add distributed and fsdp owner labels

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -31,7 +31,8 @@ pip install -r /pytorch/requirements.txt
 pip install auditwheel==6.2.0 wheel
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
     export USE_SYSTEM_NCCL=1
@@ -45,5 +46,6 @@ else
         export USE_NVIDIA_PYPI_LIBS=1
     fi
 
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -317,7 +317,7 @@ if __name__ == "__main__":
     ).decode()
 
     print("Building PyTorch wheel")
-    build_vars = ""
+    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
     # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
     if enable_cuda:
         build_vars += "MAX_JOBS=5 "

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -241,7 +241,7 @@ def wait_for_connection(addr, port, timeout=15, attempt_cnt=5):
         try:
             with socket.create_connection((addr, port), timeout=timeout):
                 return
-        except (ConnectionRefusedError, TimeoutError):  # noqa: PERF203
+        except (ConnectionRefusedError, socket.timeout):  # noqa: PERF203
             if i == attempt_cnt - 1:
                 raise
             time.sleep(timeout)
@@ -1004,7 +1004,7 @@ if __name__ == "__main__":
         install_condaforge_python(host, args.python_version)
         sys.exit(0)
 
-    python_version = args.python_version if args.python_version is not None else "3.10"
+    python_version = args.python_version if args.python_version is not None else "3.9"
 
     if args.use_torch_from_pypi:
         configure_system(host, compiler=args.compiler, python_version=python_version)

.ci/docker/almalinux/Dockerfile

@@ -69,8 +69,7 @@ RUN bash ./install_cuda.sh 13.0
 ENV DESIRED_CUDA=13.0
 
 FROM ${ROCM_IMAGE} as rocm
-ARG PYTORCH_ROCM_ARCH
-ENV PYTORCH_ROCM_ARCH ${PYTORCH_ROCM_ARCH}
+ENV PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
 ADD ./common/install_mkl.sh install_mkl.sh
 RUN bash ./install_mkl.sh && rm install_mkl.sh
 ENV MKLROOT /opt/intel

.ci/docker/almalinux/build.sh

@@ -36,12 +36,6 @@ case ${DOCKER_TAG_PREFIX} in
     ;;
   rocm*)
     BASE_TARGET=rocm
-    PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
-    # add gfx950 conditionally starting in ROCm 7.0
-    if [[ "$ROCM_VERSION" == *"7.0"* ]]; then
-        PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
-    fi
-    EXTRA_BUILD_ARGS="${EXTRA_BUILD_ARGS} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}"
     ;;
   *)
     echo "ERROR: Unknown docker tag ${DOCKER_TAG_PREFIX}"

.ci/docker/build.sh

@@ -262,10 +262,13 @@ case "$tag" in
     TRITON_CPU=yes
     ;;
   pytorch-linux-jammy-linter)
-    PYTHON_VERSION=3.10
+    # TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
+    # We will need to update mypy version eventually, but that's for another day. The task
+    # would be to upgrade mypy to 1.0.0 with Python 3.11
+    PYTHON_VERSION=3.9
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter)
-    PYTHON_VERSION=3.10
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
+    PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
     ;;
   pytorch-linux-jammy-aarch64-py3.10-gcc11)

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-e0dda9059d082537cee36be6c5e4fe3b18c880c0
+56392aa978594cc155fa8af48cd949f5b5f1823a

.ci/docker/ci_commit_pins/huggingface-requirements.txt

@@ -1,2 +1,2 @@
-transformers==4.56.0
+transformers==4.54.0
 soxr==0.5.0

.ci/docker/ci_commit_pins/nccl-cu12.txt

@@ -1 +1 @@
-v2.28.3-1
+v2.27.5-1

.ci/docker/ci_commit_pins/nccl-cu13.txt

@@ -1 +1 @@
-v2.28.3-1
+v2.27.7-1

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-bbb06c0334a6772b92d24bde54956e675c8c6604
+5ae38bdb0dc066c5823e34dc9797afb9de42c866

.ci/docker/common/install_executorch.sh

@@ -42,27 +42,22 @@ install_pip_dependencies() {
   # A workaround, ExecuTorch has moved to numpy 2.0 which is not compatible with the current
   # numba and scipy version used in PyTorch CI
   conda_run pip uninstall -y numba scipy
-  # Yaspin is needed for running CI test (get_benchmark_analysis_data.py)
-  pip_install yaspin==3.1.0
 
   popd
 }
 
 setup_executorch() {
+  pushd executorch
+
   export PYTHON_EXECUTABLE=python
-  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON -DEXECUTORCH_BUILD_TESTS=ON"
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
 
   as_jenkins .ci/scripts/setup-linux.sh --build-tool cmake || true
+  popd
 }
 
-if [ $# -eq 0 ]; then
-  clone_executorch
-  install_buck2
-  install_conda_dependencies
-  install_pip_dependencies
-  pushd executorch
-  setup_executorch
-  popd
-else
-  "$@"
-fi
+clone_executorch
+install_buck2
+install_conda_dependencies
+install_pip_dependencies
+setup_executorch

.ci/docker/common/install_rocm_magma.sh

@@ -12,8 +12,8 @@ function do_install() {
 
     rocm_version_nodot=${rocm_version//./}
 
-    # https://github.com/icl-utk-edu/magma/pull/65
-    MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
+    # Version 2.7.2 + ROCm related updates
+    MAGMA_VERSION=a1625ff4d9bc362906bd01f805dbbe12612953f6
     magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"
 
     rocm_dir="/opt/rocm"

.ci/docker/libtorch/build.sh

@@ -40,16 +40,12 @@ case ${DOCKER_TAG_PREFIX} in
         ;;
     rocm*)
         # we want the patch version of 6.4 instead
-        if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
+        if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
             GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete
         PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
-        # add gfx950 conditionally starting in ROCm 7.0
-        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
-        fi
         DOCKER_GPU_BUILD_ARG="--build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg ROCM_VERSION=${GPU_ARCH_VERSION}"
         ;;
     *)

.ci/docker/manywheel/build.sh

@@ -82,7 +82,7 @@ case ${image} in
         ;;
     manylinux2_28-builder:rocm*)
         # we want the patch version of 6.4 instead
-        if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
+        if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
             GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
         fi
         TARGET=rocm_final
@@ -90,10 +90,6 @@ case ${image} in
         DEVTOOLSET_VERSION="11"
         GPU_IMAGE=rocm/dev-almalinux-8:${GPU_ARCH_VERSION}-complete
         PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
-        # add gfx950 conditionally starting in ROCm 7.0
-        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
-        fi
         DOCKER_GPU_BUILD_ARG="--build-arg ROCM_VERSION=${GPU_ARCH_VERSION} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}"
         ;;
     manylinux2_28-builder:xpu)

.ci/docker/requirements-ci.txt

@@ -93,9 +93,8 @@ librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
 #Pinned versions:
 #test that import:
 
-mypy==1.16.0 ; platform_system != "Windows"
+mypy==1.16.0
 # Pin MyPy version because new errors are likely to appear with each release
-# Skip on Windows as lots of type annotations are POSIX specific
 #Description: linter
 #Pinned versions: 1.16.0
 #test that import: test_typing.py, test_type_hints.py
@@ -112,6 +111,8 @@ ninja==1.11.1.3
 #Pinned versions: 1.11.1.3
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
 
+numba==0.49.0 ; python_version < "3.9" and platform_machine != "s390x"
+numba==0.55.2 ; python_version == "3.9" and platform_machine != "s390x"
 numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
 numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
 #Description: Just-In-Time Compiler for Numerical Functions
@@ -132,7 +133,7 @@ numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
 #test_nn.py, test_namedtensor.py, test_linalg.py, test_jit_cuda_fuser.py,
 #test_jit.py, test_indexing.py, test_datapipe.py, test_dataloader.py,
 #test_binary_ufuncs.py
-numpy==1.22.4; python_version == "3.10"
+numpy==1.22.4; python_version == "3.9" or python_version == "3.10"
 numpy==1.26.2; python_version == "3.11" or python_version == "3.12"
 numpy==2.1.2; python_version >= "3.13"
 
@@ -324,6 +325,8 @@ pywavelets==1.7.0 ; python_version >= "3.12"
 lxml==5.3.0
 #Description: This is a requirement of unittest-xml-reporting
 
+# Python-3.9 binaries
+
 PyGithub==2.3.0
 
 sympy==1.13.3

.ci/docker/requirements-docs.txt

@@ -1,15 +1,8 @@
 sphinx==5.3.0
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 5.3.0
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@1657ad2fc1acdc98aa719eebecbb0128a7c13ce4#egg=pytorch_sphinx_theme2
 
-standard-imghdr==3.13.0; python_version >= "3.13"
-#Description: This is needed by Sphinx, so it needs to be added here.
-# The reasons are as follows:
-# 1) This module has been removed from the Python standard library since Python 3.13(https://peps.python.org/pep-0594/#imghdr);
-# 2) The current version of Sphinx (5.3.0) is not compatible with Python 3.13.
-# Once Sphinx is upgraded to a version compatible with Python 3.13 or later, we can remove this dependency.
-
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git@d53b0ffb9b1cda68260693ea98f3483823c88d8e#egg=pytorch_sphinx_theme2
 # TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
 # but it doesn't seem to work and hangs around idly. The initial thought that it is probably
 # something related to Docker setup. We can investigate this later.

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

@@ -41,6 +41,7 @@ def sample_vllm_test_library():
                 "pytest -v -s basic_correctness/test_cumem.py",
                 "pytest -v -s basic_correctness/test_basic_correctness.py",
                 "pytest -v -s basic_correctness/test_cpu_offload.py",
+                "VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py",
             ],
         },
         "vllm_basic_models_test": {
@@ -67,12 +68,15 @@ def sample_vllm_test_library():
                         "-v",
                         "-s",
                         "entrypoints/llm",
+                        "--ignore=entrypoints/llm/test_lazy_outlines.py",
                         "--ignore=entrypoints/llm/test_generate.py",
+                        "--ignore=entrypoints/llm/test_generate_multiple_loras.py",
                         "--ignore=entrypoints/llm/test_collective_rpc.py",
                     ]
                 ),
-                "pytest -v -s entrypoints/llm/test_generate.py",
-                "pytest -v -s entrypoints/offline_mode",
+                "pytest -v -s entrypoints/llm/test_lazy_outlines.py",
+                "pytest -v -s entrypoints/llm/test_generate.py ",
+                "VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode",
             ],
         },
         "vllm_regression_test": {

.ci/magma-rocm/Makefile

@@ -1,11 +1,11 @@
 SHELL=/usr/bin/env bash
 
 DOCKER_CMD ?= docker
-DESIRED_ROCM ?= 7.0
+DESIRED_ROCM ?= 6.4
 DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
 PACKAGE_NAME = magma-rocm
 # inherit this from underlying docker image, do not pass this env var to docker
-#PYTORCH_ROCM_ARCH ?= gfx900;gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201
+#PYTORCH_ROCM_ARCH ?= gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201
 
 DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	-v $(shell git rev-parse --show-toplevel)/.ci:/builder \
@@ -16,7 +16,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma-rocm/build_magma.sh
 
 .PHONY: all
-all: magma-rocm70
 all: magma-rocm64
 all: magma-rocm63
 
@@ -25,11 +24,6 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
-.PHONY: magma-rocm70
-magma-rocm70: DESIRED_ROCM := 7.0
-magma-rocm70:
-	$(DOCKER_RUN)
-
 .PHONY: magma-rocm64
 magma-rocm64: DESIRED_ROCM := 6.4
 magma-rocm64:

.ci/magma-rocm/build_magma.sh

@@ -6,8 +6,8 @@ set -eou pipefail
 # The script expects DESIRED_CUDA and PACKAGE_NAME to be set
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 
-# https://github.com/icl-utk-edu/magma/pull/65
-MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
+# Version 2.7.2 + ROCm related updates
+MAGMA_VERSION=a1625ff4d9bc362906bd01f805dbbe12612953f6
 
 # Folders for the build
 PACKAGE_FILES=${ROOT_DIR}/magma-rocm/package_files # metadata
@@ -20,7 +20,7 @@ mkdir -p ${PACKAGE_DIR} ${PACKAGE_OUTPUT}/linux-64 ${PACKAGE_BUILD} ${PACKAGE_RE
 
 # Fetch magma sources and verify checksum
 pushd ${PACKAGE_DIR}
-git clone https://github.com/jeffdaily/magma
+git clone https://bitbucket.org/icl/magma.git
 pushd magma
 git checkout ${MAGMA_VERSION}
 popd

.ci/manywheel/build_libtorch.sh

@@ -104,7 +104,7 @@ if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
     export ROCclr_DIR=/opt/rocm/rocclr/lib/cmake/rocclr
 fi
 
-echo "Calling -m pip install . -v --no-build-isolation at $(date)"
+echo "Calling 'python -m pip install .' at $(date)"
 
 if [[ $LIBTORCH_VARIANT = *"static"* ]]; then
     STATIC_CMAKE_FLAG="-DTORCH_STATIC=1"

.ci/pytorch/cpp_doc_push_script.sh

@@ -58,7 +58,7 @@ time python tools/setup_helpers/generate_code.py \
 
 # Build the docs
 pushd docs/cpp
-time make VERBOSE=1 html
+time make VERBOSE=1 html -j
 
 popd
 popd

.ci/pytorch/macos-build.sh

@@ -35,11 +35,10 @@ fi
 
 print_cmake_info
 if [[ ${BUILD_ENVIRONMENT} == *"distributed"* ]]; then
-  # Needed for inductor benchmarks, as lots of HF networks make `torch.distribtued` calls
-  USE_DISTRIBUTED=1 USE_OPENMP=1 WERROR=1 python setup.py bdist_wheel
+  USE_OPENMP=1 WERROR=1 python setup.py bdist_wheel
 else
-  # Explicitly set USE_DISTRIBUTED=0 to align with the default build config on mac. This also serves as the sole CI config that tests
-  # that building with USE_DISTRIBUTED=0 works at all. See https://github.com/pytorch/pytorch/issues/86448
+  # NB: we always build with distributed; USE_DISTRIBUTED turns off all
+  # backends (specifically the gloo backend), so test that this case works too
   USE_DISTRIBUTED=0 USE_OPENMP=1 MACOSX_DEPLOYMENT_TARGET=11.0 WERROR=1 BUILD_TEST=OFF USE_PYTORCH_METAL=1 python setup.py bdist_wheel --plat-name macosx_11_0_arm64
 fi
 if which sccache > /dev/null; then

.ci/pytorch/macos-test.sh

@@ -13,9 +13,13 @@ if [[ ! $(python -c "import torch; print(int(torch.backends.openmp.is_available(
 fi
 popd
 
+python -mpip install -r requirements.txt
+
 # enable debug asserts in serialization
 export TORCH_SERIALIZATION_DEBUG=1
 
+python -mpip install --no-input -r requirements.txt
+
 setup_test_python() {
   # The CircleCI worker hostname doesn't resolve to an address.
   # This environment variable makes ProcessGroupGloo default to
@@ -55,7 +59,7 @@ test_python_shard() {
 
   setup_test_python
 
-  time python test/run_test.py --verbose --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests --shard "$1" "$NUM_TEST_SHARDS"
+  time python test/run_test.py --verbose --exclude-jit-executor --exclude-distributed-tests --shard "$1" "$NUM_TEST_SHARDS"
 
   assert_git_not_dirty
 }

.ci/pytorch/test.sh

@@ -322,29 +322,23 @@ test_python_shard() {
 
   # modify LD_LIBRARY_PATH to ensure it has the conda env.
   # This set of tests has been shown to be buggy without it for the split-build
-  time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests $INCLUDE_CLAUSE --shard "$1" "$NUM_TEST_SHARDS" --verbose $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests $INCLUDE_CLAUSE --shard "$1" "$NUM_TEST_SHARDS" --verbose $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
 
   assert_git_not_dirty
 }
 
 test_python() {
   # shellcheck disable=SC2086
-  time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests $INCLUDE_CLAUSE --verbose $PYTHON_TEST_EXTRA_OPTION
+  time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests $INCLUDE_CLAUSE --verbose $PYTHON_TEST_EXTRA_OPTION
   assert_git_not_dirty
 }
 
 test_python_smoke() {
-  # Smoke tests for H100/B200
+  # Smoke tests for H100
   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 inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
 test_h100_distributed() {
   # Distributed tests at H100
   time python test/run_test.py --include distributed/_composable/test_composability/test_pp_composability.py  $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
@@ -390,7 +384,6 @@ test_dynamo_wrapped_shard() {
     --exclude-distributed-tests \
     --exclude-torch-export-tests \
     --exclude-aot-dispatch-tests \
-    --exclude-quantization-tests \
     --shard "$1" "$NUM_TEST_SHARDS" \
     --verbose \
     --upload-artifacts-while-running
@@ -1163,12 +1156,6 @@ test_distributed() {
   fi
 }
 
