(4, 32)
-VEC_INT_NEON_TEMPLATE(8, 16)
-VEC_INT_NEON_TEMPLATE(16, 8)
-
-inline int32_t Vectorized<int32_t>::reduce_max() const {
-  return vmaxvq_s32(values);
-}
-
-inline int16_t Vectorized<int16_t>::reduce_max() const {
-  return vmaxvq_s16(values);
-}
-
-inline int8_t Vectorized<int8_t>::reduce_max() const {
-  return vmaxvq_s8(values);
-}
-
-template <>
-Vectorized<int32_t> inline operator*(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  return vmulq_s32(a, b);
-}
-
-template <>
-Vectorized<int16_t> inline operator*(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  return vmulq_s16(a, b);
-}
-
-template <>
-Vectorized<int8_t> inline operator*(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  return vmulq_s8(a, b);
-}
-
-template <>
-inline Vectorized<int64_t> operator~(const Vectorized<int64_t>& a) {
-  int64x2_t val = a;
-  return ~val;
-}
-
-template <>
-inline Vectorized<int32_t> operator~(const Vectorized<int32_t>& a) {
-  return vmvnq_s32(a);
-}
-
-template <>
-inline Vectorized<int16_t> operator~(const Vectorized<int16_t>& a) {
-  return vmvnq_s16(a);
-}
-
-template <>
-inline Vectorized<int8_t> operator~(const Vectorized<int8_t>& a) {
-  return vmvnq_s8(a);
-}
-
-inline Vectorized<int64_t> Vectorized<int64_t>::operator!=(
-    const Vectorized<int64_t>& other) const {
-  return ~(*this == other);
-}
-
-inline Vectorized<int32_t> Vectorized<int32_t>::operator!=(
-    const Vectorized<int32_t>& other) const {
-  return ~(*this == other);
-}
-
-inline Vectorized<int16_t> Vectorized<int16_t>::operator!=(
-    const Vectorized<int16_t>& other) const {
-  return ~(*this == other);
-}
-
-inline Vectorized<int8_t> Vectorized<int8_t>::operator!=(
-    const Vectorized<int8_t>& other) const {
-  return ~(*this == other);
-}
-
-template <>
-Vectorized<int32_t> inline minimum(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  return vminq_s32(a, b);
-}
-
-template <>
-Vectorized<int16_t> inline minimum(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  return vminq_s16(a, b);
-}
-
-template <>
-Vectorized<int8_t> inline minimum(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  return vminq_s8(a, b);
-}
-
-template <>
-Vectorized<int32_t> inline maximum(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  return vmaxq_s32(a, b);
-}
-
-template <>
-Vectorized<int16_t> inline maximum(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  return vmaxq_s16(a, b);
-}
-
-template <>
-Vectorized<int8_t> inline maximum(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  return vmaxq_s8(a, b);
-}
-
-template <int64_t mask>
-Vectorized<int64_t> Vectorized<int64_t>::blend(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint64x2_t maskArray = {
-      (mask & 1LL) ? 0xFFFFFFFFFFFFFFFF : 0,
-      (mask & 2LL) ? 0xFFFFFFFFFFFFFFFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s64(maskArray, b.values, a.values);
-}
-
-template <int64_t mask>
-Vectorized<int32_t> Vectorized<int32_t>::blend(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint32x4_t maskArray = {
-      (mask & 1LL) ? 0xFFFFFFFF : 0,
-      (mask & 2LL) ? 0xFFFFFFFF : 0,
-      (mask & 4LL) ? 0xFFFFFFFF : 0,
-      (mask & 8LL) ? 0xFFFFFFFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s32(maskArray, b.values, a.values);
-}
-
-template <int64_t mask>
-Vectorized<int16_t> Vectorized<int16_t>::blend(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint16x8_t maskArray = {
-      (mask & 1LL) ? 0xFFFF : 0,
-      (mask & 2LL) ? 0xFFFF : 0,
-      (mask & 4LL) ? 0xFFFF : 0,
-      (mask & 8LL) ? 0xFFFF : 0,
-      (mask & 16LL) ? 0xFFFF : 0,
-      (mask & 32LL) ? 0xFFFF : 0,
-      (mask & 64LL) ? 0xFFFF : 0,
-      (mask & 128LL) ? 0xFFFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s16(maskArray, b.values, a.values);
-}
-
-template <int64_t mask>
-Vectorized<int8_t> Vectorized<int8_t>::blend(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint8x16_t maskArray = {
-      (mask & 1LL) ? 0xFF : 0,
-      (mask & 2LL) ? 0xFF : 0,
-      (mask & 4LL) ? 0xFF : 0,
-      (mask & 8LL) ? 0xFF : 0,
-      (mask & 16LL) ? 0xFF : 0,
-      (mask & 32LL) ? 0xFF : 0,
-      (mask & 64LL) ? 0xFF : 0,
