revert expectedFailure on TestFSDPWithEP::test_e2e

test_gather_object* require exact world size (4) to succeed
Likely, all NVSHMEM Triton tests would fail. Skipping all for now.

.ci/docker/build.sh

@@ -113,7 +113,6 @@ case "$tag" in
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
-    INSTALL_MINGW=yes
     ;;
   pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
     CUDA_VERSION=13.0.0
@@ -195,16 +194,13 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-xpu-n-py3 | pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks)
+  pytorch-linux-jammy-xpu-n-py3)
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     XPU_VERSION=2025.2
     NINJA_VERSION=1.9.0
     TRITON=yes
-    if [[ $tag =~ "benchmarks" ]]; then
-      INDUCTOR_BENCHMARKS=yes
-    fi
     ;;
   pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10
@@ -365,7 +361,6 @@ docker build \
        --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
        --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
-       --build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \
        -t "$tmp_tag" \
        "$@" \

.ci/docker/common/install_conda.sh

@@ -49,20 +49,12 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
     export SYSROOT_DEP="sysroot_linux-64=2.17"
   fi
 
-# Install correct Python version
-# Also ensure sysroot is using a modern GLIBC to match system compilers
-if [ "$ANACONDA_PYTHON_VERSION" = "3.14" ]; then
-  as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
-             python="3.14.0" \
-             ${SYSROOT_DEP} \
-             -c conda-forge
-else
   # Install correct Python version
   # Also ensure sysroot is using a modern GLIBC to match system compilers
   as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
              python="$ANACONDA_PYTHON_VERSION" \
              ${SYSROOT_DEP}
-fi
+
   # libstdcxx from conda default channels are too old, we need GLIBCXX_3.4.30
   # which is provided in libstdcxx 12 and up.
   conda_install libstdcxx-ng=12.3.0 --update-deps -c conda-forge

.ci/docker/common/install_cpython.sh

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

.ci/docker/common/install_cuda.sh

@@ -10,7 +10,7 @@ else
   arch_path='sbsa'
 fi
 
-NVSHMEM_VERSION=3.4.5
+NVSHMEM_VERSION=3.3.24
 
 function install_cuda {
   version=$1
@@ -150,7 +150,7 @@ function install_130 {
   CUDNN_VERSION=9.13.0.50
   echo "Installing CUDA 13.0 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
   # install CUDA 13.0 in the same container
-  install_cuda 13.0.2 cuda_13.0.2_580.95.05_linux
+  install_cuda 13.0.0 cuda_13.0.0_580.65.06_linux
 
   # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
   install_cudnn 13 $CUDNN_VERSION

.ci/docker/common/install_mingw.sh

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

.ci/docker/common/install_onnx.sh

@@ -19,8 +19,8 @@ pip_install \
   transformers==4.36.2
 
 pip_install coloredlogs packaging
-pip_install onnxruntime==1.23.1
-pip_install onnxscript==0.5.4
+pip_install onnxruntime==1.23.0
+pip_install onnxscript==0.5.3
 
 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/

.ci/docker/common/install_rocm.sh

@@ -40,7 +40,11 @@ EOF
 
     # Default url values
     rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
+    amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
+
+    # Add amdgpu repository
     UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
+    echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
 
     # Add rocm repository
     wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -

.ci/docker/common/install_rocm_magma.sh

@@ -12,8 +12,8 @@ function do_install() {
 
     rocm_version_nodot=${rocm_version//./}
 
-    # post merge of https://github.com/icl-utk-edu/magma/pull/65
-    MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
+    # https://github.com/icl-utk-edu/magma/pull/65
+    MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
     magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"
 
     rocm_dir="/opt/rocm"

.ci/docker/libtorch/build.sh

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

.ci/docker/manywheel/build.sh

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

.ci/docker/requirements-ci.txt

@@ -138,12 +138,10 @@ numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
 #test_binary_ufuncs.py
 numpy==1.22.4; 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" and python_version < "3.14"
-numpy==2.3.4; python_version >= "3.14"
+numpy==2.1.2; python_version >= "3.13"
 
 pandas==2.0.3; python_version < "3.13"
-pandas==2.2.3; python_version >= "3.13" and python_version < "3.14"
-pandas==2.3.3; python_version >= "3.14"
+pandas==2.2.3; python_version >= "3.13"
 
 #onnxruntime
 #Description: scoring engine for Open Neural Network Exchange (ONNX) models
@@ -155,8 +153,7 @@ opt-einsum==3.3
 #Pinned versions: 3.3
 #test that import: test_linalg.py
 
-optree==0.13.0 ; python_version < "3.14"
-optree==0.17.0 ; python_version >= "3.14"
+optree==0.13.0
 #Description: A library for tree manipulation
 #Pinned versions: 0.13.0
 #test that import: test_vmap.py, test_aotdispatch.py, test_dynamic_shapes.py,
@@ -255,8 +252,7 @@ scikit-image==0.22.0
 #test that import:
 
 scipy==1.10.1 ; python_version <= "3.11"
-scipy==1.14.1 ; python_version > "3.11" and python_version < "3.14"
-scipy==1.16.2 ; python_version >= "3.14"
+scipy==1.14.1 ; python_version >= "3.12"
 # Pin SciPy because of failing distribution tests (see #60347)
 #Description: scientific python
 #Pinned versions: 1.10.1
@@ -328,8 +324,7 @@ pywavelets==1.7.0 ; python_version >= "3.12"
 #Pinned versions: 1.4.1
 #test that import:
 
-lxml==5.3.0 ; python_version < "3.14"
-lxml==6.0.2 ; python_version >= "3.14"
+lxml==5.3.0
 #Description: This is a requirement of unittest-xml-reporting
 
 PyGithub==2.3.0
@@ -339,14 +334,12 @@ sympy==1.13.3
 #Pinned versions:
 #test that import:
 
-onnx==1.19.1 ; python_version < "3.14"
-# Unpin once Python 3.14 is supported. See  onnxruntime issue 26309.
-onnx==1.18.0 ; python_version == "3.14"
+onnx==1.18.0
 #Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
 
-onnxscript==0.5.4
+onnxscript==0.5.3
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -366,7 +359,7 @@ pwlf==2.2.1
 #test that import: test_sac_estimator.py
 
 # To build PyTorch itself
-pyyaml==6.0.3
+pyyaml==6.0.2
 pyzstd
 setuptools==78.1.1
 packaging==23.1

.ci/docker/ubuntu-xpu/Dockerfile

@@ -54,15 +54,12 @@ ENV OPENSSL_DIR /opt/openssl
 RUN rm install_openssl.sh
 
 ARG INDUCTOR_BENCHMARKS
-ARG ANACONDA_PYTHON_VERSION
-ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt
 
 # Install XPU Dependencies
 ARG XPU_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -100,16 +100,9 @@ COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
-# Only build aoti cpp tests when INDUCTOR_BENCHMARKS is set to True
-ENV BUILD_AOT_INDUCTOR_TEST ${INDUCTOR_BENCHMARKS}
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
 RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 
-ARG INSTALL_MINGW
-COPY ./common/install_mingw.sh install_mingw.sh
-RUN if [ -n "${INSTALL_MINGW}" ]; then bash ./install_mingw.sh; fi
-RUN rm install_mingw.sh
-
 ARG TRITON
 ARG TRITON_CPU
 

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

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

.ci/lumen_cli/pyproject.toml

@@ -6,7 +6,7 @@ dependencies = [
     "GitPython==3.1.45",
     "docker==7.1.0",
     "pytest==7.3.2",
-    "uv==0.9.6"
+    "uv==0.8.6"
 ]
 
 [tool.setuptools]

.ci/magma-rocm/Makefile

@@ -1,7 +1,7 @@
 SHELL=/usr/bin/env bash
 
 DOCKER_CMD ?= docker
-DESIRED_ROCM ?= 7.1
+DESIRED_ROCM ?= 7.0
 DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
 PACKAGE_NAME = magma-rocm
 # inherit this from underlying docker image, do not pass this env var to docker
@@ -16,7 +16,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma-rocm/build_magma.sh
 
 .PHONY: all
-all: magma-rocm71
 all: magma-rocm70
 all: magma-rocm64
 
@@ -25,11 +24,6 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
-.PHONY: magma-rocm71
-magma-rocm71: DESIRED_ROCM := 7.1
-magma-rocm71:
-	$(DOCKER_RUN)
-
 .PHONY: magma-rocm70
 magma-rocm70: DESIRED_ROCM := 7.0
 magma-rocm70:

.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)"
 
-# post merge of https://github.com/icl-utk-edu/magma/pull/65
-MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
+# https://github.com/icl-utk-edu/magma/pull/65
+MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
 
 # 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/icl-utk-edu/magma
+git clone https://github.com/jeffdaily/magma
 pushd magma
 git checkout ${MAGMA_VERSION}
 popd

.ci/pytorch/build.sh

@@ -426,7 +426,7 @@ fi
 if [[ "$BUILD_ENVIRONMENT" != *libtorch* && "$BUILD_ENVIRONMENT" != *bazel* ]]; then
   # export test times so that potential sharded tests that'll branch off this build will use consistent data
   # don't do this for libtorch as libtorch is C++ only and thus won't have python tests run on its build
-  PYTHONPATH=. python tools/stats/export_test_times.py
+  python tools/stats/export_test_times.py
 fi
 # don't do this for bazel or s390x or riscv64 as they don't use sccache
 if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then

.ci/pytorch/test.sh

@@ -460,37 +460,31 @@ test_inductor_shard() {
     --verbose
 }
 
-test_inductor_aoti_cpp() {
+test_inductor_aoti() {
+  # docker build uses bdist_wheel which does not work with test_aot_inductor
+  # TODO: need a faster way to build
   if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
     # We need to hipify before building again
     python3 tools/amd_build/build_amd.py
   fi
   if [[ "$BUILD_ENVIRONMENT" == *sm86* ]]; then
+    BUILD_COMMAND=(TORCH_CUDA_ARCH_LIST=8.6 USE_FLASH_ATTENTION=OFF python -m pip install --no-build-isolation -v -e .)
     # TODO: Replace me completely, as one should not use conda libstdc++, nor need special path to TORCH_LIB
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="/opt/conda/envs/py_3.10/lib:${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}")
   else
+    BUILD_COMMAND=(python -m pip install --no-build-isolation -v -e .)
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}")
   fi
 
+  # aoti cmake custom command requires `torch` to be installed
+  # initialize the cmake build cache and install torch
+  /usr/bin/env "${BUILD_COMMAND[@]}"
+  # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
+  /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${BUILD_COMMAND[@]}"
+
   /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
-test_inductor_aoti_cross_compile_for_windows() {
-
-  TEST_REPORTS_DIR=$(pwd)/test/test-reports
-  mkdir -p "$TEST_REPORTS_DIR"
-
-  # Set WINDOWS_CUDA_HOME environment variable
-  WINDOWS_CUDA_HOME="$(pwd)/win-torch-wheel-extracted"
-  export WINDOWS_CUDA_HOME
-
-  echo "WINDOWS_CUDA_HOME is set to: $WINDOWS_CUDA_HOME"
-  echo "Contents:"
-  ls -lah "$(pwd)/win-torch-wheel-extracted/lib/x64/" || true
-
-  python test/inductor/test_aoti_cross_compile_windows.py -k compile --package-dir "$TEST_REPORTS_DIR" --win-torch-lib-dir "$(pwd)/win-torch-wheel-extracted/torch/lib"
-}
-
 test_inductor_cpp_wrapper_shard() {
   if [[ -z "$NUM_TEST_SHARDS" ]]; then
     echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@@ -572,8 +566,6 @@ fi
 
 if [[ "${TEST_CONFIG}" == *cpu* ]]; then
   DYNAMO_BENCHMARK_FLAGS+=(--device cpu)
-elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
-  DYNAMO_BENCHMARK_FLAGS+=(--device xpu)
 else
   DYNAMO_BENCHMARK_FLAGS+=(--device cuda)
 fi
@@ -667,8 +659,6 @@ test_perf_for_dashboard() {
     device=cuda_b200
   elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
     device=rocm
-  elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
-    device=xpu
   fi
 
   for mode in "${modes[@]}"; do
@@ -910,7 +900,7 @@ test_inductor_set_cpu_affinity(){
   export LD_PRELOAD="$JEMALLOC_LIB":"$LD_PRELOAD"
   export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1"
 