-test_quantization() {
-  echo "Testing quantization"
-
-  python test/test_quantization.py
-}
-
 test_rpc() {
   echo "Testing RPC C++ tests"
   # NB: the ending test_rpc must match the current function name for the current
@@ -1563,10 +1550,14 @@ test_executorch() {
   install_torchvision
   install_torchaudio
 
-  INSTALL_SCRIPT="$(pwd)/.ci/docker/common/install_executorch.sh"
-
   pushd /executorch
-  "${INSTALL_SCRIPT}" setup_executorch
+
+  export PYTHON_EXECUTABLE=python
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
+
+  # NB: We need to rebuild ExecuTorch runner here because it depends on PyTorch
+  # from the PR
+  bash .ci/scripts/setup-linux.sh --build-tool cmake
 
   echo "Run ExecuTorch unit tests"
   pytest -v -n auto
@@ -1580,13 +1571,17 @@ test_executorch() {
 
   popd
 
+  # Test torchgen generated code for Executorch.
+  echo "Testing ExecuTorch op registration"
+  "$BUILD_BIN_DIR"/test_edge_op_registration
+
   assert_git_not_dirty
 }
 
 test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
-        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops profiler/test_memory_profiler \
+        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
         distributed/elastic/timer/api_test distributed/elastic/timer/local_timer_example distributed/elastic/timer/local_timer_test \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
@@ -1617,7 +1612,7 @@ test_operator_benchmark() {
   test_inductor_set_cpu_affinity
 
   cd benchmarks/operator_benchmark/pt_extension
-  python -m pip install . -v --no-build-isolation
+  python -m pip install .
 
   cd "${TEST_DIR}"/benchmarks/operator_benchmark
   $TASKSET python -m benchmark_all_test --device "$1" --tag-filter "$2" \
@@ -1630,25 +1625,6 @@ test_operator_benchmark() {
       --expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
 }
 
-test_operator_microbenchmark() {
-  TEST_REPORTS_DIR=$(pwd)/test/test-reports
-  mkdir -p "$TEST_REPORTS_DIR"
-  TEST_DIR=$(pwd)
-
-  cd benchmarks/operator_benchmark/pt_extension
-  python -m pip install .
-
-  cd "${TEST_DIR}"/benchmarks/operator_benchmark
-
-  for OP_BENCHMARK_TESTS in matmul mm addmm bmm; do
-    $TASKSET python -m pt.${OP_BENCHMARK_TESTS}_test --tag-filter long \
-      --output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_microbenchmark_${OP_BENCHMARK_TESTS}_compile.json" \
-      --benchmark-name "PyTorch operator microbenchmark" --use-compile
-    $TASKSET python -m pt.${OP_BENCHMARK_TESTS}_test --tag-filter long \
-      --output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_microbenchmark_${OP_BENCHMARK_TESTS}.json" \
-      --benchmark-name "PyTorch operator microbenchmark"
-  done
-}
 
 if ! [[ "${BUILD_ENVIRONMENT}" == *libtorch* || "${BUILD_ENVIRONMENT}" == *-bazel-* ]]; then
   (cd test && python -c "import torch; print(torch.__config__.show())")
@@ -1681,8 +1657,6 @@ elif [[ "${TEST_CONFIG}" == *executorch* ]]; then
   test_executorch
 elif [[ "$TEST_CONFIG" == 'jit_legacy' ]]; then
   test_python_legacy_jit
-elif [[ "$TEST_CONFIG" == 'quantization' ]]; then
-  test_quantization
 elif [[ "${BUILD_ENVIRONMENT}" == *libtorch* ]]; then
   # TODO: run some C++ tests
   echo "no-op at the moment"
@@ -1705,8 +1679,6 @@ elif [[ "${TEST_CONFIG}" == *operator_benchmark* ]]; then
     test_operator_benchmark cpu ${TEST_MODE}
 
   fi
-elif [[ "${TEST_CONFIG}" == *operator_microbenchmark* ]]; then
-  test_operator_microbenchmark
 elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
   test_inductor_distributed
 elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
@@ -1809,14 +1781,10 @@ elif [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
   test_xpu_bin
 elif [[ "${TEST_CONFIG}" == smoke ]]; then
   test_python_smoke
-elif [[ "${TEST_CONFIG}" == smoke_b200 ]]; then
-  test_python_smoke_b200
 elif [[ "${TEST_CONFIG}" == h100_distributed ]]; then
   test_h100_distributed
 elif [[ "${TEST_CONFIG}" == "h100-symm-mem" ]]; then
   test_h100_symm_mem
-elif [[ "${TEST_CONFIG}" == "b200-symm-mem" ]]; then
-  test_h100_symm_mem
 elif [[ "${TEST_CONFIG}" == h100_cutlass_backend ]]; then
   test_h100_cutlass_backend
 else

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

@@ -137,7 +137,7 @@ sccache --show-stats
 python -c "import os, glob; os.system('python -mpip install --no-index --no-deps ' + glob.glob('dist/*.whl')[0])"
 (
   if "%BUILD_ENVIRONMENT%"=="" (
-    echo NOTE: To run `import torch`, please make sure to activate the conda environment by running `call %CONDA_ROOT_DIR%\Scripts\activate.bat %CONDA_ROOT_DIR%\envs\py_tmp` in Command Prompt before running Git Bash.
+    echo NOTE: To run `import torch`, please make sure to activate the conda environment by running `call %CONDA_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3` in Command Prompt before running Git Bash.
   ) else (
     copy /Y "dist\*.whl" "%PYTORCH_FINAL_PACKAGE_DIR%"
 

.ci/pytorch/win-test-helpers/installation-helpers/activate_miniconda3.bat

@@ -3,12 +3,12 @@ if "%BUILD_ENVIRONMENT%"=="" (
 ) else (
   set CONDA_PARENT_DIR=C:\Jenkins
 )
-set CONDA_ROOT_DIR=%CONDA_PARENT_DIR%\Miniconda3
+
 
 :: Be conservative here when rolling out the new AMI with conda. This will try
 :: to install conda as before if it couldn't find the conda installation. This
 :: can be removed eventually after we gain enough confidence in the AMI
-if not exist %CONDA_ROOT_DIR% (
+if not exist %CONDA_PARENT_DIR%\Miniconda3 (
   set INSTALL_FRESH_CONDA=1
 )
 
@@ -17,14 +17,10 @@ if "%INSTALL_FRESH_CONDA%"=="1" (
   if errorlevel 1 exit /b
   if not errorlevel 0 exit /b
 
-  %TMP_DIR_WIN%\Miniconda3-latest-Windows-x86_64.exe /InstallationType=JustMe /RegisterPython=0 /S /AddToPath=0 /D=%CONDA_ROOT_DIR%
+  %TMP_DIR_WIN%\Miniconda3-latest-Windows-x86_64.exe /InstallationType=JustMe /RegisterPython=0 /S /AddToPath=0 /D=%CONDA_PARENT_DIR%\Miniconda3
   if errorlevel 1 exit /b
   if not errorlevel 0 exit /b
 )
 
 :: Activate conda so that we can use its commands, i.e. conda, python, pip
-call %CONDA_ROOT_DIR%\Scripts\activate.bat %CONDA_ROOT_DIR%
-:: Activate conda so that we can use its commands, i.e. conda, python, pip
-call conda activate py_tmp
-
-call pip install -r .ci/docker/requirements-ci.txt
+call %CONDA_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3

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

@@ -14,7 +14,7 @@ if not errorlevel 0 exit /b
 :: build\torch. Rather than changing all these references, making a copy of torch folder
 :: from conda to the current workspace is easier. The workspace will be cleaned up after
 :: the job anyway
-xcopy /s %CONDA_ROOT_DIR%\envs\py_tmp\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
+xcopy /s %CONDA_PARENT_DIR%\Miniconda3\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
 
 pushd .
 if "%VC_VERSION%" == "" (

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

@@ -25,7 +25,7 @@ echo Copying over test times file
 robocopy /E "%PYTORCH_FINAL_PACKAGE_DIR_WIN%\.additional_ci_files" "%PROJECT_DIR_WIN%\.additional_ci_files"
 
 echo Run nn tests
-python run_test.py --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests --shard "%SHARD_NUMBER%" "%NUM_TEST_SHARDS%" --verbose
+python run_test.py --exclude-jit-executor --exclude-distributed-tests --shard "%SHARD_NUMBER%" "%NUM_TEST_SHARDS%" --verbose
 if ERRORLEVEL 1 goto fail
 
 popd

.ci/pytorch/win-test.sh

@@ -38,14 +38,7 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
 fi
 
 # TODO: Move both of them to Windows AMI
-python -m pip install tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
-
-# Copied from https://github.com/pytorch/test-infra/blob/be01a40157c36cd5a48391fdf44a7bc3ebd4c7e3/aws/ami/windows/scripts/Installers/Install-Pip-Dependencies.ps1#L16 with some adjustments
-# pytest-rerunfailures==10.3 as 10.2 fails with INTERNALERROR> pluggy._manager.PluginValidationError: unknown hook 'pytest_configure_node'
-# scipy from 1.6.3 to 1.10
-# expecttest from 0.1.3 to 0.3.0
-# xdoctest from 1.0.2 to 1.3.0
-python -m pip install "future==0.18.2" "hypothesis==5.35.1" "expecttest==0.3.0" "librosa>=0.6.2" "scipy==1.10.1" "psutil==5.9.1" "pynvml==11.4.1" "pillow==9.2.0" "unittest-xml-reporting<=3.2.0,>=2.0.0" "pytest==7.1.3" "pytest-xdist==2.5.0" "pytest-flakefinder==1.1.0" "pytest-rerunfailures==10.3" "pytest-shard==0.1.2" "sympy==1.11.1" "xdoctest==1.3.0" "pygments==2.12.0" "opt-einsum>=3.3" "networkx==2.8.8" "mpmath==1.2.1" "pytest-cpp==2.3.0" "boto3==1.35.42"
+python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
 
 # Install Z3 optional dependency for Windows builds.
 python -m pip install z3-solver==4.15.1.0
@@ -59,6 +52,9 @@ python -m pip install parameterized==0.8.1
 # Install pulp for testing ilps under torch\distributed\_tools
 python -m pip install pulp==2.9.0
 
+# Install expecttest to merge https://github.com/pytorch/pytorch/pull/155308
+python -m pip install expecttest==0.3.0
+
 run_tests() {
     # Run nvidia-smi if available
     for path in '/c/Program Files/NVIDIA Corporation/NVSMI/nvidia-smi.exe' /c/Windows/System32/nvidia-smi.exe; do

.ci/pytorch/windows/internal/static_lib_test.bat

@@ -63,7 +63,7 @@ if errorlevel 1 exit /b 1
 call %CONDA_HOME%\condabin\activate.bat testenv
 if errorlevel 1 exit /b 1
 
-call conda install  -y -q -c conda-forge libuv=1.51
+call conda install  -y -q -c conda-forge libuv=1.39
 call conda install -y -q intel-openmp
 
 echo "install and test libtorch"

.ci/wheel/build_wheel.sh

@@ -177,7 +177,8 @@ source ~/${desired_python}-build/bin/activate
 retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements.txt"
 retry brew install libomp
 
-# For USE_DISTRIBUTED=1 on macOS, need libuv, which is build as part of tensorpipe submodule
+# For USE_DISTRIBUTED=1 on macOS, this enables gloo, which needs libuv, which
+# is build as part of tensorpipe submodule
 export USE_DISTRIBUTED=1
 
 export USE_MKLDNN=OFF

.circleci/scripts/functorch_doc_push_script.sh

@@ -0,0 +1,47 @@
+#!/bin/bash
+# =================== The following code **should** be executed inside Docker container ===================
+
+# Install dependencies
+sudo apt-get -y update
+sudo apt-get -y install expect-dev
+
+# This is where the local pytorch install in the docker image is located
+pt_checkout="/var/lib/jenkins/workspace"
+source "$pt_checkout/.ci/pytorch/common_utils.sh"
+echo "functorch_doc_push_script.sh: Invoked with $*"
+
+set -ex
+
+version=${DOCS_VERSION:-nightly}
+echo "version: $version"
+
+# Build functorch docs
+pushd $pt_checkout/functorch/docs
+pip -q install -r requirements.txt
+make html
+popd
+
+git clone https://github.com/pytorch/functorch -b gh-pages --depth 1 functorch_ghpages
+pushd functorch_ghpages
+
+if [ $version == "main" ]; then
+  version=nightly
+fi
+
+git rm -rf "$version" || true
+mv "$pt_checkout/functorch/docs/build/html" "$version"
+
+git add "$version" || true
+git status
+git config user.email "soumith+bot@pytorch.org"
+git config user.name "pytorchbot"
+# If there aren't changes, don't make a commit; push is no-op
+git commit -m "Generate Python docs from pytorch/pytorch@${GITHUB_SHA}" || true
+git status
+
+if [[ "${WITH_PUSH:-}" == true ]]; then
+  git push -u origin gh-pages
+fi
+
+popd
+# =================== The above code **should** be executed inside Docker container ===================

.clang-tidy

@@ -69,8 +69,6 @@ readability-string-compare,
 '
 HeaderFilterRegex: '^(aten/|c10/|torch/).*$'
 WarningsAsErrors: '*'
-LineFilter:
-  - name: '/usr/include/.*'
 CheckOptions:
   cppcoreguidelines-special-member-functions.AllowSoleDefaultDtor: true
   cppcoreguidelines-special-member-functions.AllowImplicitlyDeletedCopyOrMove: true

.github/ISSUE_TEMPLATE/ci-sev.md

@@ -1,10 +1,6 @@
 ---
 name: "⚠️ CI SEV"
 about: Tracking incidents for PyTorch's CI infra.
-title: ''
-labels: ''
-assignees: ''
-
 ---
 
 > NOTE: Remember to label this issue with "`ci: sev`"

.github/ISSUE_TEMPLATE/disable-autorevert.md

@@ -1,18 +0,0 @@
----
-name: DISABLE AUTOREVERT
-about: Disables autorevert when open
-title: "❌​\U0001F519​ [DISABLE AUTOREVERT]"
-labels: 'ci: disable-autorevert'
-assignees: ''
-
----
-
-This issue, while open, disables the autorevert functionality.
-
-More details can be found [here](https://github.com/pytorch/test-infra/blob/main/aws/lambda/pytorch-auto-revert/README.md)
-
-
-## Why are you disabling autorevert?
-
-
-## Links to any issues/commits/errors that shows the source of problem

.github/ISSUE_TEMPLATE/disable-ci-jobs.md

@@ -1,10 +1,8 @@
 ---
 name: Disable CI jobs (PyTorch Dev Infra only)
 about: Use this template to disable CI jobs
-title: DISABLED [WORKFLOW_NAME] / [PLATFORM_NAME] / [JOB_NAME]
-labels: 'module: ci'
-assignees: ''
-
+title: "DISABLED [WORKFLOW_NAME] / [PLATFORM_NAME] / [JOB_NAME]"
+labels: "module: ci"
 ---
 