-      (mask & 128LL) ? 0xFF : 0,
-      (mask & 256LL) ? 0xFF : 0,
-      (mask & 512LL) ? 0xFF : 0,
-      (mask & 1024LL) ? 0xFF : 0,
-      (mask & 2048LL) ? 0xFF : 0,
-      (mask & 4096LL) ? 0xFF : 0,
-      (mask & 8192LL) ? 0xFF : 0,
-      (mask & 16384LL) ? 0xFF : 0,
-      (mask & 32768LL) ? 0xFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s8(maskArray, b.values, a.values);
-}
-
-#define VEC_INT_NEON_OPS(vl, bit)                                             \
-  inline Vectorized<int##bit##_t>::Vectorized(int##bit##_t val) {             \
-    values = vdupq_n_s##bit(val);                                             \
-  }                                                                           \
-  inline Vectorized<int##bit##_t> Vectorized<int##bit##_t>::loadu(            \
-      const void* ptr, int64_t count) {                                       \
-    if (count == size()) {                                                    \
-      return vld1q_s##bit(reinterpret_cast<const int##bit##_t*>(ptr));        \
-    } else {                                                                  \
-      __at_align__ int##bit##_t tmp_values[size()];                           \
-      for (const auto i : c10::irange(size())) {                              \
-        tmp_values[i] = 0;                                                    \
-      }                                                                       \
-      std::memcpy(                                                            \
-          tmp_values,                                                         \
-          reinterpret_cast<const int##bit##_t*>(ptr),                         \
-          count * sizeof(int##bit##_t));                                      \
-      return vld1q_s##bit(reinterpret_cast<const int##bit##_t*>(tmp_values)); \
-    }                                                                         \
-  }                                                                           \
-  inline void Vectorized<int##bit##_t>::store(void* ptr, int64_t count)       \
-      const {                                                                 \
-    if (count == size()) {                                                    \
-      vst1q_s##bit(reinterpret_cast<int##bit##_t*>(ptr), values);             \
-    } else {                                                                  \
-      int##bit##_t tmp_values[size()];                                        \
-      vst1q_s##bit(reinterpret_cast<int##bit##_t*>(tmp_values), values);      \
-      std::memcpy(ptr, tmp_values, count * sizeof(int##bit##_t));             \
-    }                                                                         \
-  }
-
-VEC_INT_NEON_OPS(2, 64)
-VEC_INT_NEON_OPS(4, 32)
-VEC_INT_NEON_OPS(8, 16)
-VEC_INT_NEON_OPS(16, 8)
-
-template <>
-Vectorized<int64_t> inline operator*(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return x * y;
-}
-
-template <>
-Vectorized<int64_t> inline operator/(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return x / y;
-}
-
-template <>
-Vectorized<int32_t> inline operator/(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  int32x4_t x = a;
-  int32x4_t y = b;
-  return x / y;
-}
-
-inline int64_t Vectorized<int64_t>::reduce_max() const {
-  return std::max(values[0], values[1]);
-}
-
-template <>
-Vectorized<int64_t> inline minimum(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return {std::min(x[0], y[0]), std::min(x[1], y[1])};
-}
-
-template <>
-Vectorized<int64_t> inline maximum(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return {std::max(x[0], y[0]), std::max(x[1], y[1])};
-}
-
-template <typename step_t>
-inline Vectorized<int64_t> Vectorized<int64_t>::arange(
-    int64_t base,
-    step_t step) {
-  const Vectorized<int64_t> base_vec(base);
-  const Vectorized<int64_t> step_vec(step);
-  const int64x2_t step_sizes = {0, 1};
-  return base_vec.values + step_sizes * step_vec.values;
-}
-
-template <typename step_t>
-inline Vectorized<int32_t> Vectorized<int32_t>::arange(
-    int32_t base,
-    step_t step) {
-  const Vectorized<int32_t> base_vec(base);