-  if [[ "$(uname -m)" != "aarch64" ]]; then
+  if [[ "${TEST_CONFIG}" != *aarch64* ]]; then
     # Use Intel OpenMP for x86
     IOMP_LIB="$(dirname "$(which python)")/../lib/libiomp5.so"
     export LD_PRELOAD="$IOMP_LIB":"$LD_PRELOAD"
@@ -924,7 +914,7 @@ test_inductor_set_cpu_affinity(){
   cores=$((cpus / thread_per_core))
 
   # Set number of cores to 16 on aarch64 for performance runs
-  if [[ "$(uname -m)" == "aarch64" && $cores -gt 16 ]]; then
+  if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
     cores=16
   fi
   export OMP_NUM_THREADS=$cores
@@ -1625,7 +1615,6 @@ test_operator_benchmark() {
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
   TEST_DIR=$(pwd)
-  ARCH=$(uname -m)
 
   test_inductor_set_cpu_affinity
 
@@ -1640,7 +1629,7 @@ test_operator_benchmark() {
   pip_install pandas
   python check_perf_csv.py \
       --actual "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
-      --expected "${ARCH}_expected_ci_operator_benchmark_eager_float32_cpu.csv"
+      --expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
 }
 
 test_operator_microbenchmark() {
@@ -1677,7 +1666,7 @@ if [[ "${TEST_CONFIG}" == *numpy_2* ]]; then
     python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
   fi
   python test/run_test.py --include dynamo/test_functions.py dynamo/test_unspec.py test_binary_ufuncs.py test_fake_tensor.py test_linalg.py test_numpy_interop.py test_tensor_creation_ops.py test_torch.py torch_np/test_basic.py
-elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]]; then
+elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
   test_linux_aarch64
 elif [[ "${TEST_CONFIG}" == *backward* ]]; then
   test_forward_backward_compatibility
@@ -1728,8 +1717,6 @@ elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
   test_inductor_micro_benchmark
-elif [[ "${TEST_CONFIG}" == *aoti_cross_compile_for_windows* ]]; then
-  test_inductor_aoti_cross_compile_for_windows
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
   install_torchvision
   id=$((SHARD_NUMBER-1))
@@ -1761,7 +1748,7 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   else
     # Do this after checkout_install_torchbench to ensure we clobber any
     # nightlies that torchbench may pull in
-    if [[ "${TEST_CONFIG}" != *cpu* && "${TEST_CONFIG}" != *xpu* ]]; then
+    if [[ "${TEST_CONFIG}" != *cpu* ]]; then
       install_torchrec_and_fbgemm
     fi
     PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
@@ -1770,7 +1757,7 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
   PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
-    test_inductor_aoti_cpp
+    test_inductor_aoti
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision

.ci/pytorch/windows/internal/install_python.bat

@@ -7,9 +7,12 @@ if "%DESIRED_PYTHON%" == "3.13t" (
     set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.13.0/python-3.13.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.13t"
+) else if "%DESIRED_PYTHON%"=="3.14" (
+    echo Python version is set to 3.14 or 3.14t
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
 ) else if "%DESIRED_PYTHON%"=="3.14t" (
     echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0-amd64.exe"
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.14t"
 ) else (

.circleci/scripts/binary_populate_env.sh

@@ -163,13 +163,8 @@ if [[ "$(uname)" != Darwin ]]; then
   MEMORY_LIMIT_MAX_JOBS=12
   NUM_CPUS=$(( $(nproc) - 2 ))
 
-  if [[ "$(uname)" == Linux ]]; then
-    # Defaults here for **binary** linux builds so they can be changed in one place
-    export MAX_JOBS=${MAX_JOBS:-$(( ${NUM_CPUS} > ${MEMORY_LIMIT_MAX_JOBS} ? ${MEMORY_LIMIT_MAX_JOBS} : ${NUM_CPUS} ))}
-  else
-    # For other builds
-    export MAX_JOBS=${NUM_CPUS}
-  fi
+  # Defaults here for **binary** linux builds so they can be changed in one place
+  export MAX_JOBS=${MAX_JOBS:-$(( ${NUM_CPUS} > ${MEMORY_LIMIT_MAX_JOBS} ? ${MEMORY_LIMIT_MAX_JOBS} : ${NUM_CPUS} ))}
 
   cat >>"$envfile" <<EOL
   export MAX_JOBS="${MAX_JOBS}"

.claude/skills/docstring/SKILL.md

@@ -1,359 +0,0 @@
----
-name: docstring
-description: Write docstrings for PyTorch functions and methods following PyTorch conventions. Use when writing or updating docstrings in PyTorch code.
----
-
-# PyTorch Docstring Writing Guide
-
-This skill describes how to write docstrings for functions and methods in the PyTorch project, following the conventions in `torch/_tensor_docs.py` and `torch/nn/functional.py`.
-
-## General Principles
-
-- Use **raw strings** (`r"""..."""`) for all docstrings to avoid issues with LaTeX/math backslashes
-- Follow **Sphinx/reStructuredText** (reST) format for documentation
-- Be **concise but complete** - include all essential information
-- Always include **examples** when possible
-- Use **cross-references** to related functions/classes
-
-## Docstring Structure
-
-### 1. Function Signature (First Line)
-
-Start with the function signature showing all parameters:
-
-```python
-r"""function_name(param1, param2, *, kwarg1=default1, kwarg2=default2) -> ReturnType
-```
-
-**Notes:**
-- Include the function name
-- Show positional and keyword-only arguments (use `*` separator)
-- Include default values
-- Show return type annotation
-- This line should NOT end with a period
-
-### 2. Brief Description
-
-Provide a one-line description of what the function does:
-
-```python
-r"""conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1) -> Tensor
-
-Applies a 2D convolution over an input image composed of several input
-planes.
-```
-
-### 3. Mathematical Formulas (if applicable)
-
-Use Sphinx math directives for mathematical expressions:
-
-```python
-.. math::
-    \text{Softmax}(x_{i}) = \frac{\exp(x_i)}{\sum_j \exp(x_j)}
-```
-
-Or inline math: `:math:\`x^2\``
-
-### 4. Cross-References
-
-Link to related classes and functions using Sphinx roles:
-
-- `:class:\`~torch.nn.ModuleName\`` - Link to a class
-- `:func:\`torch.function_name\`` - Link to a function
-- `:meth:\`~Tensor.method_name\`` - Link to a method
-- `:attr:\`attribute_name\`` - Reference an attribute
-- The `~` prefix shows only the last component (e.g., `Conv2d` instead of `torch.nn.Conv2d`)
-
-**Example:**
-```python
-See :class:`~torch.nn.Conv2d` for details and output shape.
-```
-
-### 5. Notes and Warnings
-
-Use admonitions for important information:
-
-```python
-.. note::
-    This function doesn't work directly with NLLLoss,
-    which expects the Log to be computed between the Softmax and itself.
-    Use log_softmax instead (it's faster and has better numerical properties).
-
-.. warning::
-    :func:`new_tensor` always copies :attr:`data`. If you have a Tensor
-    ``data`` and want to avoid a copy, use :func:`torch.Tensor.requires_grad_`
-    or :func:`torch.Tensor.detach`.
-```
-
-### 6. Args Section
-
-Document all parameters with type annotations and descriptions:
-
-```python
-Args:
-    input (Tensor): input tensor of shape :math:`(\text{minibatch} , \text{in\_channels} , iH , iW)`
-    weight (Tensor): filters of shape :math:`(\text{out\_channels} , kH , kW)`
-    bias (Tensor, optional): optional bias tensor of shape :math:`(\text{out\_channels})`. Default: ``None``
-    stride (int or tuple): the stride of the convolving kernel. Can be a single number or a
-      tuple `(sH, sW)`. Default: 1
-```
-
-**Formatting rules:**
-- Parameter name in **lowercase**
-- Type in parentheses: `(Type)`, `(Type, optional)` for optional parameters
-- Description follows the type
-- For optional parameters, include "Default: ``value``" at the end
-- Use double backticks for inline code: ``` ``None`` ```
-- Indent continuation lines by 2 spaces
-
-### 7. Keyword Args Section (if applicable)
-
-Sometimes keyword arguments are documented separately:
-
-```python
-Keyword args:
-    dtype (:class:`torch.dtype`, optional): the desired type of returned tensor.
-        Default: if None, same :class:`torch.dtype` as this tensor.
-    device (:class:`torch.device`, optional): the desired device of returned tensor.
-        Default: if None, same :class:`torch.device` as this tensor.
-    requires_grad (bool, optional): If autograd should record operations on the
-        returned tensor. Default: ``False``.
-```
-
-### 8. Returns Section (if needed)
-
-Document the return value:
-
-```python
-Returns:
-    Tensor: Sampled tensor of same shape as `logits` from the Gumbel-Softmax distribution.
-        If ``hard=True``, the returned samples will be one-hot, otherwise they will
-        be probability distributions that sum to 1 across `dim`.
-```
-
-Or simply include it in the function signature line if obvious from context.
-
-### 9. Examples Section
-
-Always include examples when possible:
-
-```python
-Examples::
-
-    >>> inputs = torch.randn(33, 16, 30)
-    >>> filters = torch.randn(20, 16, 5)
-    >>> F.conv1d(inputs, filters)
-
-    >>> # With square kernels and equal stride
-    >>> filters = torch.randn(8, 4, 3, 3)
-    >>> inputs = torch.randn(1, 4, 5, 5)
-    >>> F.conv2d(inputs, filters, padding=1)
-```
-
-**Formatting rules:**
-- Use `Examples::` with double colon
-- Use `>>>` prompt for Python code
-- Include comments with `#` when helpful
-- Show actual output when it helps understanding (indent without `>>>`)
-
-### 10. External References
-
-Link to papers or external documentation:
-
-```python
-.. _Link Name:
-    https://arxiv.org/abs/1611.00712
-```
-
-Reference them in text: ```See `Link Name`_```
-
-## Method Types
-
-### Native Python Functions
-
-For regular Python functions, use a standard docstring:
-
-```python
-def relu(input: Tensor, inplace: bool = False) -> Tensor:
-    r"""relu(input, inplace=False) -> Tensor
-
-    Applies the rectified linear unit function element-wise. See
-    :class:`~torch.nn.ReLU` for more details.
-    """
-    # implementation
-```
-
-### C-Bound Functions (using add_docstr)
-
-For C-bound functions, use `_add_docstr`:
-
-```python
-conv1d = _add_docstr(
-    torch.conv1d,
-    r"""
-conv1d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1) -> Tensor
-
-Applies a 1D convolution over an input signal composed of several input
-planes.
-
-See :class:`~torch.nn.Conv1d` for details and output shape.
-
-Args:
-    input: input tensor of shape :math:`(\text{minibatch} , \text{in\_channels} , iW)`
-    weight: filters of shape :math:`(\text{out\_channels} , kW)`
-    ...
-""",
-)
-```
-
-### In-Place Variants
-
-For in-place operations (ending with `_`), reference the original:
-
-```python
-add_docstr_all(
-    "abs_",
-    r"""
-abs_() -> Tensor
-
-In-place version of :meth:`~Tensor.abs`
-""",
-)
-```
-
-### Alias Functions
-
-For aliases, simply reference the original:
-
-```python
-add_docstr_all(
-    "absolute",
-    r"""
-absolute() -> Tensor
-
-Alias for :func:`abs`
-""",
-)
-```
-
-## Common Patterns
-
-### Shape Documentation
-
-Use LaTeX math notation for tensor shapes:
-
-```python
-:math:`(\text{minibatch} , \text{in\_channels} , iH , iW)`
-```
-
-### Reusable Argument Definitions
-
-For commonly used arguments, define them once and reuse:
-
-```python
-common_args = parse_kwargs(
-    """
-    dtype (:class:`torch.dtype`, optional): the desired type of returned tensor.
-        Default: if None, same as this tensor.
-"""
-)
-
-# Then use with .format():
-r"""
-...
-
-Keyword args:
-    {dtype}
-    {device}
-""".format(**common_args)
-```
-
-### Template Insertion
-
-Insert reproducibility notes or other common text:
-
-```python
-r"""
-{tf32_note}
-
-{cudnn_reproducibility_note}
-""".format(**reproducibility_notes, **tf32_notes)
-```
-
-## Complete Example
-
-Here's a complete example showing all elements:
-
-```python
-def gumbel_softmax(
-    logits: Tensor,
-    tau: float = 1,
-    hard: bool = False,
-    eps: float = 1e-10,
-    dim: int = -1,
-) -> Tensor:
-    r"""
-    Sample from the Gumbel-Softmax distribution and optionally discretize.
-
-    Args:
-        logits (Tensor): `[..., num_features]` unnormalized log probabilities
-        tau (float): non-negative scalar temperature
-        hard (bool): if ``True``, the returned samples will be discretized as one-hot vectors,
-              but will be differentiated as if it is the soft sample in autograd. Default: ``False``
-        dim (int): A dimension along which softmax will be computed. Default: -1
-
-    Returns:
-        Tensor: Sampled tensor of same shape as `logits` from the Gumbel-Softmax distribution.
-            If ``hard=True``, the returned samples will be one-hot, otherwise they will
-            be probability distributions that sum to 1 across `dim`.
-
-    .. note::
-        This function is here for legacy reasons, may be removed from nn.Functional in the future.
-
-    Examples::
-        >>> logits = torch.randn(20, 32)
-        >>> # Sample soft categorical using reparametrization trick:
-        >>> F.gumbel_softmax(logits, tau=1, hard=False)
-        >>> # Sample hard categorical using "Straight-through" trick:
-        >>> F.gumbel_softmax(logits, tau=1, hard=True)
-
-    .. _Link 1:
-        https://arxiv.org/abs/1611.00712
-    """
-    # implementation
-```
-
-## Quick Checklist
-
-When writing a PyTorch docstring, ensure:
-
-- [ ] Use raw string (`r"""`)
-- [ ] Include function signature on first line
-- [ ] Provide brief description
-- [ ] Document all parameters in Args section with types
-- [ ] Include default values for optional parameters
-- [ ] Use Sphinx cross-references (`:func:`, `:class:`, `:meth:`)
-- [ ] Add mathematical formulas if applicable
-- [ ] Include at least one example in Examples section
-- [ ] Add warnings/notes for important caveats
-- [ ] Link to related module class with `:class:`
-- [ ] Use proper math notation for tensor shapes
-- [ ] Follow consistent formatting and indentation
-
-## Common Sphinx Roles Reference
-
-- `:class:\`~torch.nn.Module\`` - Class reference
-- `:func:\`torch.function\`` - Function reference
-- `:meth:\`~Tensor.method\`` - Method reference
-- `:attr:\`attribute\`` - Attribute reference
-- `:math:\`equation\`` - Inline math
-- `:ref:\`label\`` - Internal reference
-- ``` ``code`` ``` - Inline code (use double backticks)
-
-## Additional Notes
-
-- **Indentation**: Use 4 spaces for code, 2 spaces for continuation of parameter descriptions
-- **Line length**: Try to keep lines under 100 characters when possible
-- **Periods**: End sentences with periods, but not the signature line
-- **Backticks**: Use double backticks for code: ``` ``True`` ``None`` ``False`` ```
-- **Types**: Common types are `Tensor`, `int`, `float`, `bool`, `str`, `tuple`, `list`, etc.