 > For example, DISABLED pull / win-vs2022-cpu-py3 / test (default). Once

.github/actionlint.yaml

@@ -22,9 +22,6 @@ self-hosted-runner:
     - linux.arm64.m7g.4xlarge
     - linux.arm64.m7g.4xlarge.ephemeral
     - linux.arm64.r7g.12xlarge.memory
-    - linux.aws.h100
-    - linux.aws.h100.4
-    - linux.aws.h100.8
     - linux.4xlarge.nvidia.gpu
     - linux.8xlarge.nvidia.gpu
     - linux.16xlarge.nvidia.gpu

.github/actions/reuse-old-whl/reuse_old_whl.py

@@ -264,7 +264,7 @@ def unzip_artifact_and_replace_files() -> None:
         change_content_to_new_version(f"artifacts/dist/{old_stem}/torch/version.py")
 
         for file in Path(f"artifacts/dist/{old_stem}").glob(
-            "*.dist-info/*",
+            "*.dist-info/**",
         ):
             change_content_to_new_version(file)
 

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

@@ -6,12 +6,6 @@ inputs:
   cuda-version:
     description: which cuda version to install, 'cpu' for none
     required: true
-  python-version:
-    required: false
-    type: string
-    default: "3.10"
-    description: |
-      The python version to be used. Will be 3.10 by default
 
 runs:
   using: composite
@@ -44,24 +38,18 @@ runs:
         CONDA="C:\Jenkins\Miniconda3\condabin\conda.bat"
 
         {
-          echo "CONDA=${CONDA}";
           echo "CONDA_RUN=${CONDA} run --no-capture-output";
           echo "CONDA_BUILD=${CONDA} run conda-build";
           echo "CONDA_INSTALL=${CONDA} install";
         } >> "${GITHUB_ENV}"
 
     - name: Setup Python3
-      env:
-          PYTHON_VERSION: ${{ inputs.python-version }}
       shell: bash
       run: |
         set +e
         set -x
 
-        # Create new py_tmp env with python-version
-        ${CONDA} create -y -n py_tmp python=${PYTHON_VERSION} intel-openmp libuv
-
-        PYTHON3=$(${CONDA_RUN} -n py_tmp which python3)
+        PYTHON3=$(${CONDA_RUN} which python3)
         EXIT_CODE=$?
 
         if [[ "${EXIT_CODE}" == "0" ]]; then
@@ -74,7 +62,7 @@ runs:
           # installation, which is Python 3 based. Its Python is default to Python 3. Further, there
           # is also the Miniconda installation that is Python 2 based, and both can be installed if
           # needed. In both cases, Python binary is just called python
-          PYTHON=$(${CONDA_RUN} -n py_tmp which python)
+          PYTHON=$(${CONDA_RUN} which python)
           EXIT_CODE=$?
 
           if [[ "${EXIT_CODE}" == "0" ]]; then

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-0307428d65acf5cf1a73a70a7722e076bbb83f22
+d119fc86140785e7efc8f125c17153544d1e0f20

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-0fc62aa26a30ed7ca419d285f285cb5ba02c4394
+c77852e117bdf056c8e9a087e51d6f65cf6ba53d

.github/labeler.yml

@@ -130,6 +130,3 @@
 - torch/csrc/inductor/aoti_include/**
 - torchgen/aoti/**
 - torchgen/gen_aoti_c_shim.py
-
-"ciflow/vllm":
-- .github/ci_commit_pins/vllm.txt

.github/merge_rules.yaml

@@ -525,21 +525,6 @@
   - Lint
   - pull
 
-- name: typechecking
-  patterns:
-  - 'pyrefly.toml'
-  - 'mypy.ini'
-  - 'mypy-strict.ini'
-  approved_by:
-  - lolpack
-  - maggiemoss
-  - ndmitchell
-  - kinto0
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
 - name: superuser
   patterns:
   - '*'

.github/pytorch-probot.yml

@@ -1,44 +1,41 @@
 tracking_issue: 24422
 ciflow_tracking_issue: 64124
 ciflow_push_tags:
-- ciflow/b200
-- ciflow/b200-symm-mem
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
-- ciflow/h100
-- ciflow/h100-cutlass-backend
-- ciflow/h100-distributed
-- ciflow/h100-symm-mem
+- ciflow/triton_binaries
 - ciflow/inductor
-- ciflow/inductor-cu126
-- ciflow/inductor-micro-benchmark
-- ciflow/inductor-micro-benchmark-cpu-x86
-- ciflow/inductor-perf-compare
-- ciflow/inductor-perf-test-nightly-rocm
-- ciflow/inductor-perf-test-nightly-x86-zen
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
+- ciflow/inductor-perf-test-nightly-rocm
+- ciflow/inductor-perf-compare
+- ciflow/inductor-micro-benchmark
+- ciflow/inductor-micro-benchmark-cpu-x86
+- ciflow/inductor-perf-test-nightly-x86-zen
+- ciflow/inductor-cu126
 - ciflow/linux-aarch64
 - ciflow/mps
 - ciflow/nightly
-- ciflow/op-benchmark
 - ciflow/periodic
 - ciflow/periodic-rocm-mi300
-- ciflow/pull
-- ciflow/quantization-periodic
-- ciflow/riscv64
 - ciflow/rocm
 - ciflow/rocm-mi300
 - ciflow/s390
+- ciflow/riscv64
 - ciflow/slow
-- ciflow/torchbench
-- ciflow/triton_binaries
 - ciflow/trunk
 - ciflow/unstable
-- ciflow/vllm
-- ciflow/win-arm64
 - ciflow/xpu
+- ciflow/vllm
+- ciflow/torchbench
+- ciflow/op-benchmark
+- ciflow/pull
+- ciflow/h100
+- ciflow/h100-distributed
+- ciflow/win-arm64
+- ciflow/h100-symm-mem
+- ciflow/h100-cutlass-backend
 retryable_workflows:
 - pull
 - trunk
@@ -47,4 +44,4 @@ retryable_workflows:
 - inductor-A100-perf-nightly
 labeler_config: labeler.yml
 label_to_label_config: label_to_label.yml
-mergebot: true
+mergebot: True

.github/scripts/generate_binary_build_matrix.py

@@ -30,7 +30,7 @@ CUDA_ARCHES_CUDNN_VERSION = {
 }
 
 # NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
-ROCM_ARCHES = ["6.4", "7.0"]
+ROCM_ARCHES = ["6.3", "6.4"]
 
 XPU_ARCHES = ["xpu"]
 
@@ -53,7 +53,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
         "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
         "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
@@ -70,7 +70,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
         "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
         "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
@@ -87,7 +87,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
         "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
-        "nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | "
+        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
         "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
         "nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
         "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "

.github/scripts/generate_ci_workflows.py

@@ -135,7 +135,7 @@ ROCM_SMOKE_WORKFLOWS = [
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
             arches=["6.4"],
-            python_versions=["3.10"],
+            python_versions=["3.9"],
         ),
         ciflow_config=CIFlowConfig(
             labels={
@@ -155,7 +155,7 @@ LINUX_BINARY_SMOKE_WORKFLOWS = [
         package_type="manywheel",
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
-            arches=["13.0"],
+            arches=["12.8"],
             python_versions=["3.12"],
         ),
         branches="main",

.github/templates/linux_binary_build_workflow.yml.j2

@@ -71,15 +71,12 @@ jobs:
     with:!{{ upload.binary_env_as_input(config) }}
       {%- if "aarch64" in build_environment %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       {%- elif "s390x" in build_environment %}
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
       timeout-minutes: 420
-      {%- elif config["gpu_arch_type"] == "rocm" %}
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
       {%- elif "conda" in build_environment and config["gpu_arch_type"] == "cuda" %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.24xlarge.ephemeral

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

@@ -187,6 +187,8 @@ jobs:
 
       - name: Install nvidia driver, nvidia-docker runtime, set GPU_FLAG
         uses: pytorch/test-infra/.github/actions/setup-nvidia@main
+        with:
+          driver-version: ${{ startsWith(inputs.GPU_ARCH_VERSION, '13') && '580.65.06' || '570.133.07' }}
         if: ${{ inputs.GPU_ARCH_TYPE == 'cuda' && steps.filter.outputs.is-test-matrix-empty == 'False' }}
 
       - name: configure aws credentials

.github/workflows/_docs.yml

@@ -67,7 +67,7 @@ jobs:
             # an OOM issue when running the job, so this upgrades the runner from 4xlarge
             # to the next available tier of 12xlarge. So much memory just to generate cpp
             # doc
-            runner: ${{ inputs.runner_prefix }}linux.12xlarge.memory
+            runner: ${{ inputs.runner_prefix }}linux.12xlarge
             # TODO: Nightly cpp docs take longer and longer to finish (more than 3h now)
             # Let's try to figure out how this can be improved
             timeout-minutes: 360

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

@@ -2,12 +2,6 @@ name: Get Changed Files
 
 on:
   workflow_call:
-    inputs:
-      all_files:
-        description: "Whether to return all files instead of just changed files"
-        required: false
-        type: boolean
-        default: false
     outputs:
       changed-files:
         description: "List of changed files (space-separated) or '*' if not in a PR"
@@ -32,23 +26,17 @@ jobs:
             # Get the PR number from the github context
             PR_NUMBER="${{ github.event.number }}"
 
-            # Check if all_files is requested
-            if [ "${{ inputs.all_files }}" = "true" ]; then
-              echo "all_files input is true, returning all files"
-              echo "changed-files=*" >> "$GITHUB_OUTPUT"
-            else
-              # Use gh CLI to get changed files in the PR with explicit repo
-              CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
+            # Use gh CLI to get changed files in the PR with explicit repo
+            CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
 
-              if [ -z "$CHANGED_FILES" ]; then
-                echo "No changed files found, setting to '*'"
-                CHANGED_FILES="*"
-              fi
-
-              echo "Changed files: $CHANGED_FILES"
-              echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
+            if [ -z "$CHANGED_FILES" ]; then
+              echo "No changed files found, setting to '*'"
+              CHANGED_FILES="*"
             fi
 
+            echo "Changed files: $CHANGED_FILES"
+            echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
+
           else
             echo "Not in PR context, setting changed files to '*'"
             echo "changed-files=*" >> "$GITHUB_OUTPUT"

.github/workflows/_linux-test.yml

@@ -273,8 +273,6 @@ jobs:
           TEST_CONFIG: ${{ matrix.config }}
           SHARD_NUMBER: ${{ matrix.shard }}
           NUM_TEST_SHARDS: ${{ matrix.num_shards }}
-          EXTRA_FLAGS: ${{ matrix.extra_flags || '' }}
-          OP_BENCHMARK_TESTS: ${{ matrix.op_benchmark_tests }}
           REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
           CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
           VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}

.github/workflows/_win-build.yml

@@ -151,7 +151,7 @@ jobs:
           BUILD_WHEEL: 1
           MAX_JOBS: 8
           CUDA_VERSION: ${{ inputs.cuda-version }}
-          PYTHON_VERSION: "3.10"
+          PYTHON_VERSION: "3.9"
           SCCACHE_BUCKET: "ossci-compiler-cache"
           SCCACHE_S3_KEY_PREFIX: ${{ github.workflow }}
           SCCACHE_REGION: us-east-1

.github/workflows/_win-test.yml

@@ -184,7 +184,7 @@ jobs:
         env:
           USE_CUDA: ${{ inputs.cuda-version != 'cpu' && '1' || '0' }}
           INSTALL_WINDOWS_SDK: 1
-          PYTHON_VERSION: "3.10"
+          PYTHON_VERSION: 3.9
           CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
           VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
           TEST_SHOWLOCALS: ${{ steps.keep-going.outputs.ci-test-showlocals }}

.github/workflows/b200-symm-mem.yml

@@ -1,60 +0,0 @@
-name: Limited CI for symmetric memory tests on B200
-
-on:
-  pull_request:
-    paths:
-      - .github/workflows/b200-symm-mem.yml
-  workflow_dispatch:
-  push:
-    tags:
-      - ciflow/b200-symm-mem/*
-  schedule:
-    - cron: 22 8 * * *  # about 1:22am 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: ${{ github.ref_type }}
-
-  linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '10.0'
-      test-matrix: |
-        { include: [
-          { config: "b200-symm-mem", shard: 1, num_shards: 1, runner: "linux.dgx.b200.8" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc11-sm100-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-    secrets: inherit

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

@@ -36,7 +36,7 @@ jobs:
     runs-on: linux.9xlarge.ephemeral
     strategy:
       matrix:
-        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.3", "rocm6.4", "rocm7.0", "cpu"]
+        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.3", "rocm6.4", "cpu"]
     steps:
       - name: Build docker image
         uses: pytorch/pytorch/.github/actions/binary-docker-build@main

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

@@ -52,8 +52,8 @@ jobs:
           { tag: "cuda12.9" },
           { tag: "cuda12.8" },
           { tag: "cuda12.6" },
+          { tag: "rocm6.3"  },
           { tag: "rocm6.4"  },
-          { tag: "rocm7.0"  },
           { tag: "cpu"      },
         ]
     steps:

.github/workflows/build-magma-rocm-linux.yml

@@ -34,7 +34,7 @@ jobs:
       id-token: write
     strategy:
       matrix:
-        rocm_version: ["70", "64"]
+        rocm_version: ["64", "63"]
     steps:
       - name: Checkout PyTorch
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

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

@@ -52,8 +52,8 @@ jobs:
           { name: "manylinuxaarch64-builder",       tag: "cuda13.0",          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: "rocm6.3",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
-          { name: "manylinux2_28-builder",          tag: "rocm7.0",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cpu",               runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28_aarch64-builder",  tag: "cpu-aarch64",       runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxcxx11-abi-builder",     tag: "cpu-cxx11-abi",     runner: "linux.9xlarge.ephemeral" },

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

@@ -50,12 +50,12 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        py_vers: [ "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
+        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
         device: ["cuda", "rocm", "xpu", "aarch64"]
         docker-image: ["pytorch/manylinux2_28-builder:cpu"]
         include:
           - device: "rocm"
-            rocm_version: "7.0"
+            rocm_version: "6.4"
             runs_on: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge"
           - device: "cuda"
             rocm_version: ""
@@ -108,6 +108,9 @@ jobs:
 
           # Determine python executable for given version
           case $PY_VERS in
+          3.9)
+            PYTHON_EXECUTABLE=/opt/python/cp39-cp39/bin/python
+            ;;
           3.10)
             PYTHON_EXECUTABLE=/opt/python/cp310-cp310/bin/python
             ;;
@@ -191,7 +194,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        py_vers: [ "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
+        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
         device: ["xpu"]
     timeout-minutes: 40
     env:

.github/workflows/create_release.yml

@@ -35,7 +35,6 @@ jobs:
       contents: write
     outputs:
       pt_release_name: ${{ steps.release_name.outputs.pt_release_name }}
-      pt_pep517_release_name: ${{ steps.release_name.outputs.pt_pep517_release_name }}
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
         with:
@@ -54,12 +53,8 @@ jobs:
           tag_or_branch="${tag_or_branch#refs/heads/}"
           # replace directory separators with _ in branch name
           tag_or_branch="${tag_or_branch//\//_}"
-          torch_version="$(python -c 'from tools.generate_torch_version import get_torch_version; print(get_torch_version())')"
-          {
-            echo "PT_RELEASE_NAME=pytorch-$tag_or_branch";
-            echo "PT_RELEASE_FILE=pytorch-$tag_or_branch.tar.gz";
-            echo "PT_PEP517_RELEASE_FILE=torch-${torch_version}.tar.gz";
-          } >> "$GITHUB_ENV"
+          echo "PT_RELEASE_NAME=pytorch-$tag_or_branch" >> "$GITHUB_ENV"
+          echo "PT_RELEASE_FILE=pytorch-$tag_or_branch.tar.gz" >> "$GITHUB_ENV"
       - name: Checkout optional submodules
         run: python3 tools/optional_submodules.py
       - name: Copy docs requirements for inclusion
@@ -69,47 +64,30 @@ jobs:
           cp .ci/docker/requirements-docs.txt docs/requirements.txt
       - name: Create source distribution
         run: |
-          # Create new folder with specified name so extracting the archive yields that
-          rm -rf "/tmp/$PT_RELEASE_NAME"
-          cp -r "$PWD" "/tmp/$PT_RELEASE_NAME"
-          mv "/tmp/$PT_RELEASE_NAME" .
-          # Cleanup
-          rm -rf "$PT_RELEASE_NAME"/{.circleci,.ci}
-          find "$PT_RELEASE_NAME" -name '.git*' -exec rm -rv {} \; || true
-          # Create archive
-          tar -czf "$PT_RELEASE_FILE" "$PT_RELEASE_NAME"
-          echo "Created source archive $PT_RELEASE_FILE with content: $(ls -a "$PT_RELEASE_NAME")"
-      - name: Create PEP 517 compatible source distribution
-        run: |
-          pip install build==1.2.2.post1 || exit 1
-          python -m build --sdist || exit 1
-          cd dist || exit 1
+            # Create new folder with specified name so extracting the archive yields that
+            rm -rf "/tmp/$PT_RELEASE_NAME"
+            cp -r "$PWD" "/tmp/$PT_RELEASE_NAME"
+            mv "/tmp/$PT_RELEASE_NAME" .
+            # Cleanup
+            rm -rf "$PT_RELEASE_NAME"/{.circleci,.ci}
+            find "$PT_RELEASE_NAME" -name '.git*' -exec rm -rv {} \; || true
+            # Create archive
+            tar -czf "$PT_RELEASE_FILE" "$PT_RELEASE_NAME"
+            echo "Created source archive $PT_RELEASE_FILE with content: $(ls -a "$PT_RELEASE_NAME")"
       - name: Upload source distribution for release
         if: ${{ github.event_name == 'release' }}
         uses: softprops/action-gh-release@da05d552573ad5aba039eaac05058a918a7bf631 # v2.2.2
         with:
-          files: |
-            ${{ env.PT_RELEASE_FILE }}
-            ${{ env.PT_PEP517_RELEASE_FILE }}
-      - name: Upload source distribution to GHA artifacts  # for release tags
+          files: ${{env.PT_RELEASE_FILE}}
+      - name: Upload source distribution to GHA artifacts for release tags
         if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
         uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
         with:
           name: ${{ env.PT_RELEASE_FILE }}
           path: ${{ env.PT_RELEASE_FILE }}
-      - name: Upload PEP 517 source distribution to GHA artifacts  # for release tags
-        if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
-        uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
-        with:
-          name: ${{ env.PT_PEP517_RELEASE_FILE }}
-          path: dist/${{ env.PT_PEP517_RELEASE_FILE }}
       - name: Set output
         id: release_name
-        run: |
-          {
-            echo "pt_release_name=${{ env.PT_RELEASE_FILE }}";
-            echo "pt_pep517_release_name=${{ env.PT_PEP517_RELEASE_FILE }}";
-          } >> "${GITHUB_OUTPUT}"
+        run: echo "pt_release_name=${{ env.PT_RELEASE_NAME }}.tar.gz" >> "${GITHUB_OUTPUT}"
 
   upload_source_code_to_s3:
     if: ${{ github.repository == 'pytorch/pytorch' && github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
@@ -125,9 +103,6 @@ jobs:
       - uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
         with:
           name: ${{ needs.release.outputs.pt_release_name }}
-      - uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
-        with:
-          name: ${{ needs.release.outputs.pt_pep517_release_name }}
       - name: Configure AWS credentials(PyTorch account)
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
         with:
@@ -138,9 +113,7 @@ jobs:
           s3-bucket: pytorch
           s3-prefix: source_code/test
           if-no-files-found: warn
-          path: |
-            ${{ needs.release.outputs.pt_release_name }}
-            ${{ needs.release.outputs.pt_pep517_release_name }}
+          path: ${{ needs.release.outputs.pt_release_name }}
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name }}

.github/workflows/docker-builds.yml

@@ -70,8 +70,9 @@ jobs:
           pytorch-linux-jammy-py3-clang18-asan,
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter,
-          pytorch-linux-jammy-py3-clang12-executorch,
+          pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter,
+          # Executorch pin needs update
+          # pytorch-linux-jammy-py3-clang12-executorch,
           pytorch-linux-jammy-py3.12-triton-cpu,
           pytorch-linux-noble-riscv64-py3.12-gcc14
         ]

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

@@ -62,7 +62,7 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -128,11 +128,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -174,11 +174,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -220,11 +220,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda13.0
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -265,7 +265,7 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -331,11 +331,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -377,11 +377,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -423,11 +423,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda13.0
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -468,7 +468,7 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -534,11 +534,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -580,11 +580,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -626,11 +626,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda13.0
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -671,7 +671,7 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -737,11 +737,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -783,11 +783,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -829,11 +829,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda13.0
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -874,7 +874,7 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -940,11 +940,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -986,11 +986,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1032,11 +1032,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda13.0
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1077,7 +1077,7 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
       DESIRED_PYTHON: "3.14"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -1143,11 +1143,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
       DESIRED_PYTHON: "3.14"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1189,11 +1189,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.14"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1235,11 +1235,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda13.0
       DESIRED_PYTHON: "3.14"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1280,7 +1280,7 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
       DESIRED_PYTHON: "3.14t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -1346,11 +1346,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
       DESIRED_PYTHON: "3.14t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1392,11 +1392,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.14t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1438,11 +1438,11 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: cuda13.0
       DESIRED_PYTHON: "3.14t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.r7g.12xlarge.memory
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -316,6 +316,120 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  libtorch-rocm6_3-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.3
+      GPU_ARCH_VERSION: "6.3"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.3
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: libtorch-rocm6_3-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-rocm6_3-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-rocm6_3-shared-with-deps-release-build
+      - get-label-type
+    runs-on: linux.rocm.gpu.mi250
+    timeout-minutes: 240
+    env:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.3
+      GPU_ARCH_VERSION: "6.3"
+      GPU_ARCH_TYPE: rocm
+      SKIP_ALL_TESTS: 1
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.3
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+    steps:
+      - name: Setup ROCm
+        uses: ./.github/actions/setup-rocm
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-rocm6_3-shared-with-deps-release
+          path: "${{ runner.temp }}/artifacts/"
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: ROCm set GPU_FLAG
+        run: |
+          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
+      - name: configure aws credentials
+        id: aws_creds
+        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
+        uses: aws-actions/configure-aws-credentials@v4
+        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: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        with:
+          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
+          docker-image-name: libtorch-cxx11-builder
+          custom-tag-prefix: rocm6.3
+          docker-build-dir: .ci/docker
+          working-directory: pytorch
+      - name: Pull Docker image
+        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
+        with:
+          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Test Pytorch binary
+        uses: ./pytorch/.github/actions/test-pytorch-binary
+        env:
+          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Teardown ROCm
+        uses: ./.github/actions/teardown-rocm
+  libtorch-rocm6_3-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-rocm6_3-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: rocm6.3
+      GPU_ARCH_VERSION: "6.3"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.3
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-rocm6_3-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   libtorch-rocm6_4-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -333,7 +447,6 @@ jobs:
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
       build_name: libtorch-rocm6_4-shared-with-deps-release
       build_environment: linux-binary-libtorch
     secrets:
@@ -430,118 +543,3 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
-
-  libtorch-rocm7_0-shared-with-deps-release-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.0
-      GPU_ARCH_VERSION: "7.0"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.0
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
-      build_name: libtorch-rocm7_0-shared-with-deps-release
-      build_environment: linux-binary-libtorch
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-rocm7_0-shared-with-deps-release-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - libtorch-rocm7_0-shared-with-deps-release-build
-      - get-label-type
-    runs-on: linux.rocm.gpu.mi250
-    timeout-minutes: 240
-    env:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.0
-      GPU_ARCH_VERSION: "7.0"
-      GPU_ARCH_TYPE: rocm
-      SKIP_ALL_TESTS: 1
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.0
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-    steps:
-      - name: Setup ROCm
-        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
-        name: Download Build Artifacts
-        with:
-          name: libtorch-rocm7_0-shared-with-deps-release
-          path: "${{ runner.temp }}/artifacts/"
-      - name: Checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-          submodules: recursive
-          path: pytorch
-          show-progress: false
-      - name: Clean PyTorch checkout
-        run: |
-          # Remove any artifacts from the previous checkouts
-          git clean -fxd
-        working-directory: pytorch
-      - name: ROCm set GPU_FLAG
-        run: |
-          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
-      - name: configure aws credentials
-        id: aws_creds
-        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
-        uses: aws-actions/configure-aws-credentials@v4
-        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: Calculate docker image
-        id: calculate-docker-image
-        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
-        with:
-          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
-          docker-image-name: libtorch-cxx11-builder
-          custom-tag-prefix: rocm7.0
-          docker-build-dir: .ci/docker
-          working-directory: pytorch
-      - name: Pull Docker image
-        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
-        with:
-          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      - name: Test Pytorch binary
-        uses: ./pytorch/.github/actions/test-pytorch-binary
-        env:
-          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      - name: Teardown ROCm
-        uses: ./.github/actions/teardown-rocm
-  libtorch-rocm7_0-shared-with-deps-release-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: libtorch-rocm7_0-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: rocm7.0
-      GPU_ARCH_VERSION: "7.0"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.0
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-rocm7_0-shared-with-deps-release
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml

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

@@ -42,7 +42,7 @@ jobs:
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
-  manywheel-py3_12-cuda13_0-build:
+  manywheel-py3_12-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -51,22 +51,22 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_12-cuda13_0
+      build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda13_0-test:  # Testing
+  manywheel-py3_12-cuda12_8-test:  # Testing
     if: ${{ github.repository_owner == 'pytorch' }}
     needs:
-      - manywheel-py3_12-cuda13_0-build
+      - manywheel-py3_12-cuda12_8-build
       - get-label-type
     uses: ./.github/workflows/_binary-test-linux.yml
     with:
@@ -74,13 +74,13 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: "12.8"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
       DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda13_0
+      build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner

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

@@ -44,7 +44,7 @@ jobs:
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
-  manywheel-py3_10-rocm6_4-build:
+  manywheel-py3_9-rocm6_4-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -58,17 +58,16 @@ jobs:
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
-      build_name: manywheel-py3_10-rocm6_4
+      build_name: manywheel-py3_9-rocm6_4
       build_environment: linux-binary-manywheel-rocm
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_10-rocm6_4-test:  # Testing
+  manywheel-py3_9-rocm6_4-test:  # Testing
     if: ${{ github.repository_owner == 'pytorch' }}
     needs:
-      - manywheel-py3_10-rocm6_4-build
+      - manywheel-py3_9-rocm6_4-build
       - get-label-type
     runs-on: linux.rocm.gpu.mi250
     timeout-minutes: 240
@@ -83,14 +82,14 @@ jobs:
       SKIP_ALL_TESTS: 1
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
       - uses: actions/download-artifact@v4.1.7
         name: Download Build Artifacts
         with:
-          name: manywheel-py3_10-rocm6_4
+          name: manywheel-py3_9-rocm6_4
           path: "${{ runner.temp }}/artifacts/"
       - name: Checkout PyTorch
         uses: actions/checkout@v4

.github/workflows/lint.yml

@@ -31,8 +31,6 @@ jobs:
     if: github.repository_owner == 'pytorch'
     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') }}
 
   lintrunner-clang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@@ -55,7 +53,7 @@ jobs:
     with:
       timeout: 120
       runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
-      docker-image: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter
+      docker-image: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter
       # NB: A shallow checkout won't work here because calculate-docker-image requires a full checkout
       # to run git rev-parse HEAD~:.ci/docker when a new image is needed
       fetch-depth: 0
@@ -266,10 +264,10 @@ jobs:
         with:
           submodules: false
           fetch-depth: 1
-      - name: Setup Python 3.10
+      - name: Setup Python 3.9
         uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
         with:
-          python-version: '3.10'
+          python-version: '3.9'
           architecture: x64
           cache: pip
       - name: Install dependencies

.github/workflows/operator_microbenchmark.yml

@@ -1,46 +0,0 @@
-name: operator_microbenchmark
-
-on:
-  push:
-    tags:
-      - ciflow/op-benchmark/*
-  workflow_dispatch:
-  schedule:
-    # Run at 06:00 UTC everyday
-    - cron: 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:
-  id-token: write
-  contents: read
-
-jobs:
-  opmicrobenchmark-build:
-    if: github.repository_owner == 'pytorch'
-    name: opmicrobenchmark-build
-    uses: ./.github/workflows/_linux-build.yml
-    with:
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '8.0 9.0'
-      test-matrix: |
-        { include: [
-          { config: "operator_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.aws.h100" },
-          { config: "operator_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.aws.a100" },
-        ]}
-    secrets: inherit
-
-  opmicrobenchmark-test:
-    name: opmicrobenchmark-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs: opmicrobenchmark-build
-    with:
-      timeout-minutes: 500
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
-      docker-image: ${{ needs.opmicrobenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.opmicrobenchmark-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/periodic.yml

@@ -59,14 +59,13 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.4-py3.10-gcc11
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11
-      cuda-arch-list: 7.5
       test-matrix: |
         { include: [
-          { config: "legacy_nvidia_driver", shard: 1, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 2, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 3, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 4, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 5, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 1, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 2, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 3, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 4, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 5, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
         ]}
     secrets: inherit
 
@@ -113,13 +112,13 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.test-matrix }}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc9-build:
-    name: linux-jammy-cuda12.8-py3.10-gcc9
+  linux-jammy-cuda12_8-py3_9-gcc9-build:
+    name: linux-jammy-cuda12.8-py3.9-gcc9
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9
+      build-environment: linux-jammy-cuda12.8-py3.9-gcc9
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9
       cuda-arch-list: 8.6
       test-matrix: |
@@ -129,14 +128,14 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-jammy-cuda12_8-py3_10-gcc9-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc9
+  linux-jammy-cuda12_8-py3_9-gcc9-test:
+    name: linux-jammy-cuda12.8-py3.9-gcc9
     uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-cuda12_8-py3_10-gcc9-build
+    needs: linux-jammy-cuda12_8-py3_9-gcc9-build
     with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-build.outputs.test-matrix }}
+      build-environment: linux-jammy-cuda12.8-py3.9-gcc9
+      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_9-gcc9-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_9-gcc9-build.outputs.test-matrix }}
     secrets: inherit
 
   linux-jammy-cuda12_8-py3_10-gcc9-debug-build:

.github/workflows/pull.yml

@@ -318,6 +318,32 @@ jobs:
         ]}
     secrets: inherit
 
+  linux-jammy-py3-clang12-executorch-build:
+    if: false  # Docker build needs pin update
+    name: linux-jammy-py3-clang12-executorch
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-py3-clang12-executorch
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang12-executorch
+      test-matrix: |
+        { include: [
+          { config: "executorch", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-py3-clang12-executorch-test:
+    name: linux-jammy-py3-clang12-executorch
+    uses: ./.github/workflows/_linux-test.yml
+    needs: linux-jammy-py3-clang12-executorch-build
+    if: false # Has been broken for a while
+    with:
+      build-environment: linux-jammy-py3-clang12-executorch
+      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-cuda12_8-py3_10-gcc9-inductor-build:
     name: cuda12.8-py3.10-gcc9-sm75
     uses: ./.github/workflows/_linux-build.yml
@@ -343,14 +369,14 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.test-matrix }}
     secrets: inherit
 
-  linux-jammy-xpu-n-py3_10-build:
-    name: linux-jammy-xpu-n-py3.10
+  linux-jammy-xpu-n-py3_9-build:
+    name: linux-jammy-xpu-n-py3.9
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
       sync-tag: linux-xpu-n-build
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
-      build-environment: linux-jammy-xpu-n-py3.10
+      build-environment: linux-jammy-xpu-n-py3.9
       docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3
       test-matrix: |
         { include: [