-  const Vectorized<int32_t> step_vec(step);
-  const int32x4_t step_sizes = {0, 1, 2, 3};
-  return vmlaq_s32(base_vec, step_sizes, step_vec);
-}
-
-template <typename step_t>
-inline Vectorized<int16_t> Vectorized<int16_t>::arange(
-    int16_t base,
-    step_t step) {
-  const Vectorized<int16_t> base_vec(base);
-  const Vectorized<int16_t> step_vec(step);
-  const int16x8_t step_sizes = {0, 1, 2, 3, 4, 5, 6, 7};
-  return vmlaq_s16(base_vec, step_sizes, step_vec);
-}
-
-template <typename step_t>
-inline Vectorized<int8_t> Vectorized<int8_t>::arange(int8_t base, step_t step) {
-  const Vectorized<int8_t> base_vec(base);
-  const Vectorized<int8_t> step_vec(step);
-  const int8x16_t step_sizes = {
-      0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
-  return vmlaq_s8(base_vec, step_sizes, step_vec);
-}
-
-template <>
-Vectorized<int64_t> inline operator>>(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  uint64x2_t u = vreinterpretq_u64_s64(y);
-  uint64x2_t z = {std::min(u[0], (uint64_t)63), std::min(u[1], (uint64_t)63)};
-  return x >> vreinterpretq_s64_u64(z);
-}
-
-template <>
-Vectorized<int32_t> inline operator>>(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  int32x4_t x = a;
-  int32x4_t y = b;
-  uint32x4_t bound = vdupq_n_u32(31);
-  uint32x4_t z = vminq_u32(vreinterpretq_u32_s32(y), bound);
-  return x >> vreinterpretq_s32_u32(z);
-}
-
-template <>
-Vectorized<int16_t> inline operator>>(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  int16x8_t x = a;
-  int16x8_t y = b;
-  uint16x8_t bound = vdupq_n_u16(15);
-  uint16x8_t z = vminq_u16(vreinterpretq_u16_s16(y), bound);
-  return x >> vreinterpretq_s16_u16(z);
-}
-
-template <>
-Vectorized<int8_t> inline operator>>(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  int8x16_t x = a;
-  int8x16_t y = b;
-  uint8x16_t bound = vdupq_n_u8(7);
-  int8x16_t z = vreinterpretq_s8_u8(vminq_u8(vreinterpretq_u8_s8(y), bound));
-  return x >> z;
-}
-
-template <>
-Vectorized<int64_t> inline operator<<(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t y = b;
-  uint64x2_t u = vreinterpretq_u64_s64(y);
-  uint64x2_t z = {std::min(u[0], (uint64_t)64), std::min(u[1], (uint64_t)64)};
-  return vshlq_s64(a, vreinterpretq_s64_u64(z));
-}
-
-template <>
-Vectorized<int32_t> inline operator<<(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  int32x4_t y = b;
-  uint32x4_t bound = vdupq_n_u32(32);
-  uint32x4_t z = vminq_u32(vreinterpretq_u32_s32(y), bound);
-  return vshlq_s32(a, vreinterpretq_s32_u32(z));
-}
-
-template <>
-Vectorized<int16_t> inline operator<<(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  int16x8_t y = b;
-  uint16x8_t bound = vdupq_n_u16(16);
-  uint16x8_t z = vminq_u16(vreinterpretq_u16_s16(y), bound);
-  return vshlq_s16(a, vreinterpretq_s16_u16(z));
-}
-
-template <>
-Vectorized<int8_t> inline operator<<(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  int8x16_t y = b;
-  uint8x16_t bound = vdupq_n_u8(8);
-  int8x16_t z = vreinterpretq_s8_u8(vminq_u8(vreinterpretq_u8_s8(y), bound));
-  return vshlq_s8(a, z);
-}
-
-inline Vectorized<int64_t> Vectorized<int64_t>::set(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 2) {
-    return b;
-  } else {
-    int64x2_t c = {b.values[0], a.values[1]};
-    return c;
-  }
-}
-
-inline Vectorized<int32_t> Vectorized<int32_t>::set(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 4) {
-    return b;
-  } else {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint32x4_t maskArray = {
-        (count >= 1LL) ? 0xFFFFFFFF : 0,
-        (count >= 2LL) ? 0xFFFFFFFF : 0,
-        (count >= 3LL) ? 0xFFFFFFFF : 0,
-        0};
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_s32(maskArray, b.values, a.values);
-  }
-}
-
-inline Vectorized<int16_t> Vectorized<int16_t>::set(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 8) {
-    return b;
-  } else {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint16x8_t maskArray = {
-        static_cast<uint16_t>((count >= 1LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 2LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 3LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 4LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 5LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 6LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 7LL) ? 0xFFFF : 0),