.claude/skills/skill-writer/SKILL.md

@@ -1,385 +0,0 @@
----
-name: skill-writer
-description: Guide users through creating Agent Skills for Claude Code. Use when the user wants to create, write, author, or design a new Skill, or needs help with SKILL.md files, frontmatter, or skill structure.
----
-
-# Skill Writer
-
-This Skill helps you create well-structured Agent Skills for Claude Code that follow best practices and validation requirements.
-
-## When to use this Skill
-
-Use this Skill when:
-- Creating a new Agent Skill
-- Writing or updating SKILL.md files
-- Designing skill structure and frontmatter
-- Troubleshooting skill discovery issues
-- Converting existing prompts or workflows into Skills
-
-## Instructions
-
-### Step 1: Determine Skill scope
-
-First, understand what the Skill should do:
-
-1. **Ask clarifying questions**:
-   - What specific capability should this Skill provide?
-   - When should Claude use this Skill?
-   - What tools or resources does it need?
-   - Is this for personal use or team sharing?
-
-2. **Keep it focused**: One Skill = one capability
-   - Good: "PDF form filling", "Excel data analysis"
-   - Too broad: "Document processing", "Data tools"
-
-### Step 2: Choose Skill location
-
-Determine where to create the Skill:
-
-**Personal Skills** (`~/.claude/skills/`):
-- Individual workflows and preferences
-- Experimental Skills
-- Personal productivity tools
-
-**Project Skills** (`.claude/skills/`):
-- Team workflows and conventions
-- Project-specific expertise
-- Shared utilities (committed to git)
-
-### Step 3: Create Skill structure
-
-Create the directory and files:
-
-```bash
-# Personal
-mkdir -p ~/.claude/skills/skill-name
-
-# Project
-mkdir -p .claude/skills/skill-name
-```
-
-For multi-file Skills:
-```
-skill-name/
-├── SKILL.md (required)
-├── reference.md (optional)
-├── examples.md (optional)
-├── scripts/
-│   └── helper.py (optional)
-└── templates/
-    └── template.txt (optional)
-```
-
-### Step 4: Write SKILL.md frontmatter
-
-Create YAML frontmatter with required fields:
-
-```yaml
----
-name: skill-name
-description: Brief description of what this does and when to use it
----
-```
-
-**Field requirements**:
-
-- **name**:
-  - Lowercase letters, numbers, hyphens only
-  - Max 64 characters
-  - Must match directory name
-  - Good: `pdf-processor`, `git-commit-helper`
-  - Bad: `PDF_Processor`, `Git Commits!`
-
-- **description**:
-  - Max 1024 characters
-  - Include BOTH what it does AND when to use it
-  - Use specific trigger words users would say
-  - Mention file types, operations, and context
-
-**Optional frontmatter fields**:
-
-- **allowed-tools**: Restrict tool access (comma-separated list)
-  ```yaml
-  allowed-tools: Read, Grep, Glob
-  ```
-  Use for:
-  - Read-only Skills
-  - Security-sensitive workflows
-  - Limited-scope operations
-
-### Step 5: Write effective descriptions
-
-The description is critical for Claude to discover your Skill.
-
-**Formula**: `[What it does] + [When to use it] + [Key triggers]`
-
-**Examples**:
-
-✅ **Good**:
-```yaml
-description: Extract text and tables from PDF files, fill forms, merge documents. Use when working with PDF files or when the user mentions PDFs, forms, or document extraction.
-```
-
-✅ **Good**:
-```yaml
-description: Analyze Excel spreadsheets, create pivot tables, and generate charts. Use when working with Excel files, spreadsheets, or analyzing tabular data in .xlsx format.
-```
-
-❌ **Too vague**:
-```yaml
-description: Helps with documents
-description: For data analysis
-```
-
-**Tips**:
-- Include specific file extensions (.pdf, .xlsx, .json)
-- Mention common user phrases ("analyze", "extract", "generate")
-- List concrete operations (not generic verbs)
-- Add context clues ("Use when...", "For...")
-
-### Step 6: Structure the Skill content
-
-Use clear Markdown sections:
-
-```markdown
-# Skill Name
-
-Brief overview of what this Skill does.
-
-## Quick start
-
-Provide a simple example to get started immediately.
-
-## Instructions
-
-Step-by-step guidance for Claude:
-1. First step with clear action
-2. Second step with expected outcome
-3. Handle edge cases
-
-## Examples
-
-Show concrete usage examples with code or commands.
-
-## Best practices
-
-- Key conventions to follow
-- Common pitfalls to avoid
-- When to use vs. not use
-
-## Requirements
-
-List any dependencies or prerequisites:
-```bash
-pip install package-name
-```
-
-## Advanced usage
-
-For complex scenarios, see [reference.md](reference.md).
-```
-
-### Step 7: Add supporting files (optional)
-
-Create additional files for progressive disclosure:
-
-**reference.md**: Detailed API docs, advanced options
-**examples.md**: Extended examples and use cases
-**scripts/**: Helper scripts and utilities
-**templates/**: File templates or boilerplate
-
-Reference them from SKILL.md:
-```markdown
-For advanced usage, see [reference.md](reference.md).
-
-Run the helper script:
-\`\`\`bash
-python scripts/helper.py input.txt
-\`\`\`
-```
-
-### Step 8: Validate the Skill
-
-Check these requirements:
-
-✅ **File structure**:
-- [ ] SKILL.md exists in correct location
-- [ ] Directory name matches frontmatter `name`
-
-✅ **YAML frontmatter**:
-- [ ] Opening `---` on line 1
-- [ ] Closing `---` before content
-- [ ] Valid YAML (no tabs, correct indentation)
-- [ ] `name` follows naming rules
-- [ ] `description` is specific and < 1024 chars
-
-✅ **Content quality**:
-- [ ] Clear instructions for Claude
-- [ ] Concrete examples provided
-- [ ] Edge cases handled
-- [ ] Dependencies listed (if any)
-
-✅ **Testing**:
-- [ ] Description matches user questions
-- [ ] Skill activates on relevant queries
-- [ ] Instructions are clear and actionable
-
-### Step 9: Test the Skill
-
-1. **Restart Claude Code** (if running) to load the Skill
-
-2. **Ask relevant questions** that match the description:
-   ```
-   Can you help me extract text from this PDF?
-   ```
-
-3. **Verify activation**: Claude should use the Skill automatically
-
-4. **Check behavior**: Confirm Claude follows the instructions correctly
-
-### Step 10: Debug if needed
-
-If Claude doesn't use the Skill:
-
-1. **Make description more specific**:
-   - Add trigger words
-   - Include file types
-   - Mention common user phrases
-
-2. **Check file location**:
-   ```bash
-   ls ~/.claude/skills/skill-name/SKILL.md
-   ls .claude/skills/skill-name/SKILL.md
-   ```
-
-3. **Validate YAML**:
-   ```bash
-   cat SKILL.md | head -n 10
-   ```
-
-4. **Run debug mode**:
-   ```bash
-   claude --debug
-   ```
-
-## Common patterns
-
-### Read-only Skill
-
-```yaml
----
-name: code-reader
-description: Read and analyze code without making changes. Use for code review, understanding codebases, or documentation.
-allowed-tools: Read, Grep, Glob
----
-```
-
-### Script-based Skill
-
-```yaml
----
-name: data-processor
-description: Process CSV and JSON data files with Python scripts. Use when analyzing data files or transforming datasets.
----
-
-# Data Processor
-
-## Instructions
-
-1. Use the processing script:
-\`\`\`bash
-python scripts/process.py input.csv --output results.json
-\`\`\`
-
-2. Validate output with:
-\`\`\`bash
-python scripts/validate.py results.json
-\`\`\`
-```
-
-### Multi-file Skill with progressive disclosure
-
-```yaml
----
-name: api-designer
-description: Design REST APIs following best practices. Use when creating API endpoints, designing routes, or planning API architecture.
----
-
-# API Designer
-
-Quick start: See [examples.md](examples.md)
-
-Detailed reference: See [reference.md](reference.md)
-
-## Instructions
-
-1. Gather requirements
-2. Design endpoints (see examples.md)
-3. Document with OpenAPI spec
-4. Review against best practices (see reference.md)
-```
-
-## Best practices for Skill authors
-
-1. **One Skill, one purpose**: Don't create mega-Skills
-2. **Specific descriptions**: Include trigger words users will say
-3. **Clear instructions**: Write for Claude, not humans
-4. **Concrete examples**: Show real code, not pseudocode
-5. **List dependencies**: Mention required packages in description
-6. **Test with teammates**: Verify activation and clarity
-7. **Version your Skills**: Document changes in content
-8. **Use progressive disclosure**: Put advanced details in separate files
-
-## Validation checklist
-
-Before finalizing a Skill, verify:
-
-- [ ] Name is lowercase, hyphens only, max 64 chars
-- [ ] Description is specific and < 1024 chars
-- [ ] Description includes "what" and "when"
-- [ ] YAML frontmatter is valid
-- [ ] Instructions are step-by-step
-- [ ] Examples are concrete and realistic
-- [ ] Dependencies are documented
-- [ ] File paths use forward slashes
-- [ ] Skill activates on relevant queries
-- [ ] Claude follows instructions correctly
-
-## Troubleshooting
-
-**Skill doesn't activate**:
-- Make description more specific with trigger words
-- Include file types and operations in description
-- Add "Use when..." clause with user phrases
-
-**Multiple Skills conflict**:
-- Make descriptions more distinct
-- Use different trigger words
-- Narrow the scope of each Skill
-
-**Skill has errors**:
-- Check YAML syntax (no tabs, proper indentation)
-- Verify file paths (use forward slashes)
-- Ensure scripts have execute permissions
-- List all dependencies
-
-## Examples
-
-See the documentation for complete examples:
-- Simple single-file Skill (commit-helper)
-- Skill with tool permissions (code-reviewer)
-- Multi-file Skill (pdf-processing)
-
-## Output format
-
-When creating a Skill, I will:
-
-1. Ask clarifying questions about scope and requirements
-2. Suggest a Skill name and location
-3. Create the SKILL.md file with proper frontmatter
-4. Include clear instructions and examples
-5. Add supporting files if needed
-6. Provide testing instructions
-7. Validate against all requirements
-
-The result will be a complete, working Skill that follows all best practices and validation rules.