.github/workflows/quantization-periodic.yml

@@ -1,54 +0,0 @@
-name: quantization-periodic
-
-on:
-  push:
-    tags:
-      - ciflow/quantization-periodic/*
-  workflow_dispatch:
-  schedule:
-    # run weekly
-    - cron: "45 0 * * 0"
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  get-default-label-prefix:
-    name: get-default-label-prefix
-    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: ${{ github.ref_type }}
-      opt_out_experiments: lf
-
-  periodic-quantization-build:
-    name: periodic-quantization-build
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-default-label-prefix
-    with:
-      runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
-      build-environment: linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '8.9'
-      test-matrix: |
-        { include: [
-          { config: "quantization", shard: 1, num_shards: 1, runner: "${{ needs.get-default-label-prefix.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu" },
-        ]}
-    secrets: inherit
-  periodic-test-quantization:
-    name: periodic-test-quantization
-    uses: ./.github/workflows/_linux-test.yml
-    needs: periodic-quantization-build
-    with:
-      build-environment: linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      docker-image: ${{ needs.periodic-quantization-build.outputs.docker-image }}
-      test-matrix: ${{ needs.periodic-quantization-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/test-b200.yml

@@ -1,76 +0,0 @@
-# B200 Smoke Tests CI Workflow
-#
-# This workflow runs smoke tests on B200 hardware
-#
-# Flow:
-# 1. Builds PyTorch with CUDA 12.8+ and sm100 architecture for B200
-# 2. Runs smoke tests on linux.dgx.b200 runner
-# 3. Tests executed are defined in .ci/pytorch/test.sh -> test_python_smoke() function
-#
-# Triggered by:
-# - Pull requests modifying this workflow file
-# - Manual dispatch
-# - Schedule (every 6 hours)
-# - Adding ciflow/b200 label to a PR (creates ciflow/b200/* tag)
-
-name: B200 Smoke Tests
-
-on:
-  pull_request:
-    paths:
-      - .github/workflows/test-b200.yml
-  workflow_dispatch:
-  schedule:
-    - cron: 0 4,10,16,22 * * *  # every 6 hours
-  push:
-    tags:
-      - ciflow/b200/*
-
-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: ${{ github.ref_type }}
-
-  linux-jammy-cuda12_8-py3_10-gcc11-sm100-build:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm100
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '10.0'
-      test-matrix: |
-        { include: [
-          { config: "smoke_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-        ]}
-      # config: "smoke_b200" maps to test_python_smoke_b200() in .ci/pytorch/test.sh
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc11-sm100-test:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-sm100-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-    secrets: inherit

.github/workflows/trunk.yml

@@ -259,27 +259,3 @@ jobs:
       docker-image: ${{ needs.verify-cachebench-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.verify-cachebench-cpu-build.outputs.test-matrix }}
     secrets: inherit
-
-  linux-jammy-py3-clang12-executorch-build:
-    name: linux-jammy-py3-clang12-executorch
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3-clang12-executorch
-      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang12-executorch
-      test-matrix: |
-        { include: [
-          { config: "executorch", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-py3-clang12-executorch-test:
-    name: linux-jammy-py3-clang12-executorch
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-py3-clang12-executorch-build
-    with:
-      build-environment: linux-jammy-py3-clang12-executorch
-      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

.github/workflows/unstable.yml

@@ -53,3 +53,27 @@ jobs:
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
+
+  linux-jammy-py3_9-clang9-xla-build:
+    name: linux-jammy-py3_9-clang9-xla
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-py3.9-clang9-xla
+      docker-image-name: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/xla_base:v1.3-lite
+      test-matrix: |
+        { include: [
+          { config: "xla", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.12xlarge" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-py3_9-clang9-xla-test:
+    name: linux-jammy-py3_9-clang9-xla
+    uses: ./.github/workflows/_linux-test.yml
+    needs: linux-jammy-py3_9-clang9-xla-build
+    with:
+      build-environment: linux-jammy-py3.9-clang9-xla
+      docker-image: ${{ needs.linux-jammy-py3_9-clang9-xla-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-py3_9-clang9-xla-build.outputs.test-matrix }}
+    secrets: inherit

.github/workflows/vllm.yml

@@ -36,8 +36,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # When building vLLM, uv doesn't like that we rename wheel without changing the wheel metadata
-      allow-reuse-old-whl: false
       build-additional-packages: "vision audio"
       build-external-packages: "vllm"
       build-environment: linux-jammy-cuda12.8-py3.12-gcc11

.gitignore

@@ -82,7 +82,6 @@ torch/return_types.pyi
 torch/nn/functional.pyi
 torch/utils/data/datapipes/datapipe.pyi
 torch/csrc/autograd/generated/*
-torch/csrc/functionalization/generated/*
 torch/csrc/lazy/generated/*.[!m]*
 torch_compile_debug/
 # Listed manually because some files in this directory are not generated
@@ -260,9 +259,6 @@ gen
 .pytest_cache
 aten/build/*
 
-# Linker scripts for prioritized text optimization
-cmake/linker_script.ld
-
 # Bram
 plsdontbreak
 

.lintrunner.toml

@@ -49,7 +49,7 @@ init_command = [
     'mccabe==0.7.0',
     'pycodestyle==2.14.0',
     'pyflakes==3.4.0',
-    'torchfix==0.4.0 ; python_version >= "3.10" and python_version < "3.13"',
+    'torchfix==0.4.0 ; python_version >= "3.9" and python_version < "3.13"',
 ]
 
 
@@ -123,7 +123,6 @@ is_formatter = true
 code = 'MYPY'
 include_patterns = [
     'setup.py',
-    'functorch/dim/**/*.py',
     'torch/**/*.py',
     'torch/**/*.pyi',
     'caffe2/**/*.py',
@@ -153,7 +152,7 @@ init_command = [
     'python3',
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
-    'numpy==1.26.4 ; python_version >= "3.10" and python_version <= "3.11"',
+    'numpy==1.26.4 ; python_version >= "3.9" and python_version <= "3.11"',
     'numpy==2.1.0 ; python_version >= "3.12"',
     'expecttest==0.3.0',
     'mypy==1.16.0',
@@ -196,7 +195,6 @@ exclude_patterns = [
     'tools/test/gen_operators_yaml_test.py',
     'tools/test/gen_oplist_test.py',
     'tools/test/test_selective_build.py',
-    'tools/experimental/dynamic_shapes/torchfuzz/**',
 ]
 command = [
     'python3',
@@ -966,6 +964,7 @@ exclude_patterns = [
     'test/jit/**',  # should be run through test/test_jit.py
     'test/ao/sparsity/**',  # should be run through test/test_ao_sparsity.py
     'test/fx/**',  # should be run through test/test_fx.py
+    'test/bottleneck_test/**',  # excluded by test/run_test.py
     'test/package/**',  # excluded by test/run_test.py
     'test/distributed/argparse_util_test.py',
     'test/distributed/bin/test_script.py',
@@ -1411,6 +1410,8 @@ exclude_patterns = [
     'torch/utils/benchmark/utils/timer.py',
     'torch/utils/benchmark/utils/valgrind_wrapper/__init__.py',
     'torch/utils/benchmark/utils/valgrind_wrapper/timer_interface.py',
+    'torch/utils/bottleneck/__init__.py',
+    'torch/utils/bottleneck/__main__.py',
     'torch/utils/bundled_inputs.py',
     'torch/utils/checkpoint.py',
     'torch/utils/collect_env.py',
@@ -1453,7 +1454,7 @@ init_command = [
     '--dry-run={{DRYRUN}}',
     'usort==1.0.8.post1',
     'isort==6.0.1',
-    'ruff==0.13.1',  # sync with RUFF
+    'ruff==0.12.9',  # sync with RUFF
 ]
 is_formatter = true
 
@@ -1587,7 +1588,7 @@ init_command = [
     'python3',
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
-    'ruff==0.13.1',  # sync with PYFMT
+    'ruff==0.12.9',  # sync with PYFMT
 ]
 is_formatter = true
 

BUILD.bazel

@@ -22,7 +22,6 @@ COMMON_COPTS = [
     "-DHAVE_SHM_UNLINK=1",
     "-D_FILE_OFFSET_BITS=64",
     "-DUSE_FBGEMM",
-    "-DUSE_DISTRIBUTED",
     "-DAT_PER_OPERATOR_HEADERS",
     "-DATEN_THREADING=NATIVE",
     "-DNO_CUDNN_DESTROY_HANDLE",
@@ -91,8 +90,6 @@ generated_cpu_cpp = [
     "aten/src/ATen/NativeMetaFunctions.h",
     "aten/src/ATen/RegistrationDeclarations.h",
     "aten/src/ATen/VmapGeneratedPlumbing.h",
-    "aten/src/ATen/ViewMetaClasses.h",
-    "aten/src/ATen/ViewMetaClasses.cpp",
     "aten/src/ATen/core/aten_interned_strings.h",
     "aten/src/ATen/core/enum_tag.h",
     "aten/src/ATen/core/TensorBody.h",
@@ -813,7 +810,7 @@ cc_library(
     name = "torch_python",
     srcs = libtorch_python_core_sources
         + if_cuda(libtorch_python_cuda_sources)
-        + if_cuda(libtorch_python_distributed_sources)
+        + libtorch_python_distributed_sources
         + GENERATED_AUTOGRAD_PYTHON,
     hdrs = glob([
         "torch/csrc/generic/*.cpp",
@@ -835,6 +832,36 @@ pybind_extension(
     ],
 )
 
+cc_library(
+    name = "functorch",
+    hdrs = glob([
+        "functorch/csrc/dim/*.h",
+    ]),
+    srcs = glob([
+        "functorch/csrc/dim/*.cpp",
+    ]),
+    deps = [
+        ":aten_nvrtc",
+        ":torch_python",
+        "@pybind11",
+    ],
+)
+
+pybind_extension(
+    name = "functorch/_C",
+    copts=[
+        "-DTORCH_EXTENSION_NAME=_C"
+    ],
+    srcs = [
+        "functorch/csrc/init_dim_only.cpp",
+    ],
+    deps = [
+        ":functorch",
+        ":torch_python",
+        ":aten_nvrtc",
+    ],
+)
+
 cc_binary(
     name = "torch/bin/torch_shm_manager",
     srcs = [
@@ -875,6 +902,7 @@ py_library(
     ],
     data = [
         ":torch/_C.so",
+        ":functorch/_C.so",
         ":torch/bin/torch_shm_manager",
     ],
 )
@@ -1077,7 +1105,6 @@ test_suite(
         "aten/src/ATen/templates/LazyNonNativeIr.h",
         "aten/src/ATen/templates/RegisterDispatchKey.cpp",
         "aten/src/ATen/templates/RegisterDispatchDefinitions.ini",
-        "aten/src/ATen/templates/ViewMetaClassesPythonBinding.cpp",
         "aten/src/ATen/native/native_functions.yaml",
         "aten/src/ATen/native/tags.yaml",
         "aten/src/ATen/native/ts_native_functions.yaml",

CMakeLists.txt

@@ -1,4 +1,5 @@
 cmake_minimum_required(VERSION 3.27 FATAL_ERROR)
+# cmake_policy(SET CMP0022 NEW) cmake_policy(SET CMP0023 NEW)
 
 # Use compiler ID "AppleClang" instead of "Clang" for XCode. Not setting this
 # sometimes makes XCode C compiler gets detected as "Clang", even when the C++
@@ -180,8 +181,9 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(ppc64le)")
   set(CPU_POWER ON)
 endif()
 
-# For non-supported platforms, turn USE_DISTRIBUTED off by default. It is not
-# tested and likely won't work without additional changes.
+# For non-supported platforms, turn USE_DISTRIBUTED off by default.
+# NB: USE_DISTRIBUTED simply disables the backend; distributed code
+# still gets built
 if(NOT LINUX AND NOT WIN32)
   set(USE_DISTRIBUTED
       OFF
@@ -261,11 +263,11 @@ option(USE_PYTORCH_METAL "Use Metal for PyTorch iOS build" OFF)
 option(USE_PYTORCH_METAL_EXPORT "Export Metal models on MacOSX desktop" OFF)
 option(USE_NATIVE_ARCH "Use -march=native" OFF)
 cmake_dependent_option(USE_MPS "Use MPS for macOS build" ON "MPS_FOUND" OFF)
-option(USE_DISTRIBUTED "Use distributed" ON)
+option(USE_DISTRIBUTED "Enable default distributed backends" ON)
 cmake_dependent_option(USE_NCCL "Use NCCL" ON
                        "USE_DISTRIBUTED;USE_CUDA OR USE_ROCM;UNIX;NOT APPLE" OFF)
 cmake_dependent_option(USE_XCCL "Use XCCL" ON
-                       "USE_XPU;UNIX;NOT APPLE" OFF)
+                       "USE_DISTRIBUTED;USE_XPU;UNIX;NOT APPLE" OFF)
 cmake_dependent_option(USE_RCCL "Use RCCL" ON USE_NCCL OFF)
 cmake_dependent_option(USE_RCCL "Use RCCL" ON "USE_NCCL;NOT WIN32" OFF)
 cmake_dependent_option(USE_STATIC_NCCL "Use static NCCL" OFF "USE_NCCL" OFF)
@@ -378,13 +380,6 @@ cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
                        OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                         "CPU_AARCH64" OFF)
-# prioritized text linker, ON by default for AArch64+Linux, option visible to all AArch64, x86 and ppc64le.
-set(USE_PRIORITIZED_TEXT_DEFAULT OFF)
-if(LINUX AND CPU_AARCH64)
-  set(USE_PRIORITIZED_TEXT_DEFAULT ON)
-endif()
-cmake_dependent_option(USE_PRIORITIZED_TEXT_FOR_LD "Use prioritized text linker for ld."
-  "${USE_PRIORITIZED_TEXT_DEFAULT}" "CPU_INTEL OR CPU_AARCH64 OR CPU_POWER" OFF)
 
 option(USE_MIMALLOC "Use mimalloc" OFF)
 # Enable third party mimalloc library to improve memory allocation performance
@@ -437,12 +432,11 @@ if(WIN32)
       PATH_SUFFIXES lib
       NO_DEFAULT_PATH)
     if(NOT libuv_tmp_LIBRARY)
-      set(USE_DISTRIBUTED OFF)
       set(USE_GLOO OFF)
       message(
         WARNING
-          "Libuv is not installed in current conda env. Set USE_DISTRIBUTED to OFF. "
-          "Please run command 'conda install -c conda-forge libuv=1.51' to install libuv."
+          "Libuv is not installed in current conda env. Set USE_GLOO to OFF. "
+          "Please run command 'conda install -c conda-forge libuv=1.39' to install libuv."
       )
     else()
       set(ENV{libuv_ROOT} ${libuv_tmp_LIBRARY}/../../)
@@ -663,11 +657,6 @@ endif(MSVC)
 
 string(APPEND CMAKE_CUDA_FLAGS " -Xfatbin -compress-all")
 
-# Set linker max-page-size to 64KiB on AArch64 Linux
-if(LINUX AND CPU_AARCH64)
-  add_link_options_if_supported("-z,max-page-size=0x10000")
-endif()
-
 # Set INTERN_BUILD_MOBILE for all mobile builds. Components that are not
 # applicable to mobile are disabled by this variable. Setting
 # `BUILD_PYTORCH_MOBILE_WITH_HOST_TOOLCHAIN` environment variable can force it
@@ -1390,6 +1379,10 @@ endif()
 include(cmake/Summary.cmake)
 caffe2_print_configuration_summary()
 