-        0};
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_s16(maskArray, b.values, a.values);
-  }
-}
-
-inline Vectorized<int8_t> Vectorized<int8_t>::set(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 16) {
-    return b;
-  } else {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint8x16_t maskArray = {
-        static_cast<uint8_t>((count >= 1LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 2LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 3LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 4LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 5LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 6LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 7LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 8LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 9LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 10LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 11LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 12LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 13LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 14LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 15LL) ? 0xFF : 0),
-        0};
-
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_s8(maskArray, b.values, a.values);
-  }
-}
-
-template <>
-Vectorized<int16_t> inline operator/(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  Vectorized<int32_t> highBitsA = vmovl_high_s16(a);
-  Vectorized<int32_t> highBitsB = vmovl_high_s16(b);
-  Vectorized<int32_t> lowBitsA = vmovl_s16(vget_low_s16(a));
-  Vectorized<int32_t> lowBitsB = vmovl_s16(vget_low_s16(b));
-  int32x4_t highBitsResult = highBitsA / highBitsB;
-  int32x4_t lowBitsResult = lowBitsA / lowBitsB;
-  return vuzp1q_s16(
-      vreinterpretq_s16_s32(lowBitsResult),
-      vreinterpretq_s16_s32(highBitsResult));
-}
-
-template <>
-Vectorized<int8_t> inline operator/(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  Vectorized<int16_t> highBitsA = vmovl_high_s8(a);
-  Vectorized<int16_t> highBitsB = vmovl_high_s8(b);
-  Vectorized<int16_t> lowBitsA = vmovl_s8(vget_low_s8(a));
-  Vectorized<int16_t> lowBitsB = vmovl_s8(vget_low_s8(b));
-  int16x8_t highBitsResult = highBitsA / highBitsB;
-  int16x8_t lowBitsResult = lowBitsA / lowBitsB;
-  return vuzp1q_s8(
-      vreinterpretq_s8_s16(lowBitsResult),
-      vreinterpretq_s8_s16(highBitsResult));
-}
-
-template <>
-Vectorized<int64_t> inline clamp(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& min,
-    const Vectorized<int64_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int32_t> inline clamp(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& min,
-    const Vectorized<int32_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int16_t> inline clamp(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& min,
-    const Vectorized<int16_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int8_t> inline clamp(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& min,
-    const Vectorized<int8_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int64_t> inline clamp_max(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int32_t> inline clamp_max(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int16_t> inline clamp_max(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int8_t> inline clamp_max(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int64_t> inline clamp_min(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& min) {
-  return maximum(min, a);
-}
-
-template <>
-Vectorized<int32_t> inline clamp_min(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& min) {
-  return maximum(min, a);
-}
-
-template <>
-Vectorized<int16_t> inline clamp_min(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& min) {
-  return maximum(min, a);
-}
-
-template <>
-Vectorized<int8_t> inline clamp_min(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& min) {
-  return maximum(min, a);
-}
-
-} // namespace CPU_CAPABILITY
-} // namespace at::vec