.flake8

@@ -7,12 +7,16 @@ max-line-length = 120
 # C408 ignored because we like the dict keyword argument syntax
 # E501 is not flexible enough, we're using B950 instead
 ignore =
-    E203,E305,E402,E501,E704,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
     # shebang has extra meaning in fbcode lints, so I think it's not worth trying
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
     B007,B008,B017,B019,B023,B028,B903,B905,B906,B907,B908,B910
+    # these ignores are from flake8-comprehensions; please fix!
+    C407,
+    # these ignores are from flake8-logging-format; please fix!
+    G100,G101,G200
     # these ignores are from flake8-simplify. please fix or ignore with commented reason
     SIM105,SIM108,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
     # SIM104 is already covered by pyupgrade ruff

.github/actions/diskspace-cleanup/action.yml

@@ -27,9 +27,7 @@ runs:
             docker system prune -af
             diskspace_new=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
             if [[ "$diskspace_new" -gt "$diskspace_cutoff" ]] ; then
-                diskspace_cutoff_int=$((diskspace_cutoff + 0))
-                difference=$((100 - diskspace_cutoff_int))
-                echo "Error: Available diskspace is less than $difference percent. Not enough diskspace."
+                echo "Error: Available diskspace is less than $diskspace_cutoff percent. Not enough diskspace."
                 echo "$msg"
                 exit 1
             else

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

@@ -124,10 +124,3 @@ runs:
       id: login-ecr
       continue-on-error: true
       uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
-
-    - name: Preserve github env variables for use in docker
-      shell: bash
-      run: |
-        env | grep '^GITHUB' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
-        env | grep '^CI' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
-        env | grep '^RUNNER' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-3b0e7a6f192ca2715e7e6cbe5db007aea7165fe2
+1b013f5b5a87a1882eb143c26d79d091150d6a37

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-218d2ab791d437309f91e0486eb9fa7f00badc17
+faffd5cf673615583da6517275e361cb3dbc77e6

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-df6798dfb931ce7c7fe5bed2447cd1092a5981af
+0fa6e3129e61143224663e1ec67980d12b7ec4eb

.github/ci_configs/vllm/Dockerfile

@@ -283,9 +283,6 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
         uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
     fi
 
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system --pre apache-tvm-ffi==0.1.0b15
-
 # Install the vllm wheel from previous stage
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/vllm/*.whl --verbose
@@ -298,8 +295,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 
-# TODO(elainewy): remove this once vllm commit is updated, and install flashinfer from pip
-# see https://github.com/pytorch/pytorch/pull/165274#issuecomment-3408531784
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
 ARG FLASHINFER_GIT_REF="v0.2.14.post1"
 

.github/label_to_label.yml

@@ -15,11 +15,6 @@
   - "module: reinplacing"
   then:
   - "module: pt2-dispatcher"
-- any:
-  - "vllm-compile"
-  then:
-  - "module: vllm"
-  - "oncall: pt2"
 - any:
   - "module: vmap"
   then:
@@ -32,6 +27,10 @@
   - "module: pt2 optimizer"
   then:
   - "module: dynamo"
+- any:
+  - "module: flex attention"
+  then:
+  - "module: higher order operators"
 - any:
   - "module: aotinductor"
   then:

.github/labeler.yml

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

.github/merge_rules.yaml

@@ -540,26 +540,6 @@
   - Lint
   - pull
 
-- name: PrivateUse1
-  patterns:
-  - torch/accelerator/**
-  - torch/utils/backend_registration.py
-  - torch/csrc/acc/**
-  - torch/csrc/DeviceAccelerator.*
-  - torch/csrc/profiler/standalone/privateuse1_observer.*
-  - aten/src/ATen/DeviceAccelerator.*
-  - aten/src/ATen/core/GeneratorForPrivateuseone.*
-  - aten/src/ATen/detail/PrivateUse1HooksInterface.*
-  - docs/source/accelerator/**
-  - test/cpp_extensions/open_registration_extension/torch_openreg/**
-  approved_by:
-  - albanD
-  - fffrog
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
 - name: superuser
   patterns:
   - '*'

.github/pytorch-probot.yml

@@ -19,7 +19,6 @@ ciflow_push_tags:
 - ciflow/inductor-perf-test-nightly-rocm-mi300
 - ciflow/inductor-perf-test-nightly-rocm-mi355
 - ciflow/inductor-perf-test-nightly-x86-zen
-- ciflow/inductor-perf-test-nightly-xpu
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
 - ciflow/linux-aarch64
@@ -27,7 +26,6 @@ ciflow_push_tags:
 - ciflow/nightly
 - ciflow/op-benchmark
 - ciflow/periodic
-- ciflow/periodic-rocm-mi200
 - ciflow/periodic-rocm-mi300
 - ciflow/pull
 - ciflow/quantization-periodic
@@ -35,7 +33,6 @@ ciflow_push_tags:
 - ciflow/rocm
 - ciflow/rocm-mi300
 - ciflow/rocm-mi355
-- ciflow/rocm-navi31
 - ciflow/s390
 - ciflow/slow
 - ciflow/torchbench

.github/scripts/generate_binary_build_matrix.py

@@ -11,24 +11,18 @@ architectures:
     * Latest XPU
 """
 
-import json
 import os
-import re
-from pathlib import Path
 from typing import Optional
 
 
-SCRIPT_DIR = Path(__file__).absolute().parent
-REPO_ROOT = SCRIPT_DIR.parent.parent
-
-
+# NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
 CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
     "12.9": "12.9.1",
-    "13.0": "13.0.2",
+    "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
@@ -37,7 +31,8 @@ CUDA_ARCHES_CUDNN_VERSION = {
     "13.0": "9",
 }
 
-ROCM_ARCHES = ["7.0", "7.1"]
+# NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
+ROCM_ARCHES = ["6.4", "7.0"]
 
 XPU_ARCHES = ["xpu"]
 
@@ -61,7 +56,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "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==3.4.5; 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'"
@@ -78,44 +73,44 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "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==3.4.5; 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'"
     ),
     "12.9": (
-        "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "13.0": (
-        "nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | "
+        "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==13.1.0.3; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.61; 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==12.0.4.66; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.3.3; 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==3.4.5; platform_system == 'Linux' | "
-        "nvidia-nvtx==13.0.85; platform_system == 'Linux' | "
-        "nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | "
-        "nvidia-cufile==1.15.1.6; 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'"
     ),
     "xpu": (
         "intel-cmplr-lib-rt==2025.2.1 | "
@@ -142,48 +137,9 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
 }
 
 
-# Used by tools/nightly.py
-PYTORCH_NIGHTLY_PIP_INDEX_URL = "https://download.pytorch.org/whl/nightly"
-NIGHTLY_SOURCE_MATRIX = {
-    "cpu": dict(
-        name="cpu",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cpu",
-        supported_platforms=["Linux", "macOS", "Windows"],
-        accelerator="cpu",
-    )
-}
-CUDA_NIGHTLY_SOURCE_MATRIX = {
-    f"cuda-{major}.{minor}": dict(
-        name=f"cuda-{major}.{minor}",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cu{major}{minor}",
-        supported_platforms=["Linux", "Windows"],
-        accelerator="cuda",
-    )
-    for major, minor in (map(int, version.split(".")) for version in CUDA_ARCHES)
-}
-ROCM_NIGHTLY_SOURCE_MATRIX = {
-    f"rocm-{major}.{minor}": dict(
-        name=f"rocm-{major}.{minor}",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/rocm{major}.{minor}",
-        supported_platforms=["Linux"],
-        accelerator="rocm",
-    )
-    for major, minor in (map(int, version.split(".")) for version in ROCM_ARCHES)
-}
-XPU_NIGHTLY_SOURCE_MATRIX = {
-    "xpu": dict(
-        name="xpu",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/xpu",
-        supported_platforms=["Linux"],
-        accelerator="xpu",
-    )
-}
-NIGHTLY_SOURCE_MATRIX.update(CUDA_NIGHTLY_SOURCE_MATRIX)
-NIGHTLY_SOURCE_MATRIX.update(ROCM_NIGHTLY_SOURCE_MATRIX)
-NIGHTLY_SOURCE_MATRIX.update(XPU_NIGHTLY_SOURCE_MATRIX)
-
-
 def get_nccl_wheel_version(arch_version: str) -> str:
+    import re
+
     requirements = map(
         str.strip, re.split("[;|]", PYTORCH_EXTRA_INSTALL_REQUIREMENTS[arch_version])
     )
@@ -191,14 +147,17 @@ def get_nccl_wheel_version(arch_version: str) -> str:
 
 
 def read_nccl_pin(arch_version: str) -> str:
-    nccl_pin_path = (
-        REPO_ROOT
-        / ".ci"
-        / "docker"
-        / "ci_commit_pins"
-        / f"nccl-cu{arch_version[:2]}.txt"
+    from pathlib import Path
+
+    nccl_pin_path = os.path.join(
+        Path(__file__).absolute().parents[2],
+        ".ci",
+        "docker",
+        "ci_commit_pins",
+        f"nccl-cu{arch_version[:2]}.txt",
     )
-    return nccl_pin_path.read_text().strip()
+    with open(nccl_pin_path) as f:
+        return f.read().strip()
 
 
 def validate_nccl_dep_consistency(arch_version: str) -> None:
@@ -206,8 +165,7 @@ def validate_nccl_dep_consistency(arch_version: str) -> None:
     wheel_ver = get_nccl_wheel_version(arch_version)
     if not nccl_release_tag.startswith(f"v{wheel_ver}"):
         raise RuntimeError(
-            f"{arch_version} NCCL release tag version {nccl_release_tag} "
-            f"does not correspond to wheel version {wheel_ver}"
+            f"{arch_version} NCCL release tag version {nccl_release_tag} does not correspond to wheel version {wheel_ver}"
         )
 