+if(BUILD_FUNCTORCH)
+  add_subdirectory(functorch)
+endif()
+
 # Parse custom debug info
 if(DEFINED USE_CUSTOM_DEBINFO)
   string(REPLACE ";" " " SOURCE_FILES "${USE_CUSTOM_DEBINFO}")
@@ -1428,57 +1421,3 @@ if(BUILD_BUNDLE_PTXAS AND USE_CUDA)
   install(PROGRAMS "${PROJECT_BINARY_DIR}/ptxas"
           DESTINATION "${CMAKE_INSTALL_BINDIR}")
 endif()
-
-if(USE_PRIORITIZED_TEXT_FOR_LD)
-  add_compile_options(
-    $<$<COMPILE_LANGUAGE:C,CXX>:-ffunction-sections>
-    $<$<COMPILE_LANGUAGE:C,CXX>:-fdata-sections>
-  )
-  set(LINKER_SCRIPT_FILE_OUT "${CMAKE_SOURCE_DIR}/cmake/linker_script.ld")
-  set(LINKER_SCRIPT_FILE_IN "${CMAKE_SOURCE_DIR}/cmake/prioritized_text.txt")
-
-  add_custom_command(
-    OUTPUT "${LINKER_SCRIPT_FILE_OUT}"
-    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py --filein "${LINKER_SCRIPT_FILE_IN}" --fout "${LINKER_SCRIPT_FILE_OUT}"
-    DEPENDS ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py "${LINKER_SCRIPT_FILE_IN}"
-    COMMENT "Generating prioritized text linker files"
-    VERBATIM
-  )
-
-  add_custom_target(generate_linker_script DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
-
-  if(BUILD_PYTHON)
-    set(LINKER_OPT_TARGETS torch_python)
-  endif()
-
-  if(NOT BUILD_LIBTORCHLESS)
-    list(APPEND LINKER_OPT_TARGETS torch_cpu c10)
-    if(USE_CUDA)
-      list(APPEND LINKER_OPT_TARGETS torch_cuda c10_cuda)
-    endif()
-    if(USE_XPU)
-      list(APPEND LINKER_OPT_TARGETS torch_xpu c10_xpu)
-    endif()
-    if(USE_ROCM)
-      list(APPEND LINKER_OPT_TARGETS torch_hip c10_hip)
-    endif()
-  endif()
-
-  foreach(tgt IN LISTS LINKER_OPT_TARGETS)
-    if(TARGET ${tgt})
-      add_dependencies("${tgt}" generate_linker_script)
-      target_link_options_if_supported(${tgt} "-T,${LINKER_SCRIPT_FILE_OUT}")
-      set_property(TARGET ${tgt} APPEND PROPERTY LINK_DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
-    else()
-       message(WARNING "Requested target '${tgt}' for linker script optimization was not found.")
-    endif()
-  endforeach()
-
-else()
-  if(LINUX AND CPU_AARCH64)
-    message(WARNING [[
-    It is strongly recommend to enable linker script optimization for all AArch64 Linux builds.
-    To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
-    ]])
-  endif()
-endif()

MANIFEST.in

@@ -1,61 +1,20 @@
 # Reference: https://setuptools.pypa.io/en/latest/userguide/miscellaneous.html
 
-# Include individual top-level files
-include CITATION.cff
-include CODEOWNERS
-include Dockerfile
-include LICENSE
-include MANIFEST.in
-include Makefile
-include NOTICE
-include .bc-linter.yml
-include .clang-format .clang-tidy
-include .cmakelintrc
-include .coveragerc
-include .dockerignore
-include .editorconfig
-include .flake8
-include .gdbinit
-include .lintrunner.toml
-include .lldbinit
-include codex_setup.sh
-include docker.Makefile
-include pyrefly.toml
-include ubsan.supp
-
-# Include bazel and BUCK related files
-include BUILD.bazel BUCK.oss
-include WORKSPACE
-include *.bzl
-include .bazelignore .bazelrc .bazelversion
-
-# Include general configuration files
-include *.ini
-# Include important top-level information
-include *.md
-# Include technical text files at the moment, comprises
-# version.txt, CMakeLists.txt, requirements.txt
-include *.txt
-
-# Include ctags configuration
-include .ctags.d/*.ctags
-
-# Include subfolders completely
-graft .devcontainer
-graft .vscode
+# Include source files in SDist
+include CMakeLists.txt
+include *.bzl *.bazel .bazel* BUILD *.BUILD BUILD.* WORKSPACE
+include BUCK BUCK.*
+include requirements*.txt
+include version.txt
+include [Mm]akefile *.[Mm]akefile [Mm]akefile.*
+include [Dd]ockerfile *.[Dd]ockerfile [Dd]ockerfile.* .dockerignore
 graft android
 graft aten
-graft benchmarks
 graft binaries
 graft c10
 graft caffe2
 graft cmake
-graft docs
 graft functorch
-graft ios
-graft mypy_plugins
-graft scripts
-graft test
 graft third_party
 graft tools
 graft torch
@@ -63,37 +22,29 @@ graft torchgen
 # FIXME: torch-xla build during codegen will fail if include this file in wheel
 exclude torchgen/BUILD.bazel
 
-# The following exclusions omit parts from third-party dependencies that
-# contain invalid symlinks[1] and that are not needed for pytorch, such as
-# bindings for unused languages
-prune third_party/flatbuffers/java
-prune third_party/flatbuffers/kotlin
-prune third_party/ittapi/rust
-prune third_party/nccl/pkg/debian
-prune third_party/opentelemetry-cpp/third_party/prometheus-cpp/cmake/project-import-*
-
-# The following document is also an invalid symlink[1] and superfluous
-exclude third_party/flatbuffers/docs/source/CONTRIBUTING.md
-
-# Omit autogenerated code
-prune torchgen/packaged
-
-# Omit caches, compiled, and scm related content
-prune */__pycache__
-prune **/.github
-prune **/.gitlab
-global-exclude *.o *.obj *.so *.dylib *.a *.pxd *.dll *.lib
-global-exclude *.py[cod] *.swp *~
-global-exclude .git .git-blame-ignore-revs .gitattributes .gitignore .gitmodules
-global-exclude .gitlab-ci.yml
+# Misc files and directories in SDist
+include *.md
+include CITATION.cff
+include LICENSE NOTICE
+include mypy*.ini
+graft benchmarks
+graft docs
+graft mypy_plugins
+graft scripts
 
 # Misc files needed for custom setuptools command
 include .gitignore
 include .gitmodules
 
-# [1] Invalid symlinks for the purposes of Python source distributions are,
-# according to the source distribution format[2] links pointing outside the
-# destination directory or links with a `..` component, which is those of
-# concern here.
+# Include test suites in SDist
+graft test
+include pytest.ini
+include .coveragerc
 
-# [2] https://packaging.python.org/en/latest/specifications/source-distribution-format/#source-distribution-archive-features
+# Prune generated/compiled files
+prune torchgen/packaged
+prune */__pycache__
+global-exclude *.o *.obj *.so *.a *.dylib *.pxd *.dll *.lib *.py[cod]
+
+prune */.git
+global-exclude .git *~ *.swp

README.md

@@ -161,7 +161,7 @@ They require JetPack 4.2 and above, and [@dusty-nv](https://github.com/dusty-nv)
 
 #### Prerequisites
 If you are installing from source, you will need:
-- Python 3.10 or later
+- Python 3.9 or later
 - A compiler that fully supports C++17, such as clang or gcc (gcc 9.4.0 or newer is required, on Linux)
 - Visual Studio or Visual Studio Build Tool (Windows only)
 