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) {
-  auto s8x8 = vget_low_s8(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) {
-  auto s8x8 = vget_low_s8(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

@@ -16,8 +16,6 @@
 #include <c10/util/irange.h>
 #include <c10/core/ScalarType.h>
 
-#include <ATen/cuda/detail/BLASConstants.h>
-
 #ifdef USE_ROCM
 #include <c10/cuda/CUDAStream.h>
 #include <hipblaslt/hipblaslt-ext.hpp>
@@ -110,7 +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':
@@ -130,7 +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).
@@ -144,7 +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,
@@ -1956,15 +1954,13 @@ void scaled_gemm(
     const void *result_scale_ptr,
     int64_t result_ld,
     ScalarType result_dtype,
-    bool use_fast_accum,
-    const std::optional<Tensor>& alpha) {
+    bool use_fast_accum) {
   // Note: see `cublasCommonArgs` for various non-intuitive manupulations
   // of input arguments to this function.
   const auto computeType = CUBLAS_COMPUTE_32F;
   const auto scaleType = CUDA_R_32F;
-  // Note: alpha_val may change later depending on user-passed argument
-  float alpha_val = 1.0;
-  float beta_val = 0.0;
+  const float alpha_val = 1.0;
+  const float beta_val = 0.0;
   CuBlasLtMatmulDescriptor computeDesc(computeType, scaleType);
   computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_TRANSA, _cublasOpFromChar(transa));
   computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_TRANSB, _cublasOpFromChar(transb));
@@ -2035,33 +2031,6 @@ void scaled_gemm(
     computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_EPILOGUE, CUBLASLT_EPILOGUE_BIAS);
     computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE, ScalarTypeToCudaDataType(bias_dtype));
   }
-
-  // Handle user-passed alpha
-  float *alpha_ptr = &alpha_val;
-  float *beta_ptr = &beta_val;
-
-  if (alpha.has_value()) {
-    auto& a = alpha.value();
-
-    // if device-tensor
-    if (a.is_cuda()) {
-      // NOTE: there are lifetime requirements on device-side pointers for alpha/beta -- the value must be
-      //       valid & correct until the cublas call finishes (not is scheduled like host-side values). Thus
-      //       we need to use allocations for alpha/beta that have some guarantees on lifetime - a statically
-      //       managed 4B buffer for alpha that we'll copy the passed alpha value into, and constant memory
-      //       for beta respectively.
-      float *user_alpha_ptr = at::cuda::detail::get_user_alpha_ptr();
-      at::Tensor user_alpha = at::from_blob(user_alpha_ptr, {1}, TensorOptions().device(kCUDA).dtype(kFloat));
-      user_alpha.copy_(a);
-      // Tell cublasLt we're using device-side pointers for alpha/beta
-      auto pointer_mode = CUBLASLT_POINTER_MODE_DEVICE;
-      computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_POINTER_MODE, pointer_mode);
-      alpha_ptr = user_alpha.data_ptr<float>();
-      beta_ptr = at::cuda::detail::get_cublas_device_zero();
-    } else {
-      alpha_val = a.item<float>();
-    }
-  }
     // For other data types, use the get_scale_mode function based on scaling type
     // The SCALE_MODE attrs only exist in cuBLAS 12.8+/ROCm 7.0 or in recent hipblaslt,
     // but we must invoke get_scale_mode anyways to trigger the version checks.
@@ -2079,7 +2048,6 @@ void scaled_gemm(
   cublasLtMatmulHeuristicResult_t heuristicResult = {};
   int returnedResult = 0;
   cublasLtHandle_t ltHandle = at::cuda::getCurrentCUDABlasLtHandle();
-
   TORCH_CUDABLAS_CHECK(cublasLtMatmulAlgoGetHeuristic(
       ltHandle,
       computeDesc.descriptor(),
@@ -2120,10 +2088,10 @@ void scaled_gemm(
         auto is_valid_status = hipblaslt_ext::matmulIsAlgoSupported(
                 ltHandle,
                 computeDesc.descriptor(),
-                alpha_ptr,
+                &alpha_val,
                 Adesc.descriptor(),
                 Bdesc.descriptor(),
-                beta_ptr,
+                &beta_val,
                 Cdesc.descriptor(),
                 Ddesc.descriptor(),
                 all_algos[i].algo,
@@ -2142,14 +2110,17 @@ void scaled_gemm(
   cublasStatus_t cublasStatus = cublasLtMatmul(
       ltHandle,
       computeDesc.descriptor(),
-      alpha_ptr,
+      &alpha_val,
       mat1_ptr,
       Adesc.descriptor(),
       mat2_ptr,
       Bdesc.descriptor(),
-      beta_ptr,
-      // NOTE: always use result_ptr here, because cuBLASLt w/device beta=0 can't handle nullptr either
+      &beta_val,
+#ifdef USE_ROCM
       result_ptr, // unused, since beta_val is 0, but hipblaslt can't handle nullptr
+#else
+      nullptr,
+#endif // ifdef USE_ROCM
       Cdesc.descriptor(),
       result_ptr,
       Ddesc.descriptor(),