 
@@ -283,11 +241,7 @@ def generate_libtorch_matrix(
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
         elif os == "windows":
-            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
-            # in 2.10
-            windows_cuda_arches = CUDA_ARCHES.copy()
-            windows_cuda_arches.remove("12.9")
-            arches += windows_cuda_arches
+            arches += CUDA_ARCHES
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -351,11 +305,7 @@ def generate_wheels_matrix(
         if os == "linux":
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
-            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
-            # in 2.10
-            windows_cuda_arches = CUDA_ARCHES.copy()
-            windows_cuda_arches.remove("12.9")
-            arches += windows_cuda_arches + XPU_ARCHES
+            arches += CUDA_ARCHES + XPU_ARCHES
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -454,14 +404,7 @@ def generate_wheels_matrix(
     return ret
 
 
-arch_version = ""
-for arch_version in CUDA_ARCHES:
-    validate_nccl_dep_consistency(arch_version)
-del arch_version
-
-
-if __name__ == "__main__":
-    # Used by tools/nightly.py
-    (SCRIPT_DIR / "nightly_source_matrix.json").write_text(
-        json.dumps(NIGHTLY_SOURCE_MATRIX, indent=4) + "\n"
-    )
+validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
+validate_nccl_dep_consistency("12.8")
+validate_nccl_dep_consistency("12.6")

.github/scripts/trymerge.py

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

.github/templates/macos_binary_build_workflow.yml.j2

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

.github/templates/windows_binary_build_workflow.yml.j2

@@ -79,9 +79,9 @@ jobs:
     runs-on: "windows-11-arm64-preview"
     {%- else %}
     {%- if branches == "nightly" %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     {%- else %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge.nonephemeral"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
     {%- endif %}
     {%- endif %}
     timeout-minutes: !{{ common.timeout_minutes_windows_binary }}

.github/workflows/_linux-build.yml

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

.github/workflows/_linux-test.yml

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

.github/workflows/_win-build.yml

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

.github/workflows/_xpu-test.yml

@@ -38,10 +38,6 @@ on:
         default: ""
         description: |
           List of tests to include (empty string implies default list)
-      dashboard-tag:
-        required: false
-        type: string
-        default: ""
       disable-monitor:
         description: |
           [Experimental] Disable utilization monitoring for tests.
@@ -62,11 +58,6 @@ on:
         required: false
         type: number
         default: 1
-    secrets:
-      HUGGING_FACE_HUB_TOKEN:
-        required: false
-        description: |
-          HF Auth token to avoid rate limits when downloading models or datasets from hub
 permissions:
   id-token: write
   contents: read
@@ -205,8 +196,6 @@ jobs:
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
           PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
           TESTS_TO_INCLUDE: ${{ inputs.tests-to-include }}
-          DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
         timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
         run: |
           # Fetch aws credential from IMDs
@@ -257,8 +246,6 @@ jobs:
             -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
             -e TESTS_TO_INCLUDE \
             -e ZE_AFFINITY_MASK \
-            -e HUGGING_FACE_HUB_TOKEN \
-            -e DASHBOARD_TAG \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --ulimit core=0 \

.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", "rocm7.0", "rocm7.1", "cpu"]
+        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.4", "rocm7.0", "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: "rocm6.4"  },
           { tag: "rocm7.0"  },
-          { tag: "rocm7.1"  },
           { tag: "cpu"      },
         ]
     steps:

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

@@ -34,7 +34,7 @@ jobs:
       id-token: write
     strategy:
       matrix:
-        rocm_version: ["71", "70"]
+        rocm_version: ["70", "64"]
     steps:
       - name: Checkout PyTorch
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

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

@@ -54,8 +54,8 @@ jobs:
           { name: "manylinuxaarch64-builder",       tag: "cuda12.9",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.8",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm7.0",           runner: "linux.9xlarge.ephemeral" },
-          { name: "manylinux2_28-builder",          tag: "rocm7.1",           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: "manylinux2_28-builder",          tag: "xpu",               runner: "linux.9xlarge.ephemeral" },

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

@@ -55,7 +55,7 @@ jobs:
         docker-image: ["pytorch/manylinux2_28-builder:cpu"]
         include:
           - device: "rocm"
-            rocm_version: "7.1"
+            rocm_version: "7.0"
             runs_on: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge"
           - device: "cuda"
             rocm_version: ""

.github/workflows/docker-builds.yml

@@ -57,7 +57,6 @@ jobs:
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
           pytorch-linux-jammy-py3.10-clang12,
           pytorch-linux-jammy-py3.13-clang12,
-          pytorch-linux-jammy-py3.14-clang12,
           pytorch-linux-jammy-rocm-n-py3,
           pytorch-linux-noble-rocm-n-py3,
           pytorch-linux-jammy-rocm-n-py3-benchmarks,
@@ -67,7 +66,6 @@ jobs:
           pytorch-linux-jammy-py3.12-halide,
           pytorch-linux-jammy-xpu-n-1-py3,
           pytorch-linux-jammy-xpu-n-py3,
-          pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks,
           pytorch-linux-jammy-py3-clang18-asan,
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,

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

@@ -132,7 +132,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -178,7 +178,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -224,7 +224,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -270,7 +270,7 @@ jobs:
       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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; 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 }}
@@ -381,7 +381,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -427,7 +427,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -473,7 +473,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -519,7 +519,7 @@ jobs:
       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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; 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 }}
@@ -630,7 +630,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -676,7 +676,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -722,7 +722,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -768,7 +768,7 @@ jobs:
       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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; 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 }}
@@ -879,7 +879,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -925,7 +925,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -971,7 +971,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1017,7 +1017,7 @@ jobs:
       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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; 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 }}
@@ -1128,7 +1128,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -1174,7 +1174,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -1220,7 +1220,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1266,7 +1266,7 @@ jobs:
       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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; 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 }}
@@ -1377,7 +1377,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -1423,7 +1423,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -1469,7 +1469,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1515,7 +1515,7 @@ jobs:
       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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; 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 }}
@@ -1626,7 +1626,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -1672,7 +1672,7 @@ jobs:
       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==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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 }}
@@ -1718,7 +1718,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1764,7 +1764,7 @@ jobs:
       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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; 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

@@ -384,6 +384,124 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  libtorch-rocm6_4-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ 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:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-rocm6_4-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-rocm6_4-shared-with-deps-release-build
+      - get-label-type
+    runs-on: linux.rocm.gpu.mi250
+    timeout-minutes: 240
+    env:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      SKIP_ALL_TESTS: 1
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - name: Setup ROCm
+        uses: ./.github/actions/setup-rocm
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-rocm6_4-shared-with-deps-release
+          path: "${{ 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: rocm6.4
+          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_4-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-rocm6_4-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: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-rocm6_4-shared-with-deps-release
+    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
@@ -501,121 +619,3 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
-
-  libtorch-rocm7_1-shared-with-deps-release-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
-      build_name: libtorch-rocm7_1-shared-with-deps-release
-      build_environment: linux-binary-libtorch
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-rocm7_1-shared-with-deps-release-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - libtorch-rocm7_1-shared-with-deps-release-build
-      - get-label-type
-    runs-on: linux.rocm.gpu.mi250
-    timeout-minutes: 240
-    env:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      SKIP_ALL_TESTS: 1
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-    permissions:
-      id-token: write
-      contents: read
-    steps:
-      - name: Setup ROCm
-        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
-        name: Download Build Artifacts
-        with:
-          name: libtorch-rocm7_1-shared-with-deps-release
-          path: "${{ 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: rocm7.1
-          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_1-shared-with-deps-release-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: libtorch-rocm7_1-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: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-rocm7_1-shared-with-deps-release
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml

.github/workflows/generated-macos-arm64-binary-libtorch-release-nightly.yml

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

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

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

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

@@ -44,7 +44,7 @@ jobs:
   libtorch-cpu-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -291,7 +291,7 @@ jobs:
   libtorch-cuda12_6-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -541,7 +541,7 @@ jobs:
   libtorch-cuda12_8-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -788,10 +788,260 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-debug-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-debug-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-debug-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-debug-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-debug-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+      build_name: libtorch-cuda12_9-shared-with-deps-debug
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

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

@@ -44,7 +44,7 @@ jobs:
   libtorch-cpu-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -291,7 +291,7 @@ jobs:
   libtorch-cuda12_6-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -541,7 +541,7 @@ jobs:
   libtorch-cuda12_8-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -788,10 +788,260 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-release-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

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

@@ -88,27 +88,27 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
         ]}
     secrets: inherit
 

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

@@ -1,148 +0,0 @@
-name: inductor-perf-nightly-xpu
-
-on:
-  push:
-    tags:
-      - ciflow/inductor-perf-test-nightly-xpu/*
-  schedule:
-    - cron: 30 17 * * *
-  workflow_dispatch:
-    inputs:
-      training:
-        description: Run training (on by default)?
-        required: false
-        type: boolean
-        default: true
-      inference:
-        description: Run inference (on by default)?
-        required: false
-        type: boolean
-        default: true
-      default:
-        description: Run inductor_default?
-        required: false
-        type: boolean
-        default: false
-      dynamic:
-        description: Run inductor_dynamic_shapes?
-        required: false
-        type: boolean
-        default: false
-      cppwrapper:
-        description: Run inductor_cpp_wrapper?
-        required: false
-        type: boolean
-        default: false
-      cudagraphs:
-        description: Run inductor_cudagraphs?
-        required: false
-        type: boolean
-        default: false
-      freezing_cudagraphs:
-        description: Run inductor_cudagraphs with freezing for inference?
-        required: false
-        type: boolean
-        default: false
-      aotinductor:
-        description: Run aot_inductor for inference?
-        required: false
-        type: boolean
-        default: false
-      maxautotune:
-        description: Run inductor_max_autotune?
-        required: false
-        type: boolean
-        default: false
-      benchmark_configs:
-        description: The list of configs used the benchmark
-        required: false
-        type: string
-        default: inductor_huggingface_perf,inductor_timm_perf,inductor_torchbench_perf,cachebench
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-      opt_out_experiments: lf
-
-  xpu-n-py3_10-inductor-benchmark-build:
-    name: xpu-n-py3.10-inductor-benchmark
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-xpu-n-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks
-      runner: linux.c7i.12xlarge
-      test-matrix: |
-        { include: [
-          { config: "inductor_huggingface_perf_xpu", shard: 1, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 2, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 3, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 4, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 5, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
-        ]}
-    secrets: inherit
-
-  xpu-n-py3_10-inductor-benchmark-test-nightly:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name != 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-benchmark
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-jammy-xpu-n-py3.10
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-false-cppwrapper-true-aotinductor-true-freezing_cudagraphs-false-cudagraphs_low_precision-false
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      # Disable monitor in perf tests for more investigation
-      disable-monitor: true
-      monitor-log-interval: 10
-      monitor-data-collect-interval: 2
-    secrets: inherit
-
-  xpu-n-py3_10-inductor-benchmark-test:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name == 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-test
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-jammy-xpu-n-py3.10
-      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit

.github/workflows/inductor-periodic.yml

@@ -88,6 +88,7 @@ jobs:
     with:
       build-environment: linux-jammy-rocm-py3_10
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
+      sync-tag: rocm-build
       test-matrix: |
         { include: [
           { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },

.github/workflows/lint.yml

@@ -118,9 +118,9 @@ jobs:
         CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
         echo "Running all other linters"
         if [ "$CHANGED_FILES" = '*' ]; then
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY --all-files" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT --all-files" .github/scripts/lintrunner.sh
         else
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT ${CHANGED_FILES}" .github/scripts/lintrunner.sh
         fi
 
   quick-checks:

.github/workflows/operator_benchmark.yml

@@ -30,9 +30,9 @@ permissions:
   contents: read
 
 jobs:
-  x86-opbenchmark-build:
+  opbenchmark-build:
     if: github.repository_owner == 'pytorch'
-    name: x86-opbenchmark-build
+    name: opbenchmark-build
     uses: ./.github/workflows/_linux-build.yml
     with:
       build-environment: linux-jammy-py3.10-gcc11-build
@@ -43,36 +43,12 @@ jobs:
         ]}
     secrets: inherit
 
-  x86-opbenchmark-test:
-    name: x86-opbenchmark-test
+  opbenchmark-test:
+    name: opbenchmark-test
     uses: ./.github/workflows/_linux-test.yml
-    needs: x86-opbenchmark-build
+    needs: opbenchmark-build
     with:
       build-environment: linux-jammy-py3.10-gcc11-build
-      docker-image: ${{ needs.x86-opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.x86-opbenchmark-build.outputs.test-matrix }}
-    secrets: inherit
-
-  aarch64-opbenchmark-build:
-    if: github.repository_owner == 'pytorch'
-    name: aarch64-opbenchmark-build
-    uses: ./.github/workflows/_linux-build.yml
-    with:
-      build-environment: linux-jammy-aarch64-py3.10
-      runner: linux.arm64.m7g.4xlarge
-      docker-image-name: ci-image:pytorch-linux-jammy-aarch64-py3.10-gcc11
-      test-matrix: |
-        { include: [
-          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.arm64.m8g.4xlarge" },
-        ]}
-    secrets: inherit
-
-  aarch64-opbenchmark-test:
-    name: aarch64-opbenchmark-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs: aarch64-opbenchmark-build
-    with:
-      build-environment: linux-jammy-aarch64-py3.10
-      docker-image: ${{ needs.aarch64-opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.aarch64-opbenchmark-build.outputs.test-matrix }}
+      docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/periodic-rocm-mi200.yml

@@ -1,84 +0,0 @@
-name: periodic-rocm-mi200
-
-on:
-  schedule:
-    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
-    # Also run less frequently on weekends.
-    - cron: 45 0,8,16 * * 1-5
-    - cron: 45 4 * * 0,6
-    - cron: 45 4,12,20 * * 1-5
-    - cron: 45 12 * * 0,6
-    - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
-  push:
-    tags:
-      - ciflow/periodic/*
-      - ciflow/periodic-rocm-mi200/*
-    branches:
-      - release/*
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}-${{ github.event.schedule }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  llm-td:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/llm_td_retrieval.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  target-determination:
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    needs: llm-td
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch'
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-rocm-py3_10-build:
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      test-matrix: |
-        { include: [
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/periodic.yml

@@ -147,16 +147,15 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-debug
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9
-      cuda-arch-list: 8.9
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 2, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 3, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 4, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 5, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 6, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 7, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 1, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 2, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 3, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 4, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 5, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 6, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 7, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
         ]}
     secrets: inherit
 
@@ -204,6 +203,37 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.test-matrix }}
     secrets: inherit
 
+  linux-jammy-rocm-py3_10-build:
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
+      test-matrix: |
+        { include: [
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+    secrets: inherit
+
   linux-jammy-cuda12_8-py3-gcc11-slow-gradcheck-build:
     name: linux-jammy-cuda12.8-py3-gcc11-slow-gradcheck
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/pull.yml

@@ -347,8 +347,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # This should sync with the build in xpu.yml but xpu uses a larger runner
-      # sync-tag: linux-xpu-n-build
+      sync-tag: linux-xpu-n-build
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
       build-environment: linux-jammy-xpu-n-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3

.github/workflows/rocm-mi300.yml

@@ -45,6 +45,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-noble-rocm-py3.12-mi300
       docker-image-name: ci-image:pytorch-linux-noble-rocm-n-py3
+      sync-tag: rocm-build
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1" },

.github/workflows/rocm-mi355.yml

@@ -42,14 +42,15 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-noble-rocm-py3.12-mi355
       docker-image-name: ci-image:pytorch-linux-noble-rocm-n-py3
+      sync-tag: rocm-build
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
         ]}
     secrets: inherit
 

.github/workflows/rocm-navi31.yml

@@ -1,75 +0,0 @@
-name: rocm-navi31
-
-on:
-  push:
-    tags:
-      - ciflow/rocm-navi31/*
-  workflow_dispatch:
-  schedule:
-    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
-    # Also run less frequently on weekends.
-    - cron: 45 */2 * * 1-5
-    - cron: 45 4,12 * * 0,6
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  target-determination:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-rocm-py3_10-build:
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3_10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-      tests-to-include: >-
-         ${{ github.event_name == 'schedule' && 'test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
-         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
-         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
-         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
-         inductor/test_flex_attention inductor/test_max_autotune' || '' }}
-    secrets: inherit

.github/workflows/rocm.yml

@@ -26,23 +26,11 @@ jobs:
       id-token: write
       contents: read
 
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
   linux-jammy-rocm-py3_10-build:
     if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
     name: linux-jammy-rocm-py3.10
     uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
     with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-rocm-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       sync-tag: rocm-build
@@ -71,3 +59,29 @@ jobs:
       docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
     secrets: inherit
+
+  linux-jammy-rocm-py3_10-gfx1100-test:
+    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' }}
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3_10-gfx1100
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: |
+        { include: [
+          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+        ]}
+      tests-to-include: >
+         test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
+         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
+         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
+         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
+         inductor/test_flex_attention inductor/test_max_autotune
+    secrets: inherit

.github/workflows/trunk-tagging.yml

@@ -58,10 +58,8 @@ jobs:
           else
             COMMIT_SHA="${{ github.sha }}"
           fi
-          {
-            echo "sha=${COMMIT_SHA}"
-            echo "tag_name=trunk/${COMMIT_SHA}"
-          } >> "${GITHUB_OUTPUT}"
+          echo "sha=${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
+          echo "tag_name=trunk/${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
 
       - name: Validate commit SHA
         run: |
@@ -89,7 +87,7 @@ jobs:
             echo "✅ Commit ${COMMIT_SHA} is valid (automatic push trigger)"
           fi
 
-      - name: Create and push tag(s) with retry
+      - name: Create and push tag with retry
         id: check_tag
         env:
           TAG_NAME: ${{ steps.commit.outputs.tag_name }}
@@ -114,23 +112,14 @@ jobs:
             return 1
           }
 
-          # Counters for summary reporting
-          created_count=0
-          skipped_count=0
-          failed_count=0
+          # Exit early if tag already exists
+          if check_tag_exists; then
+            echo "✅ Tag already exists - no action needed"
+            echo "exists=true" >> "${GITHUB_OUTPUT}"
+            exit 0
+          fi
 
-          # Always write outputs once on exit
-          finish() {
-            set +e
-            if [ -n "${GITHUB_OUTPUT:-}" ]; then
-              {
-                echo "created_count=${created_count}"
-                echo "skipped_count=${skipped_count}"
-                echo "failed_count=${failed_count}"
-              } >> "${GITHUB_OUTPUT}"
-            fi
-          }
-          trap finish EXIT
+          echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
 
           # Retry configuration
           MAX_RETRIES=5
@@ -205,111 +194,31 @@ jobs:
             }
           }
 
-          # New behavior for push events: enumerate commits in the push and tag each one.
-          # For workflow_dispatch, retain existing single-SHA behavior.
-
-          # Always fetch tags once up front to improve idempotency in loops
-          git fetch origin --tags --quiet || true
-
-          if [ "${{ github.event_name }}" = "push" ]; then
-            BEFORE_SHA="${{ github.event.before }}"
-            AFTER_SHA="${{ github.sha }}"  # same as event.after
-
-            # List commits introduced by this push (old..new), oldest first for stable ordering
-            commits_file="$(mktemp)"
-            git rev-list --reverse "${BEFORE_SHA}..${AFTER_SHA}" > "${commits_file}"
-
-            if [ ! -s "${commits_file}" ]; then
-              echo "No new commits found between ${BEFORE_SHA}..${AFTER_SHA}; nothing to tag."
-              rm -f "${commits_file}"
-              exit 0
-            fi
-
-            commit_count="$(wc -l < "${commits_file}" | tr -d ' ')"
-            echo "Found ${commit_count} commit(s) to tag for push:"
-            while IFS= read -r sha; do
-              printf '  %s\n' "${sha}"
-            done < "${commits_file}"
-
-            while IFS= read -r sha; do
-              TAG_NAME="trunk/${sha}"
-              COMMIT_SHA="${sha}"
-
-              # If tag already exists locally or remotely, skip (idempotent)
-              if check_tag_exists; then
-                echo "✅ Tag ${TAG_NAME} already exists - skipping"
-                skipped_count=$((skipped_count + 1))
-                continue
-              fi
-
-              echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
-
-              if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
-                created_count=$((created_count + 1))
-              else
-                echo "Tag creation failed after all retry attempts for ${TAG_NAME}"
-                failed_count=$((failed_count + 1))
-              fi
-            done < "${commits_file}"
-
-            rm -f "${commits_file}"
-
-            if [ "${failed_count}" -gt 0 ]; then
-              exit 1
-            fi
+          # Execute with retry
+          if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
+            echo "exists=false" >> "${GITHUB_OUTPUT}"
             exit 0
           else
-            # workflow_dispatch path (single SHA tagging preserved)
-
-            # Exit early if tag already exists
-            if check_tag_exists; then
-              echo "✅ Tag already exists - no action needed"
-              skipped_count=1
-              exit 0
-            fi
-
-            echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
-
-            if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
-              created_count=1
-              exit 0
-            else
-              echo "Tag creation failed after all retry attempts"
-              failed_count=1
-              exit 1
-            fi
+            echo "Tag creation failed after all retry attempts"
+            exit 1
           fi
 