@@ -275,7 +275,7 @@ conda install pkg-config libuv
 pip install mkl-static mkl-include
 # Add these packages if torch.distributed is needed.
 # Distributed package support on Windows is a prototype feature and is subject to changes.
-conda install -c conda-forge libuv=1.51
+conda install -c conda-forge libuv
 ```
 
 #### Install PyTorch

aten/src/ATen/CMakeLists.txt

@@ -317,20 +317,10 @@ IF(USE_FBGEMM_GENAI)
         -greedy-reverse-local-assignment=1
         -fhip-new-launch-api)
 
-      # Only compile for gfx942 for now.
-      # This is rather hacky, I could not figure out a clean solution :(
-      set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
-      string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
-      if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-        list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
-      endif()
-      set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
-
       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)
 

aten/src/ATen/Context.cpp

@@ -180,7 +180,7 @@ void Context::setUserEnabledNNPACK(bool e) {
 }
 
 bool Context::allowTF32CuDNN(const std::string& op) const {
-  if (op.empty()){
+  if (op.size() == 0){
     bool allow_tf32_rnn = float32Precision("cuda", "rnn") == "tf32";
     bool allow_tf32_conv = float32Precision("cuda", "conv") == "tf32";
     TORCH_CHECK(
@@ -281,6 +281,9 @@ bool Context::userEnabledOverrideableSDP() const {
 
 static constexpr const auto cublas_config_var_name = "CUBLAS_WORKSPACE_CONFIG";
 static constexpr const std::array<const char*, 2> cublas_deterministic_configs = {":4096:8", ":16:8"};
+#ifdef USE_ROCM
+static constexpr const auto hipblaslt_allow_tf32 = "HIPBLASLT_ALLOW_TF32";
+#endif
 
 bool Context::checkCuBLASConfigDeterministic() {
   // If using CUDA 10.2 or greater, need to make sure CuBLAS workspace config
@@ -340,6 +343,12 @@ void Context::setImmediateMiopen(bool b) {
 }
 
 bool Context::allowTF32CuBLAS() const {
+#ifdef USE_ROCM
+    const auto allow_tf32 = c10::utils::check_env(hipblaslt_allow_tf32);
+    if (allow_tf32 != true) {
+      return false;
+    }
+#endif
   bool legacy_allow_tf32 = float32_matmul_precision != at::Float32MatmulPrecision::HIGHEST;
   bool allow_tf32_new = float32Precision("cuda", "matmul") == "tf32";
   TORCH_CHECK(
@@ -353,6 +362,14 @@ bool Context::allowTF32CuBLAS() const {
 }
 
 void Context::setAllowTF32CuBLAS(bool b) {
+#ifdef USE_ROCM
+  const auto allow_tf32 = c10::utils::check_env(hipblaslt_allow_tf32);
+  if (allow_tf32 != true) {
+    C10_LOG_FIRST_N(INFO, 10) << "torch.backends.cuda.matmul.allow_tf32 is not supported on ROCm by default. "
+                              << "Please set environment variable HIPBLASLT_ALLOW_TF32=1 to enable it.";
+    return;
+  }
+#endif
   float32_matmul_precision = b ? at::Float32MatmulPrecision::HIGH : at::Float32MatmulPrecision::HIGHEST;
   setFloat32Precision("cuda", "matmul", b ? "tf32" : "ieee");
 }
@@ -426,7 +443,7 @@ void Context::setFloat32Precision(const std::string& backend, const std::string&
     std::string msg;
     auto iterp = _fp32_precisions.find(backend);
     TORCH_CHECK(iterp != _fp32_precisions.end());
-    for (const auto& p : iterp->second) {
+    for (auto p : iterp->second) {
       msg += p;
       msg += " ";
     }

aten/src/ATen/DLConvertor.cpp

@@ -401,13 +401,30 @@ T* toDLPackImpl(const Tensor& src) {
   // 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;
+  bool need_normalize_strides = false;
+  int64_t expected_stride = 1;
+  for (int i = src.dim() - 1; i >= 0; i--) {
+    // detect if we do not meet continuous pattern
+    // and the size is 1, so there is opportunity to normalize
+    if (src.stride(i) != expected_stride && src.size(i) == 1) {
+      need_normalize_strides = true;
+      break;
+    }
+    expected_stride *= src.size(i);
+  }
+
   // 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());
+    auto strides = src.strides().vec();
+    for (int i = 0; i < src.dim(); i++) {
+      if (shape[i] < 2) {
+        strides[i] = 1;
+      }
+    }
+    view = src.as_strided(shape, strides, src.storage_offset());
   }
 
   ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);

aten/src/ATen/ExpandUtils.h

@@ -468,7 +468,7 @@ inline Tensor _sum_to(
       // if we assume no reduction due to unbacked we ensure that at runtime.
       TORCH_MAYBE_SYM_CHECK(
           sym_eq(shape[i - leading_dims], sizes[i]),
-          "non-reduction path was assumed due to unbacked symbols expected those two sizes to be the same:",
+          "non-reduction path was assumed due to unabcked symbols expected those two sizes to be the same:",
           shape[i - leading_dims],
           ", ",
           sizes[i])

aten/src/ATen/FunctionalStorageImpl.cpp

@@ -9,6 +9,11 @@
 
 namespace at::functionalization {
 
+ViewMeta ViewMeta::to_out_idx(int64_t out_idx) {
+  if (out_idx == this->out_index) return *this;
+  return ViewMeta(forward_fn, reverse_fn, has_symbolic_inputs, is_multi_output, is_as_strided, out_idx);
+}
+
 // Note [Functionalization: Alias Removal Part 2]
 // See Note [Functionalization: Alias Removal] for more details.
 // This function applies a single update from one of the views to the StorageImpl.
@@ -37,12 +42,12 @@ namespace at::functionalization {
 static const Tensor apply_update(const FunctionalStorageImpl::Update& update, const Tensor& base) {
   at::Tensor t = update.new_val;
   TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
-  if (update.view_metas.empty()) { return t; }
+  if (update.view_metas.empty()) return t;
 
   std::vector<at::Tensor> tmp_values({base});
   tmp_values.reserve(update.view_metas.size());
   for (size_t i = 0; i < update.view_metas.size() - 1; ++i) {
-    at::Tensor next_view = update.view_metas[i]->forward(tmp_values.back());
+    at::Tensor next_view = update.view_metas[i].forward_fn(tmp_values.back(), update.view_metas[i].out_index);
     // NB: We only actually need tmp_values for ops like select/slice/diagonal/squeeze/as_strided
     // All of these ops require additional information to recover the sizes of the original tensor.
     // If need to, we could probably apply this optimization and only bother computing tmp_values
@@ -50,8 +55,9 @@ static const Tensor apply_update(const FunctionalStorageImpl::Update& update, co
     tmp_values.push_back(std::move(next_view));
   }
   for(int64_t i = static_cast<int64_t>(update.view_metas.size()) - 1; i >= 0; --i) {
+    int64_t out_idx = update.view_metas[i].out_index;
     // Each view inverse is implemented in ViewInverses.cpp.
-    t = update.view_metas[i]->reverse(tmp_values[i], t);
+    t = update.view_metas[i].reverse_fn(tmp_values[i], t, out_idx);
   }
   TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
   return t;
@@ -105,13 +111,13 @@ FunctionalStorageImpl::FunctionalStorageImpl(const Tensor& base)
   TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(base_));
 }
 
-void FunctionalStorageImpl::add_update(const Tensor& updated_val, const std::vector<std::shared_ptr<ViewMeta>>& metas) {
+void FunctionalStorageImpl::add_update(const Tensor& updated_val, const std::vector<ViewMeta>& metas) {
   TORCH_CHECK(!frozen_, "cannot mutate tensors with frozen storage");
 
   if (metas.size() > 1) {
     for (size_t i = 1; i < metas.size(); ++i) {
       // Skipping this check for XLA. Would be good to add it back, but it is failing XLA CI
-      TORCH_CHECK(updated_val.device().type() == c10::DeviceType::XLA || !metas[i]->is_as_strided,
+      TORCH_CHECK(updated_val.device().type() == c10::DeviceType::XLA || !metas[i].is_as_strided,
 "During torch.compile, encountered a mutation on a view chain of length ", metas.size(), ", where view ", i,
 " was an as_strided() call. as_strided() is non-compositional, and therefore is not possible to functionalize properly today,"
 "so this behavior is banned in compile. As a workaround, you can either remove the mutation from the model code, or you "

aten/src/ATen/FunctionalStorageImpl.h

@@ -8,89 +8,44 @@ namespace at::functionalization {
 
 // See Note [Functionalization Pass In Core]
 
-enum class InverseReturnMode {
-  /// Specifies that functional inverses should always return a view.
-  AlwaysView,
-  /// Specifies that functional inverses should always return a non-view / copy.
-  NeverView,
-  /// Specifies that functional inverses should return a view unless a (copying)
-  /// scatter
-  /// inverse exists, in which case that will be used instead.
-  /// This avoids as_strided() calls that can be difficult for subclasses to
-  /// handle.
-  ViewOrScatterInverse,
-};
-
-#define FUNCTIONALIZATION_VIEWMETA_NAME(TYPE) \
-  static const char* name() {                 \
-    return #TYPE;                             \
-  }
-
-#define FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(...) \
-  using SerializableTuple = std::tuple<__VA_ARGS__>
-
 // ViewMeta is a class used by the functionalization pass to navigate between
 // a base tensor and a view tensor.
 // For example, if I call `b = a.view1(...)`
-// the functionalization pass will generate and store a ViewMeta specialization
-// for `view1` operation on b that looks like:
+// the functionalization pass will generate and store a ViewMeta on b that looks
+// like:
 //
-// struct TORCH_API view1_ViewMeta : public ViewMeta {
-//   FUNCTIONALIZATION_VIEWMETA_NAME(view1_ViewMeta);
-//   FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(
-//       bool /* reapply_views */,
-//       const std::vector<int64_t>&);
-//
-//   view1_ViewMeta(const SerializableTuple& tpl)
-//       : view1_ViewMeta(std::get<0>(tpl), std::get<1>(tpl)) {}
-//
-//   view1_ViewMeta(bool reapply_views, const std::vector<int64_t>& size)
-//       : ViewMeta(/*has_symbolic_inputs=*/false),
-//         reapply_views(reapply_views),
-//         size(size) {}
-//
-//   Tensor forward(const Tensor& base) override {
-//       return base.view1(...);
+// ViewMeta(
+//   [<captures>](const Tensor& base, int64_t mutated_view_idx) {
+//     return base.view1(...);
+//   },
+//   [<captures>](const at::Tensor& base, const at::Tensor& mutated_view,
+//   int64_t mutated_view_idx) -> at::Tensor {
+//     return at::functionalization::impl::view1_inverse(base, mutated_view,
+//     ...);
 //   }
 //
-//   Tensor reverse(const Tensor& base, const Tensor& mutated_view) override {
-//       return at::functionalization::impl::view1_inverse(base, mutated_view,
-//       ...);
-//   }
+// The forward_fn lambda describes how to replay view1 on a tensor.
 //
-//   SerializableTuple to_serializable_tuple() {
-//     return std::make_tuple(reapply_views, size);
-//   }
-//
-//   bool reapply_views;
-//   std::vector<int64_t> size;
-// };
-//
-// The forward function describes how to replay view1 on a tensor.
-//
-// The reverse function describes how, given a tensor that is already a view,
+// The reverse_fn lambda describes how, given a tensor that is already a view,
 // how to get the corresponding base tensor. See Note [Functionalization Pass:
 // View Inverses] for details.
-//
-// `SerializedTuple` is a typedef that defines an `std::tuple<...>` type
-// representing the `ViewMeta` instance state. Methods that take in/return such
-// a type are used for supporting pickle serialization.
 struct ViewMeta {
   ViewMeta(
+      std::function<Tensor(const Tensor&, int64_t)> forward,
+      std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse,
       bool has_symbolic_inputs,
       bool is_multi_output = false,
       bool is_as_strided = false,
       int64_t out_idx = 0)
-      : out_index(out_idx),
+      : forward_fn(std::move(forward)),
+        reverse_fn(std::move(reverse)),
+        out_index(out_idx),
         is_multi_output(is_multi_output),
         is_as_strided(is_as_strided),
         has_symbolic_inputs(has_symbolic_inputs) {}
 
-  virtual ~ViewMeta() = default;
-
-  virtual Tensor forward(const Tensor& base) = 0;
-  virtual Tensor reverse(const Tensor& base, const Tensor& mutated_view) = 0;
-
+  std::function<Tensor(const Tensor&, int64_t)> forward_fn;
+  std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse_fn;
   // See Note [out_idx in ViewMeta]
   int64_t out_index;
 
@@ -102,17 +57,10 @@ struct ViewMeta {
   // Tells us if this view operation has any symbolic inputs
   bool has_symbolic_inputs;
 
-  // Returns a new ViewMeta with the same forward/reverse
+  // Returns a copy of the current ViewMeta, if out_idx matches the current
+  // out_index. Otherwise, returns a new ViewMeta with the same forward/reverse
   // functions, but a new out index.
-  //
-  // This method should be implemented by those `ViewMeta` that have more than
-  // one output.
-  virtual std::shared_ptr<ViewMeta> to_out_index(int64_t out_index) {
-    TORCH_CHECK_NOT_IMPLEMENTED(
-        false,
-        "ViewMeta::to_out_index not implemented. ",
-        "Likely because there's only one output.");
-  }
+  ViewMeta to_out_idx(int64_t out_idx);
 };
 
 // FunctionalStorageImpl is a subclass of StorageImpl used by the
@@ -145,14 +93,14 @@ struct TORCH_API FunctionalStorageImpl : public c10::StorageImpl {
     // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
     const at::Tensor new_val;
     // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
-    const std::vector<std::shared_ptr<ViewMeta>> view_metas;
+    const std::vector<ViewMeta> view_metas;
   };
 
   explicit FunctionalStorageImpl(const Tensor& value);
 
   void add_update(
       const Tensor& updated_val,
-      const std::vector<std::shared_ptr<ViewMeta>>& view_metas);
+      const std::vector<ViewMeta>& view_metas);
   bool apply_updates();
   const Tensor& base() {
     return base_;

aten/src/ATen/FunctionalTensorWrapper.cpp

@@ -129,19 +129,17 @@ void FunctionalTensorWrapper::freeze_storage() const {
 // - view_value: The output tensor that we need to wrap.
 // - base: The "base" of the view that `view_value` was generated from.
 // See Note [Functionalization: Alias Removal Part 2] for more details on the mutation replay logic.
-FunctionalTensorWrapper::FunctionalTensorWrapper(
-    const Tensor& view_value,
-    const FunctionalTensorWrapper* base,
-    const std::shared_ptr<functionalization::ViewMeta>& meta)
-    : c10::TensorImpl(
-          c10::DispatchKeySet(DispatchKey::Functionalize),
-          view_value.dtype(),
-          base->storage().data_ptr().device()),
-      value_(view_value),
-      is_multi_output_view_(
-          base->is_multi_output_view_ || meta->is_multi_output),
-      was_storage_changed_(base->was_storage_changed_),
-      is_symbolic_(base->is_symbolic_) {
+FunctionalTensorWrapper::FunctionalTensorWrapper(const Tensor& view_value, const FunctionalTensorWrapper* base, const functionalization::ViewMeta& meta)
+  : c10::TensorImpl(
+      c10::DispatchKeySet(DispatchKey::Functionalize),
+      view_value.dtype(),
+      view_value.device()
+    ),
+    value_(view_value),
+    is_multi_output_view_(base->is_multi_output_view_ || meta.is_multi_output),
+    was_storage_changed_(base->was_storage_changed_),
+    is_symbolic_(base->is_symbolic_)
+{
   TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(value_));
   TORCH_INTERNAL_ASSERT(!value_.key_set().has(c10::DispatchKey::Functionalize));
   set_constructor_metadata();
@@ -150,10 +148,11 @@ FunctionalTensorWrapper::FunctionalTensorWrapper(
       view_metas_ = base->view_metas_;  // copy
   }
   view_metas_.push_back(meta);
-  maybe_mark_symbolic(meta.get());
+  maybe_mark_symbolic(meta);
   storage_ = base->storage_; // alias this tensor's storage with the base tensor's
 }
 
+
 functionalization::FunctionalStorageImpl* FunctionalTensorWrapper::functional_storage_impl() const {
   return static_cast<functionalization::FunctionalStorageImpl*>(storage_.unsafeGetStorageImpl());
 }
@@ -177,18 +176,18 @@ bool FunctionalTensorWrapper::is_up_to_date() const {
 }
 
 // See Note [Functionalization Pass - Inplace View Ops]
-void FunctionalTensorWrapper::mutate_view_meta(const std::shared_ptr<at::functionalization::ViewMeta>& meta) {
+void FunctionalTensorWrapper::mutate_view_meta(const at::functionalization::ViewMeta& meta) {
   view_metas_.push_back(meta);
   // Manually track the fact that this tensor received a metadata mutation!
   has_metadata_mutation_ = true;
   // Mark this tensor as being symbolic if there are any symbolic inputs used by the view operation.
-  maybe_mark_symbolic(meta.get());
+  maybe_mark_symbolic(meta);
   // Note [Functionalization Pass - Inplace View Ops]
   // So, these ops are special - they're mutation AND view ops. They get special codegen.
   // An example is transpose_, e.g. `a.transpose_()`
   // Calling transpose_() should ensure that a gets an alias, and append the new ViewMeta to a's current list of ViewMetas.
   at::AutoDispatchSkipFunctionalize guard;
-  value_ = meta->forward(value_);
+  value_ = meta.forward_fn(value_, meta.out_index);
   TORCH_INTERNAL_ASSERT(!value_.key_set().has(c10::DispatchKey::Functionalize));
 }
 
@@ -369,8 +368,15 @@ void FunctionalTensorWrapper::sync_() {
   regenerate_from_base();
 }
 
-const std::vector<std::shared_ptr<functionalization::ViewMeta>>& FunctionalTensorWrapper::view_metas() const {
-  return view_metas_;
+Tensor FunctionalTensorWrapper::apply_view_metas(const Tensor& base) {
+  auto t = base;
+
+  // Reapply views to get the viewed tensor from the base in alias_
+  for (auto& view_meta: view_metas_) {
+    t = view_meta.forward_fn(t, view_meta.out_index);
+  }
+
+  return t;
 }
 
 void FunctionalTensorWrapper::regenerate_from_base() {
@@ -379,7 +385,7 @@ void FunctionalTensorWrapper::regenerate_from_base() {
   auto t = storage_impl->base();
 
   TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
-  t = at::functionalization::impl::apply_view_meta_sequence(t, view_metas_);
+  t = apply_view_metas(t);
   TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
 
   replace_(t, /*from_lazy_regenerate=*/true);
@@ -479,10 +485,7 @@ void FunctionalTensorWrapper::shallow_copy_from(const c10::intrusive_ptr<TensorI
 
 
 c10::Device FunctionalTensorWrapper::device_custom() const {
-  // The storage pointer already uses the underlying tensor custom device (if
-  // applicable) to extract the device. So, we dont have to recurse again by
-  // doing value_.unsafeGetTensorImpl()->device().
-  return storage().data_ptr().device();
+  return value_.unsafeGetTensorImpl()->device();
 }
 at::IntArrayRef FunctionalTensorWrapper::sizes_custom() const {
   return value_.unsafeGetTensorImpl()->sizes();
@@ -721,11 +724,11 @@ bool isFunctionalTensor(const std::optional<Tensor>& t) {
 }
 
 bool isFunctionalTensor(const c10::List<::std::optional<Tensor>>& t_list) {
-  if (t_list.empty()) { return false; }
+  if (t_list.empty()) return false;
   auto functional_count = 0;
   for (const auto i : c10::irange(t_list.size())) {
     auto const & e= t_list[i];
-    if (!e.has_value() || !e->defined()) { continue; }
+    if (!e.has_value() || !e->defined()) continue;
     if (isFunctionalTensor(e)) {
       ++functional_count;
     }
@@ -735,10 +738,10 @@ bool isFunctionalTensor(const c10::List<::std::optional<Tensor>>& t_list) {
 
 template <typename T>
 static bool isFunctionalTensorIListRef(c10::IListRef<T> list) {
-  if (list.size() == 0) { return false; }
+  if (list.size() == 0) return false;
   auto functional_count = 0;
   for (const auto& tensor : list) {