aten/src/ATen/cuda/CUDABlas.h

@@ -161,8 +161,7 @@ void scaled_gemm(
     const void* result_scale_ptr,
     int64_t result_ld,
     ScalarType result_dtype,
-    bool use_fast_accum,
-    const std::optional<Tensor>& alpha);
+    bool use_fast_accum);
 
 #define CUDABLAS_BGEMM_ARGTYPES(Dtype)  CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(Dtype, Dtype)
 

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());
@@ -325,9 +325,9 @@ uint64_t CUDAGeneratorImpl::seed() {
  */
 c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
   // The RNG state comprises the seed, and an offset used for Philox.
-  constexpr size_t seed_size = sizeof(uint64_t);
-  constexpr size_t offset_size = sizeof(int64_t);
-  constexpr size_t total_size = seed_size + offset_size;
+  static const size_t seed_size = sizeof(uint64_t);
+  static const size_t offset_size = sizeof(int64_t);
+  static const size_t total_size = seed_size + offset_size;
 
   auto state_tensor = at::detail::empty_cpu({(int64_t)total_size}, ScalarType::Byte, std::nullopt, std::nullopt, std::nullopt, std::nullopt);
   auto rng_state = state_tensor.data_ptr<uint8_t>();
@@ -346,9 +346,9 @@ c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
  * and size of the internal state.
  */
 void CUDAGeneratorImpl::set_state(const c10::TensorImpl& new_state) {
-  constexpr size_t seed_size = sizeof(uint64_t);
-  constexpr size_t offset_size = sizeof(int64_t);
-  constexpr size_t total_size = seed_size + offset_size;
+  static const size_t seed_size = sizeof(uint64_t);
+  static const size_t offset_size = sizeof(int64_t);
+  static const size_t total_size = seed_size + offset_size;
 
   detail::check_rng_state(new_state);
 

aten/src/ATen/cuda/CachingHostAllocator.cpp

@@ -183,6 +183,11 @@ struct CUDACachingHostAllocatorImpl
     return true;
   }
 
+  bool pinned_use_background_threads() override {
+    return c10::cuda::CUDACachingAllocator::CUDAAllocatorConfig::
+        pinned_use_background_threads();
+  }
+
   EventPool::Event create_event_internal(DeviceIndex idx) {
     // Leak the event pool to avoid shutdown issue.
     static auto* event_pool = new EventPool();

aten/src/ATen/cuda/cub.cuh

@@ -177,6 +177,7 @@ inline void segmented_sort_pairs(
   }
 }
 
+#if CUB_SUPPORTS_UNIQUE_BY_KEY()
 template <typename KeysInputIteratorT, typename ValuesInputIteratorT, typename ValuesOutputIteratorT, typename NumSelectedIteratorT>
 inline void unique_by_key(
   KeysInputIteratorT keys_in, ValuesInputIteratorT values_in,
@@ -192,6 +193,7 @@ inline void unique_by_key(
   CUB_WRAPPER(NO_ROCM(at_cuda_detail)::cub::DeviceSelect::UniqueByKey,
     keys_in, values_in, keys_out_, values_out, num_selected, num_input_items, c10::cuda::getCurrentCUDAStream());
 }
+#endif
 
 namespace impl {
 
@@ -577,6 +579,7 @@ inline void exclusive_scan(InputIteratorT input, OutputIteratorT output, ScanOpT
 #endif
 }
 
+#if CUB_SUPPORTS_SCAN_BY_KEY()
 
 template <typename KeysInputIteratorT, typename ValuesInputIteratorT, typename ValuesOutputIteratorT>
 inline void inclusive_sum_by_key(KeysInputIteratorT keys, ValuesInputIteratorT input, ValuesOutputIteratorT output, int64_t num_items) {
@@ -604,6 +607,7 @@ inline void inclusive_scan_by_key(KeysInputIteratorT keys, ValuesInputIteratorT
 #endif
 }
 
+#endif
 
 template <typename InputIteratorT, typename OutputIteratorT, typename NumSelectedIteratorT>
 void unique(InputIteratorT input, OutputIteratorT output,