       - name: Tag creation summary
         if: always()
         run: |
-          if [ "${{ github.event_name }}" = "push" ]; then
-            echo "Trigger: push on main"
-            echo "Created: ${{ steps.check_tag.outputs.created_count }}"
-            echo "Skipped (already existed): ${{ steps.check_tag.outputs.skipped_count }}"
-            echo "Failed: ${{ steps.check_tag.outputs.failed_count }}"
-            if [ "${{ steps.check_tag.outputs.failed_count }}" = "0" ]; then
-              echo "✅ Completed tagging for push range ${{ github.event.before }}..${{ github.sha }}"
-            else
-              echo "❌ Some tags failed to create for push range ${{ github.event.before }}..${{ github.sha }}"
-            fi
+          if [ "${{ steps.check_tag.outputs.exists }}" = "true" ]; then
+            echo "✅ Tag ${{ steps.commit.outputs.tag_name }} already existed - no action needed"
+          elif [ "${{ job.status }}" = "success" ]; then
+            echo "✅ Successfully created tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
           else
-            if [ "${{ steps.check_tag.outputs.failed_count }}" = "0" ]; then
-              if [ "${{ steps.check_tag.outputs.created_count }}" = "0" ]; then
-                echo "✅ Tag ${{ steps.commit.outputs.tag_name }} already existed - no action needed"
-              else
-                echo "✅ Successfully created tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
-              fi
-            else
-              echo "❌ Failed to create tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
-            fi
-
-            echo ""
-            echo "Tag details:"
-            echo "  Name: ${{ steps.commit.outputs.tag_name }}"
-            echo "  Commit: ${{ steps.commit.outputs.sha }}"
-            echo "  Trigger: ${{ github.event_name }}"
-            if [ -n "${{ github.event.inputs.commit_sha }}" ]; then
-              echo "  Manual commit: ${{ github.event.inputs.commit_sha }}"
-            fi
+            echo "❌ Failed to create tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
+          fi
+
+          echo ""
+          echo "Tag details:"
+          echo "  Name: ${{ steps.commit.outputs.tag_name }}"
+          echo "  Commit: ${{ steps.commit.outputs.sha }}"
+          echo "  Trigger: ${{ github.event_name }}"
+          if [ -n "${{ github.event.inputs.commit_sha }}" ]; then
+            echo "  Manual commit: ${{ github.event.inputs.commit_sha }}"
           fi

.github/workflows/trunk.yml

@@ -190,40 +190,6 @@ jobs:
       runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
     secrets: inherit
 
-  linux-jammy-rocm-py3_10-build:
-    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
-    secrets: inherit
-
   inductor-build:
     name: inductor-build
     uses: ./.github/workflows/_linux-build.yml
@@ -234,23 +200,6 @@ jobs:
       cuda-arch-list: '8.0'
     secrets: inherit
 
-  # Test cross-compiled models with Windows libs extracted from wheel
-  cross-compile-linux-test:
-    name: cross-compile-linux-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-build
-      - get-label-type
-      - win-vs2022-cuda12_8-py3-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
-      test-matrix: |
-        { include: [
-          { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
-        ]}
-    secrets: inherit
-
   verify-cachebench-cpu-build:
     name: verify-cachebench-cpu-build
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/upload-test-stats.yml

@@ -6,7 +6,6 @@ on:
       - pull
       - trunk
       - periodic
-      - periodic-rocm-mi200
       - periodic-rocm-mi300
       - inductor
       - unstable

.github/workflows/xpu.yml

@@ -59,18 +59,14 @@ jobs:
       runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 2, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 3, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 4, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 5, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 6, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 7, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 8, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 9, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 10, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 11, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 12, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 1, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 2, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 3, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 4, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 5, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 6, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 7, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 8, num_shards: 8, runner: "linux.idc.xpu" },
         ]}
     secrets: inherit
 

.gitignore

@@ -143,7 +143,6 @@ scripts/release_notes/*.json
 sccache-stats*.json
 lint.json
 merge_record.json
-.github/scripts/nightly_source_matrix.json
 
 # These files get copied over on invoking setup.py
 torchgen/packaged/*
@@ -375,7 +374,6 @@ third_party/ruy/
 third_party/glog/
 
 # Virtualenv
-.venv/
 venv/
 
 # Log files

.lintrunner.toml

@@ -209,46 +209,6 @@ command = [
     '@{{PATHSFILE}}'
 ]
 
-
-[[linter]]
-code = 'PYREFLY'
-include_patterns = [
-    'torch/**/*.py',
-    'torch/**/*.pyi',
-    'torchgen/**/*.py',
-    'torchgen/**/*.pyi',
-    'functorch/**/*.py',
-    'functorch/**/*.pyi',
-]
-exclude_patterns = []
-command = [
-    'python3',
-    'tools/linter/adapters/pyrefly_linter.py',
-    '--config=pyrefly.toml',
-]
-init_command = [
-    'python3',
-    'tools/linter/adapters/pip_init.py',
-    '--dry-run={{DRYRUN}}',
-    'numpy==2.1.0 ; python_version >= "3.12"',
-    'expecttest==0.3.0',
-    'pyrefly==0.36.2',
-    'sympy==1.13.3',
-    'types-requests==2.27.25',
-    'types-pyyaml==6.0.2',
-    'types-tabulate==0.8.8',
-    'types-protobuf==5.29.1.20250403',
-    'types-setuptools==79.0.0.20250422',
-    'types-jinja2==2.11.9',
-    'types-colorama==0.4.6',
-    'filelock==3.18.0',
-    'junitparser==2.1.1',
-    'rich==14.1.0',
-    'optree==0.17.0',
-    'types-openpyxl==3.1.5.20250919',
-    'types-python-dateutil==2.9.0.20251008'
-]
-
 [[linter]]
 code = 'CLANGTIDY'
 include_patterns = [
@@ -833,7 +793,8 @@ exclude_patterns = [
 command = [
     'python3',
     'tools/linter/adapters/grep_linter.py',
-    '--pattern=(cudaSetDevice|cudaGetDevice)\\(',
+    '--pattern=cudaSetDevice(',
+    '--pattern=cudaGetDevice(',
     '--linter-name=RAWCUDADEVICE',
     '--error-name=raw CUDA API usage',
     """--error-description=\
@@ -1137,8 +1098,11 @@ command = [
 [[linter]]
 code = 'WORKFLOWSYNC'
 include_patterns = [
-    '.github/workflows/*.yml',
-    '.github/workflows/*.yaml',
+    '.github/workflows/pull.yml',
+    '.github/workflows/trunk.yml',
+    '.github/workflows/periodic.yml',
+    '.github/workflows/mac-mps.yml',
+    '.github/workflows/slow.yml',
 ]
 command = [
     'python3',

CMakeLists.txt

@@ -374,7 +374,7 @@ cmake_dependent_option(
   "Build the lazy Torchscript backend, not compatible with mobile builds" ON
   "NOT INTERN_BUILD_MOBILE" OFF)
 cmake_dependent_option(BUILD_FUNCTORCH "Build Functorch" ON "BUILD_PYTHON" OFF)
-cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin folder"
+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)

CODEOWNERS

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

README.md

@@ -1,4 +1,4 @@
-![PyTorch Logo](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/pytorch-logo-dark.png)
+![PyTorch Logo](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/pytorch-logo-dark.png)
 
 --------------------------------------------------------------------------------
 
@@ -72,7 +72,7 @@ Elaborating Further:
 
 If you use NumPy, then you have used Tensors (a.k.a. ndarray).
 
-![Tensor illustration](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/tensor_illustration.png)
+![Tensor illustration](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/tensor_illustration.png)
 
 PyTorch provides Tensors that can live either on the CPU or the GPU and accelerates the
 computation by a huge amount.
@@ -99,7 +99,7 @@ from several research papers on this topic, as well as current and past work suc
 While this technique is not unique to PyTorch, it's one of the fastest implementations of it to date.
 You get the best of speed and flexibility for your crazy research.
 
-![Dynamic graph](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/dynamic_graph.gif)
+![Dynamic graph](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/dynamic_graph.gif)
 
 ### Python First
 

SECURITY.md

@@ -31,9 +31,9 @@ Be careful when running untrusted models. This classification includes models cr
 
 **Prefer to execute untrusted models within a secure, isolated environment such as a sandbox** (e.g., containers, virtual machines). This helps protect your system from potentially malicious code. You can find further details and instructions in [this page](https://developers.google.com/code-sandboxing).
 
-**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) has a significantly larger surface of attack but is more flexible in what it can serialize. See the documentation for more details.
+**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) with `weights_only=True` is also secure to our knowledge even though it offers significantly larger surface of attack. Loading un-trusted checkpoint with `weights_only=False` MUST never be done.
+
 
-Even for more secure serialization formats, unexpected inputs to the downstream system can cause diverse security threats (e.g. denial of service, out of bound reads/writes) and thus we recommend extensive validation of any untrusted inputs.
 
 Important Note: The trustworthiness of a model is not binary. You must always determine the proper level of caution depending on the specific model and how it matches your use case and risk tolerance.
 

aten/src/ATen/CMakeLists.txt

@@ -38,7 +38,7 @@ set_bool(AT_HIPSPARSELT_ENABLED CAFFE2_USE_HIPSPARSELT)
 
 configure_file(Config.h.in "${CMAKE_CURRENT_SOURCE_DIR}/Config.h")
 # TODO: Do not generate CUDAConfig.h for ROCm BUILDS
-# At the moment, `jit_macros.h` include CUDAConfig.h for both CUDA and HIP builds
+# At the moment, `jit_macors.h` include CUDAConfig.h for both CUDA and HIP builds
 if(USE_CUDA OR USE_ROCM)
   configure_file(cuda/CUDAConfig.h.in "${CMAKE_CURRENT_SOURCE_DIR}/cuda/CUDAConfig.h")
 endif()
@@ -260,7 +260,7 @@ IF(USE_FBGEMM_GENAI)
   if(USE_CUDA)
     # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
     # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
-    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*(mx8mx8bf16_grouped|f4f4bf16_grouped|f4f4bf16).*")
+    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
     file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
@@ -289,16 +289,14 @@ IF(USE_FBGEMM_GENAI)
 
     set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
-    set(fbgemm_genai_cuh
+    set(fbgemm_genai_mx8mx8bf16_grouped
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
-      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/f4f4bf16_grouped/"
-      "${FBGEMM_GENAI_SRCS}/"
     )
 
     target_include_directories(fbgemm_genai PRIVATE
       ${FBGEMM_THIRD_PARTY}/cutlass/include
       ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${fbgemm_genai_cuh}
+      ${fbgemm_genai_mx8mx8bf16_grouped}
       ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
       ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
     )
@@ -315,14 +313,13 @@ IF(USE_FBGEMM_GENAI)
 
     # Add additional HIPCC compiler flags for performance
     set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+      -mllvm
+      -amdgpu-coerce-illegal-types=1
       -mllvm
       -enable-post-misched=0
       -mllvm
       -greedy-reverse-local-assignment=1
       -fhip-new-launch-api)
-    if(DEFINED ROCM_VERSION_DEV AND ROCM_VERSION_DEV VERSION_LESS "7.2.0")
-        list(PREPEND FBGEMM_GENAI_EXTRA_HIPCC_FLAGS -mllvm -amdgpu-coerce-illegal-types=1)
-      endif()
 
     # Only compile for gfx942 for now.
     # This is rather hacky, I could not figure out a clean solution :(

aten/src/ATen/Context.h

@@ -19,7 +19,6 @@
 #include <ATen/detail/MPSHooksInterface.h>
 #include <ATen/detail/MTIAHooksInterface.h>
 #include <ATen/detail/PrivateUse1HooksInterface.h>
-#include <ATen/detail/XLAHooksInterface.h>
 #include <ATen/detail/XPUHooksInterface.h>
 #include <c10/core/QEngine.h>
 #include <c10/core/impl/DeviceGuardImplInterface.h>
@@ -89,8 +88,6 @@ class TORCH_API Context {
       return at::detail::getHIPHooks();
     } else if (opt_device_type == at::kHPU) {
       return at::detail::getHPUHooks();
-    } else if (opt_device_type == at::kXLA) {
-      return at::detail::getXLAHooks();
     } else {
       TORCH_CHECK(
           false,
@@ -199,7 +196,7 @@ class TORCH_API Context {
     return c10::impl::hasDeviceGuardImpl(c10::DeviceType::IPU);
   }
   static bool hasXLA() {
-    return detail::getXLAHooks().hasXLA();
+    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::XLA);
   }
   static bool hasXPU() {
     return detail::getXPUHooks().hasXPU();

aten/src/ATen/FunctionalTensorWrapper.cpp

@@ -122,7 +122,7 @@ void FunctionalTensorWrapper::freeze_storage() const {
 //          |   have their own storages, but backends like functorch      |
 //         \/   are allowed to re-alias underneath the pass               \/
 // . - - - - - - - - - - - - - .                             . - - - - - - - - - - - - - - - .
-// |    underlying_storage     |                             |      underlying_storage       |
+// |    underyling_storage     |                             |      underyling_storage       |
 // . - - - - - - - - - - - - - .                             . - - - - - - - - - - - - - - - .
 //
 // This constructor is only used by view ops.