-    if (!tensor.defined()) { continue; }
+    if (!tensor.defined()) continue;
     if (isFunctionalTensor(tensor)) {
       ++functional_count;
     }
@@ -756,28 +759,20 @@ void freeze_functional_tensor(const Tensor& tensor) {
   functional_base_impl->freeze_storage();
 }
 
-Tensor create_functional_tensor_with_view_meta(
-    const at::Tensor& view_to_wrap,
-    const at::Tensor& base,
-    const std::shared_ptr<functionalization::ViewMeta>& meta,
-    int64_t out_idx) {
+Tensor create_functional_tensor_with_view_meta(const at::Tensor& view_to_wrap, const at::Tensor& base, functionalization::ViewMeta meta, int64_t out_idx) {
   TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(view_to_wrap));
   TORCH_INTERNAL_ASSERT(at::functionalization::impl::isFunctionalTensor(base));
   auto functional_base_impl = at::functionalization::impl::unsafeGetFunctionalWrapper(base);
-  auto meta_ = meta;
   if (out_idx != 0) {
     // Note [out_idx in ViewMeta]
     // When a view op outputs multiple tensors, each output needs its own separate ViewMeta.
     // Each ViewMeta also tracks the index of the particular output tensor, which is needed in the reverse function.
-    meta_ = meta->to_out_index(out_idx);
+    meta = meta.to_out_idx(out_idx);
   }
-  return at::detail::make_tensor<FunctionalTensorWrapper>(view_to_wrap, functional_base_impl, meta_);
+  return at::detail::make_tensor<FunctionalTensorWrapper>(view_to_wrap, functional_base_impl, meta);
 }
 
-std::vector<Tensor> create_functional_tensor_with_view_meta(
-    ITensorListRef view_to_wrap,
-    const at::Tensor& base,
-    const std::shared_ptr<functionalization::ViewMeta>& meta) {
+std::vector<Tensor> create_functional_tensor_with_view_meta(ITensorListRef view_to_wrap, const at::Tensor& base, const functionalization::ViewMeta& meta) {
   std::vector<Tensor> outputs(view_to_wrap.size());
   int64_t i = 0;
   for (const auto& tensor : view_to_wrap) {
@@ -787,22 +782,12 @@ std::vector<Tensor> create_functional_tensor_with_view_meta(
   return outputs;
 }
 
-void mutate_view_meta(const at::Tensor& self, const std::shared_ptr<functionalization::ViewMeta>& meta) {
+void mutate_view_meta(const at::Tensor& self, const functionalization::ViewMeta& meta) {
   TORCH_INTERNAL_ASSERT(at::functionalization::impl::isFunctionalTensor(self));
   auto self_impl = at::functionalization::impl::unsafeGetFunctionalWrapper(self);
   self_impl->mutate_view_meta(meta);
 }
 
-Tensor apply_view_meta_sequence(
-    const Tensor& base,
-    const std::vector<std::shared_ptr<functionalization::ViewMeta>>& sequence) {
-  Tensor r = base;
-  for (auto& vm : sequence) {
-    r = vm->forward(r);
-  }
-  return r;
-}
-
 // Note [Propagating strides in the functionalization pass]
 // In order to properly compute stride information, the functionalization pass
 // calls each {view} reference implementations with meta tensors.
@@ -896,7 +881,7 @@ void functionalize_op_helper(const c10::OperatorHandle& op, torch::jit::Stack* s
     const auto& ivalue = returns[idx];
     if (ivalue.isTensor()) {
       const auto& t = ivalue.toTensor();
-      if (!t.defined()) { continue; }
+      if (!t.defined()) continue;
       at::functionalization::impl::sync(t);
       auto t_new = c10::IValue(at::functionalization::impl::from_functional_tensor(t));
       (*stack)[returns_begin + idx] = t_new;

aten/src/ATen/FunctionalTensorWrapper.h

@@ -56,7 +56,7 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
   explicit FunctionalTensorWrapper(
       const Tensor& view_value,
       const FunctionalTensorWrapper* base,
-      const std::shared_ptr<functionalization::ViewMeta>& meta);
+      const functionalization::ViewMeta& meta);
 
   // Get the underlying, actual tensor, that doesn't know anything about
   // functionalization.
@@ -99,17 +99,17 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
         ->are_all_mutations_under_no_grad_or_inference_mode();
   }
 
-  void maybe_mark_symbolic(functionalization::ViewMeta* meta) {
-    is_symbolic_ = is_symbolic_ | meta->has_symbolic_inputs;
+  void maybe_mark_symbolic(const functionalization::ViewMeta& meta) {
+    is_symbolic_ = is_symbolic_ | meta.has_symbolic_inputs;
   }
 
   bool is_symbolic() const {
     return is_symbolic_;
   }
 
-  // Retrieves the ViewMeta sequence of this tensor.
-  const std::vector<std::shared_ptr<functionalization::ViewMeta>>& view_metas()
-      const;
+  // Runs the forward_fn of every ViewMeta collected in the current instance
+  // to some other base.
+  Tensor apply_view_metas(const Tensor& base);
 
   // Sync's the underlying tensor with its alias, if it's out of date. This
   // involves two steps: 1) Apply any pending updates/mutations to the alias 2)
@@ -146,8 +146,7 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
   // from the base tensor. This method is used by inplace-view ops like
   // transpose_. It appends a ViewMeta to the existing stack, and refreshes the
   // tensor by replaying the views off of the alias.
-  void mutate_view_meta(
-      const std::shared_ptr<at::functionalization::ViewMeta>& meta);
+  void mutate_view_meta(const at::functionalization::ViewMeta& meta);
 
   // Custom implementation of self.set_(src)
   void set__impl(const FunctionalTensorWrapper* other);
@@ -286,7 +285,7 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
   bool is_symbolic_ = false;
 
   size_t generation_ = 0;
-  std::vector<std::shared_ptr<at::functionalization::ViewMeta>> view_metas_;
+  std::vector<at::functionalization::ViewMeta> view_metas_;
 
  protected:
   static void copy_tensor_metadata(
@@ -378,20 +377,16 @@ TORCH_API void propagate_xla_data_direct(
 Tensor create_functional_tensor_with_view_meta(
     const Tensor& view_to_wrap,
     const Tensor& base,
-    const std::shared_ptr<functionalization::ViewMeta>& meta,
+    functionalization::ViewMeta meta,
     int64_t out_idx = 0);
 std::vector<Tensor> create_functional_tensor_with_view_meta(
     ITensorListRef view_to_wrap,
     const Tensor& base,
-    const std::shared_ptr<functionalization::ViewMeta>& meta);
+    const functionalization::ViewMeta& meta);
 
 void mutate_view_meta(
     const Tensor& self,
-    const std::shared_ptr<functionalization::ViewMeta>& meta);
-
-TORCH_API Tensor apply_view_meta_sequence(
-    const Tensor& base,
-    const std::vector<std::shared_ptr<functionalization::ViewMeta>>& sequence);
+    const functionalization::ViewMeta& meta);
 
 void set_sizes_strides_offset(const Tensor& out, const Tensor& meta_out);
 void set_sizes_strides_offset(

aten/src/ATen/FunctionalizeFallbackKernel.cpp

@@ -1,5 +1,3 @@
-#include <ATen/FunctionalizeFallbackKernel.h>
-
 #include <ATen/core/dispatch/Dispatcher.h>
 #include <ATen/core/LegacyTypeDispatch.h>
 #include <ATen/EmptyTensor.h>
@@ -9,6 +7,7 @@
 #include <torch/library.h>
 #include <c10/util/irange.h>
 #include <c10/util/strides.h>
+#include <ATen/EmptyTensor.h>
 
 #ifndef AT_PER_OPERATOR_HEADERS
 #include <ATen/ATen.h>
@@ -29,31 +28,6 @@
 #include <utility>
 #endif
 
-namespace at::functionalization {
-
-Tensor resize__ViewMeta::forward(const Tensor& base) {
-  if (reapply_views) {
-    return base.as_strided(size, c10::contiguous_strides(size));
-  } else {
-    return at::as_strided_copy(base, size, c10::contiguous_strides(size));
-  }
-}
-
-Tensor resize__ViewMeta::reverse(const Tensor& base, const Tensor& mutated_view) {
-  return base.as_strided_scatter(
-      mutated_view, size, c10::contiguous_strides(size));
-}
-
-Tensor _unsafe_view_ViewMeta::forward(const Tensor& base) {
-  return at::_unsafe_view_symint(base, size);
-}
-
-Tensor _unsafe_view_ViewMeta::reverse(const Tensor& base, const Tensor& mutated_view) {
-  return at::_unsafe_view_symint(mutated_view, base.sym_sizes());
-}
-
-} // namespace at::functionalization
-
 namespace {
   void functionalizeFallback(const c10::OperatorHandle& op, c10::DispatchKeySet dispatchKeySet [[maybe_unused]], torch::jit::Stack* stack) {
     const auto& schema = op.schema();
@@ -132,9 +106,7 @@ namespace {
       const auto& ivalue = returns[idx];
       if (ivalue.isTensor() && should_wrap_outputs) {
         const auto& t = ivalue.toTensor();
-        if (!t.defined()) {
-          continue;
-        }
+        if (!t.defined()) continue;
         auto t_new = c10::IValue(at::functionalization::impl::to_functional_tensor(t));
         (*stack)[returns_begin + idx] = t_new;
       } else if (ivalue.isTensorList() && should_wrap_outputs) {
@@ -197,8 +169,19 @@ static const at::Tensor & resize__functionalization(c10::DispatchKeySet dispatch
   // The output of resizing is equivalent to taking a slice of a larger tensor.
   // We have to emulate this "slicing" with an as_strided call.
   auto reapply_views = at::functionalization::impl::getFunctionalizationReapplyViewsTLS();
-  auto view_meta = std::make_shared<at::functionalization::resize__ViewMeta>(
-      reapply_views, size.vec());
+  at::functionalization::ViewMeta view_meta = at::functionalization::ViewMeta(
+    [reapply_views = reapply_views, size = size.vec()](const at::Tensor & base, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
+      if (reapply_views) {
+        return base.as_strided(size, c10::contiguous_strides(size));
+      } else {
+        return at::as_strided_copy(base, size, c10::contiguous_strides(size));
+      }
+    },
+    [size = size.vec()](const at::Tensor & base, const at::Tensor & mutated_view, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
+      return base.as_strided_scatter(mutated_view, size, c10::contiguous_strides(size));
+    },
+    /*has_symbolic_inputs=*/false
+  );
   at::functionalization::impl::mutate_view_meta(self, view_meta);
   return self;
 }
@@ -317,11 +300,17 @@ static at::Tensor _unsafe_view_functionalize(const at::Tensor & self, at::SymInt
     tmp_output = at::_unsafe_view_symint(self_, size);
   }
 
-  bool has_symbolic_inputs = std::any_of(
-      size.begin(), size.end(), [=](auto& s) { return s.is_symbolic(); });
-  auto view_meta =
-      std::make_shared<at::functionalization::_unsafe_view_ViewMeta>(
-          has_symbolic_inputs, size.vec());
+  bool has_symbolic_inputs = std::any_of(size.begin(), size.end(), [=](auto& s) { return s.is_symbolic(); });
+
+  at::functionalization::ViewMeta view_meta = at::functionalization::ViewMeta(
+    [size = size.vec()](const at::Tensor & base, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
+      return at::_unsafe_view_symint(base, size);
+    },
+    [size = size.vec()](const at::Tensor & base, const at::Tensor & mutated_view, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
+      return at::_unsafe_view_symint(mutated_view, base.sym_sizes());
+    },
+    /*has_symbolic_inputs=*/has_symbolic_inputs
+  );
 
   auto out = at::functionalization::impl::create_functional_tensor_with_view_meta(tmp_output, self, std::move(view_meta));
   // See  Note [Propagating strides in the functionalization pass]

aten/src/ATen/FunctionalizeFallbackKernel.h

@@ -1,58 +0,0 @@
-#pragma once
-
-#include <ATen/FunctionalStorageImpl.h>
-
-namespace at::functionalization {
-
-// `ViewMeta` implementation for `resize_` operation.
-struct TORCH_API resize__ViewMeta : public ViewMeta {
-  FUNCTIONALIZATION_VIEWMETA_NAME(resize__ViewMeta)
-  FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(
-      bool /* reapply_views */,
-      const std::vector<int64_t>&);
-
-  resize__ViewMeta(const SerializableTuple& tpl)
-      : resize__ViewMeta(std::get<0>(tpl), std::get<1>(tpl)) {}
-
-  resize__ViewMeta(bool reapply_views, const std::vector<int64_t>& size)
-      : ViewMeta(/*has_symbolic_inputs=*/false),
-        reapply_views(reapply_views),
-        size(size) {}
-
-  Tensor forward(const Tensor& base) override;
-  Tensor reverse(const Tensor& base, const Tensor& mutated_view) override;
-
-  SerializableTuple to_serializable_tuple() {
-    return std::make_tuple(reapply_views, size);
-  }
-
-  bool reapply_views;
-  std::vector<int64_t> size;
-};
-
-// `ViewMeta` implementation for `_unsafe_view` operation.
-struct TORCH_API _unsafe_view_ViewMeta : public ViewMeta {
-  FUNCTIONALIZATION_VIEWMETA_NAME(_unsafe_view_ViewMeta)
-  FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(
-      bool /* has_symbolic_inputs */,
-      const std::vector<c10::SymInt>&);
-
-  _unsafe_view_ViewMeta(const SerializableTuple& tpl)
-      : _unsafe_view_ViewMeta(std::get<0>(tpl), std::get<1>(tpl)) {}
-
-  _unsafe_view_ViewMeta(
-      bool has_symbolic_inputs,
-      const std::vector<c10::SymInt>& size)
-      : ViewMeta(has_symbolic_inputs), size(size) {}
-
-  Tensor forward(const Tensor& base) override;
-  Tensor reverse(const Tensor& base, const Tensor& mutated_view) override;
-
-  SerializableTuple to_serializable_tuple() {
-    return std::make_tuple(has_symbolic_inputs, size);
-  }
-
-  std::vector<c10::SymInt> size;
-};
-
-} // namespace at::functionalization

aten/src/ATen/InferSize.h

@@ -45,39 +45,7 @@ inline void infer_size_impl(
     }
   }
 
-  if (infer_dim) {
-    // numel is the product of known sizes, it has to be divisible by newsize.
-    // and newsize should be positive unless newsize == numel (we throw
-    // different) error message in that case.
-    if constexpr (std::is_same_v<NumelType, c10::SymInt>) {
-      auto v = newsize.maybe_as_int();
-      if (v and *v == 0) {
-        // Avoid div by 0 when sym_eq(numel % newsize, 0) is constructed!
-        // which may happen when newsize is not a symbol! if its a symbol
-        // division won't happen anyway during compile.
-        TORCH_MAYBE_SYM_CHECK(
-            numel == newsize,
-            "shape '",
-            shape,
-            "' is invalid for input of size ",
-            numel);
-      } else {
-        auto cond = sym_gt(newsize, 0)
-                        .sym_and(sym_eq(numel % newsize, 0))
-                        .sym_or(sym_eq(numel, newsize));
-        TORCH_MAYBE_SYM_CHECK(
-            cond, "shape '", shape, "' is invalid for input of size ", numel);
-      }
-
-    } else {
-      TORCH_CHECK(
-          (newsize > 0 && (numel % newsize == 0)) || numel == newsize,
-          "shape '",
-          shape,
-          "' is invalid for input of size ",
-          numel);
-    }
-
+  auto set_infer_dim = [&]() {
     // We have a degree of freedom here to select the dimension size; follow
     // NumPy semantics and just bail.  However, a nice error message is needed
     // because users often use `view` as a way to flatten & unflatten
@@ -86,15 +54,19 @@ inline void infer_size_impl(
     // works yet
     //   empty_tensor.view(-1, 0)
     // doesn't.
-    TORCH_MAYBE_SYM_CHECK(
+    TORCH_CHECK(
         newsize != 0,
         "cannot reshape tensor of 0 elements into shape ",
         shape,
         " because the unspecified dimension size -1 can be any "
         "value and is ambiguous");
-
     res[*infer_dim] = numel / newsize;
     return;
+  };
+
+  if (infer_dim && newsize > 0 && numel % newsize == 0) {
+    set_infer_dim();
+    return;
   }
 
   TORCH_MAYBE_SYM_CHECK(
@@ -103,6 +75,9 @@ inline void infer_size_impl(
       shape,
       "' is invalid for input of size ",
       numel);
+  if (infer_dim) {
+    set_infer_dim();
+  }
 }
 
 inline std::vector<int64_t> infer_size(IntArrayRef shape, int64_t numel) {

aten/src/ATen/Version.cpp

@@ -103,9 +103,7 @@ std::string get_cpu_capability() {
 #elif defined(HAVE_ZVECTOR_CPU_DEFINITION)
     case native::CPUCapability::ZVECTOR:
       return "Z VECTOR";
-#elif defined(HAVE_SVE_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
-    case native::CPUCapability::SVE128:
-      return "SVE128";
+#elif defined(HAVE_SVE256_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
     case native::CPUCapability::SVE256:
       return "SVE256";
 #else

aten/src/ATen/core/PythonOpRegistrationTrampoline.cpp

@@ -1,22 +1,32 @@
 #include <ATen/core/PythonOpRegistrationTrampoline.h>
-#include <c10/core/impl/PyInterpreterHooks.h>
 
-// TODO: delete this
 namespace at::impl {
 
-c10::impl::PyInterpreter* PythonOpRegistrationTrampoline::interpreter_ = nullptr;
+// The strategy is that all python interpreters attempt to register themselves
+// as the main interpreter, but only one wins.  Only that interpreter is
+// allowed to interact with the C++ dispatcher.  Furthermore, when we execute
+// logic on that interpreter, we do so hermetically, never setting pyobj field
+// on Tensor.
+
+std::atomic<c10::impl::PyInterpreter*>
+    PythonOpRegistrationTrampoline::interpreter_{nullptr};
 
 c10::impl::PyInterpreter* PythonOpRegistrationTrampoline::getInterpreter() {
-  return c10::impl::getGlobalPyInterpreter();
+  return PythonOpRegistrationTrampoline::interpreter_.load();
 }
 
 bool PythonOpRegistrationTrampoline::registerInterpreter(
     c10::impl::PyInterpreter* interp) {
-  if (interpreter_ != nullptr) {
+  c10::impl::PyInterpreter* expected = nullptr;
+  interpreter_.compare_exchange_strong(expected, interp);
+  if (expected != nullptr) {
+    // This is the second (or later) Python interpreter, which means we need
+    // non-trivial hermetic PyObject TLS
+    c10::impl::HermeticPyObjectTLS::init_state();
     return false;
+  } else {
+    return true;
   }
-  interpreter_ = interp;
-  return true;
 }
 
 } // namespace at::impl

aten/src/ATen/core/PythonOpRegistrationTrampoline.h

@@ -2,21 +2,19 @@
 
 #include <ATen/core/dispatch/Dispatcher.h>
 
-// TODO: We can get rid of this
+// TODO: this can probably live in c10
 
 
 namespace at::impl {
 
-// Manages the single Python interpreter instance for PyTorch.
 class TORCH_API PythonOpRegistrationTrampoline final {
-  static c10::impl::PyInterpreter* interpreter_;
+  static std::atomic<c10::impl::PyInterpreter*> interpreter_;
 
 public:
-  // Register the Python interpreter. Returns true on first registration,
-  // false if an interpreter was already registered.
+  //  Returns true if you successfully registered yourself (that means
+  //  you are in the hot seat for doing the operator registrations!)
   static bool registerInterpreter(c10::impl::PyInterpreter*);
 
-  // Returns the registered interpreter via the global PyInterpreter hooks.
   // Returns nullptr if no interpreter has been registered yet.
   static c10::impl::PyInterpreter* getInterpreter();
 };

aten/src/ATen/core/jit_type.h

@@ -1234,7 +1234,7 @@ struct TORCH_API TupleType : public NamedType {
   std::shared_ptr<FunctionSchema> schema_;
 };
 
-// the common supertype of all Enums, only used in operator registration.
+// the common supertype of all Enums, only used in operator registraion.
 // EnumType <: AnyEnumType for all Enums
 struct AnyEnumType;
 using AnyEnumTypePtr = SingletonTypePtr<AnyEnumType>;

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

@@ -102,31 +102,8 @@ struct VecReduceAllSIMD<float, Op> {
 #endif // defined(__GNUC__) && (__GNUC__ > 5) && !defined(_MSC_VER) &&
        // !defined(C10_MOBILE)
 
-#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
+#if defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
+    !defined(CPU_CAPABILITY_SVE)
 template <typename Op>
 struct VecReduceAllSIMD<float, Op> {
   static inline float apply(
@@ -163,8 +140,35 @@ 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

@@ -1,21 +1,9 @@
 #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

@@ -2,6 +2,7 @@
 
 #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>