aten/src/ATen/TensorIterator.cpp

@@ -1534,7 +1534,7 @@ void TensorIteratorBase::build(TensorIteratorConfig& config) {
 
   // XLA and lazy tensors don't have storage, so they don't have an underlying data pointer.
   // Nothing beyond this point is important for meta functions, so it's fine to exit early here.
-  // Extend the condition to MAIA tensors as MAIA tensors also don't have storage.
+  // Extend the condition to MAIA tesnors as MAIA tensors also don't have storage.
   if (privateuse1_without_storage  ||
       common_device_.type() == DeviceType::XLA  ||
       common_device_.type() == DeviceType::IPU  ||

aten/src/ATen/core/CachingHostAllocator.h

@@ -39,7 +39,7 @@ struct HostBlock {
 };
 
 template <typename B>
-struct alignas(hardware_destructive_interference_size) FreeBlockList {
+struct alignas(64) FreeBlockList {
   std::mutex mutex_;
   std::deque<B*> list_;
 };
@@ -94,11 +94,11 @@ struct PinnedReserveSegment {
 struct TORCH_API HostStats {
   // COUNT: total allocations (active)
   Stat active_requests;
-  // SUM: bytes allocated/reserved by this memory allocator. (active)
+  // SUM: bytes allocated/reserved by this memory alocator. (active)
   Stat active_bytes;
   // COUNT: total allocations (active + free)
   Stat allocations;
-  // SUM: bytes allocated/reserved by this memory allocator. This accounts
+  // SUM: bytes allocated/reserved by this memory alocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
 
@@ -122,12 +122,12 @@ struct TORCH_API HostStats {
 // Struct containing memory allocator summary statistics for host, as they
 // are staged for reporting. This is a temporary struct that is used to
 // avoid locking the allocator while collecting stats.
-struct alignas(hardware_destructive_interference_size) HostStatsStaged {
+struct alignas(64) HostStatsStaged {
   std::mutex timing_mutex_;
   // COUNT: total allocations (active + free)
   // LOCK: access to this stat is protected by the allocator's blocks_mutex_
   Stat allocations;
-  // SUM: bytes allocated/reserved by this memory allocator. This accounts
+  // SUM: bytes allocated/reserved by this memory alocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
   // COUNT: number of allocations per bucket (active)
@@ -455,7 +455,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetAccumulatedStats() {
-    // Resetting accumulated memory stats requires concurrently holding both the
+    // Reseting accumulated memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {
@@ -482,7 +482,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetPeakStats() {
-    // Resetting peak memory stats requires concurrently holding both the
+    // Reseting peak memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {
@@ -669,7 +669,7 @@ struct CachingHostAllocatorImpl {
     TORCH_CHECK_NOT_IMPLEMENTED(false, "Not implemented for query_event");
   }
 
-  alignas(hardware_destructive_interference_size) std::mutex blocks_mutex_;
+  alignas(64) std::mutex blocks_mutex_;
   ska::flat_hash_set<B*> blocks_; // block list
   ska::flat_hash_map<void*, B*> ptr_to_block_;
 
@@ -677,17 +677,17 @@ struct CachingHostAllocatorImpl {
   // size. This allows us to quickly find a free block of the right size.
   // We use deque to store per size free list and guard the list with its own
   // mutex.
-  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>>
-      free_list_{MAX_SIZE_INDEX};
+  alignas(64) std::vector<FreeBlockList<B>> free_list_ =
+      std::vector<FreeBlockList<B>>(MAX_SIZE_INDEX);
 
-  alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
+  alignas(64) std::mutex events_mutex_;
   std::deque<std::pair<E, B*>> events_; // event queue paired with block
 
   // Indicates whether the object is active.
   // Set to false in the destructor to signal background threads to stop.
   std::atomic<bool> active_{true};
 protected:
-  alignas(hardware_destructive_interference_size) HostStatsStaged stats_;
+  alignas(64) HostStatsStaged stats_;
 };
 
 struct TORCH_API HostAllocator : public at::Allocator {

aten/src/ATen/core/Generator.h

@@ -59,7 +59,9 @@ struct TORCH_API Generator {
 
   explicit Generator(c10::intrusive_ptr<c10::GeneratorImpl> gen_impl)
    : impl_(std::move(gen_impl)) {
-    TORCH_CHECK(impl_.get(), "GeneratorImpl with nullptr is not supported");
+    if (impl_.get() == nullptr) {
+      throw std::runtime_error("GeneratorImpl with nullptr is not supported");
+    }
   }
 
   bool operator==(const Generator& rhs) const {

aten/src/ATen/core/PhiloxRNGEngine.h

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

aten/src/ATen/core/TensorBase.h

@@ -111,7 +111,9 @@ class TORCH_API TensorBase {
   explicit TensorBase(
       c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl)
       : impl_(std::move(tensor_impl)) {
-    TORCH_CHECK(impl_.get(), "TensorImpl with nullptr is not supported");
+    if (impl_.get() == nullptr) {
+      throw std::runtime_error("TensorImpl with nullptr is not supported");
+    }
   }
   TensorBase(const TensorBase&) = default;
   TensorBase(TensorBase&&) noexcept = default;

aten/src/ATen/core/VariableFallbackKernel.cpp

@@ -109,10 +109,6 @@ TORCH_LIBRARY_IMPL(_, AutogradHPU, m) {
   m.fallback(AUTOGRAD_FALLBACK);
 }
 
-TORCH_LIBRARY_IMPL(_, AutogradPrivateUse1, m) {
-  m.fallback(AUTOGRAD_FALLBACK);
-}
-
 #undef AUTOGRAD_FALLBACK
 
 } // namespace

aten/src/ATen/core/class_type.h

@@ -148,7 +148,7 @@ struct TORCH_API ClassType : public NamedType {
 
   void checkNotExist(const std::string& name, const std::string& what) const;
 
-  // Attributes are stored in a specific slot at runtime for efficiency.
+  // Attributes are stored in a specific slot at runtime for effiency.
   // When emitting instructions we specify the slot so that attribute access is
   // a constant lookup
   std::optional<size_t> findAttributeSlot(const std::string& name) const {
@@ -412,7 +412,7 @@ struct TORCH_API ClassType : public NamedType {
   // Holds method attributes
   std::weak_ptr<CompilationUnit> compilation_unit_;
 
-  // Holds all attributes, attribute details are found on ClassAttribute
+  // Holds all atrributes, attribute details are found on ClassAttribute
   std::vector<ClassAttribute> attributes_;
   // Construct mirroring attributes_, only around due to the fact that `containedTypes()` method returns an ArrayRef.
   // Never fill this without using the appropriate provideNewClassAttribute method

aten/src/ATen/core/dispatch/Dispatcher.cpp

@@ -442,17 +442,11 @@ RegistrationHandleRAII Dispatcher::registerFallback(DispatchKey dispatchKey, Ker
 
   auto idx = getDispatchTableIndexForDispatchKey(dispatchKey);
   TORCH_CHECK(idx >= 0 && static_cast<uint64_t>(idx) < backendFallbackKernels_.size(), "idx=", idx);
-  // NB: Perserve BC for registering fallback for AutogradPrivateUse1 multiple time,
-  // refer to https://github.com/pytorch/pytorch/issues/163979 for more informations.
   TORCH_CHECK(
-      dispatchKey == DispatchKey::AutogradPrivateUse1 ||
-          !backendFallbackKernels_[idx].kernel.isValid(),
-      "Tried to register multiple backend fallbacks for the same dispatch key ",
-      dispatchKey,
-      "; previous registration ",
-      backendFallbackKernels_[idx].debug,
-      ", new registration ",
-      debug);
+    !backendFallbackKernels_[idx].kernel.isValid(),
+    "Tried to register multiple backend fallbacks for the same dispatch key ", dispatchKey, "; previous registration ",
+    backendFallbackKernels_[idx].debug, ", new registration ", debug
+  );
   // NB: inferred function schema is always nullptr for fallbacks, as fallbacks
   // cannot be unboxed
   backendFallbackKernels_[idx] = impl::AnnotatedKernel(std::move(kernel), nullptr, std::move(debug));
@@ -537,7 +531,7 @@ int64_t Dispatcher::sequenceNumberForRunningRecordFunction(DispatchKey dispatchK
 
   // Note: this records a sequence number for both Autograd keys, and for
   // non-Autograd keys where the dispatchKeySet still contains an autograd key.
-  // This means that we might collect the same sequence number two different
+  // This means that we might collect the same sequence nubmer two different
   // events if they all occurred above Autograd and still had the Autograd
   // dispatch key in the dispatch key set.
   // However, this usually doesn't happen: normally the first call will

aten/src/ATen/core/dispatch/Dispatcher.h

@@ -585,7 +585,7 @@ class TORCH_API OperatorHandle {
 
   // We need to store this iterator in order to make
   // Dispatcher::cleanup() fast -- it runs a lot on program
-  // termination (and presumably library unloading).
+  // termination (and presuambly library unloading).
   std::list<Dispatcher::OperatorDef>::iterator operatorIterator_;
 };
 

aten/src/ATen/core/dispatch/OperatorEntry.cpp

@@ -365,7 +365,7 @@ std::pair<const AnnotatedKernel&, const char*> OperatorEntry::computeDispatchTab
   //          For autograd keys, we only use kernel from CompositeImplicitAutograd when there's no direct registration
   //          to its corresponding backend key or CompositeExplicitAutograd. See Note [CompositeExplicitAutograd and CompositeImplicitAutograd].
   //          For AutogradOther, we eagerly return ambiguousAutogradOtherKernel() if there's registration to any of
-  //          its backends and ask backend extender to request a dedicated Autograd key for the backend.
+  //          its backends and ask backend extender to request a decicated Autograd key for the backend.
   //          See Note [Ambiguity in AutogradOther kernel] for more details.
   //          A CompositeExplicitAutograd kernel prevents CompositeImplicitAutograd kernel being used for Autograd keys, but it doesn't
   //          cause confusion for AutogradOther. It's pretty straightforward to use Autograd (if available)

aten/src/ATen/core/function_schema.cpp

@@ -261,7 +261,7 @@ std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema) {
     //
     // There are 2 cases
     // 1. something like 'aten::items.str(Dict(str, t) self) -> ((str, t)[])'.
-    // without the extra parenthesis, the c++ scheme parser can not parse it.
+    // without the extra parenthesis, the c++ schem parser can not parse it.
     // 2. something like '-> ((str, str))'. Need extra parenthesis so the return
     // type is a single tuple rather than two strings.
     // PR (https://github.com/pytorch/pytorch/pull/23204) has more context about

aten/src/ATen/core/interned_strings.cpp

@@ -68,7 +68,11 @@ Symbol InternedStrings::_symbol(const std::string& s) {
     return it->second;
 
   auto pos = s.find("::");
-  TORCH_CHECK(pos != std::string::npos, "all symbols must have a namespace, <namespace>::<string>, but found: ", s);
+  if (pos == std::string::npos) {
+    std::stringstream ss;
+    ss << "all symbols must have a namespace, <namespace>::<string>, but found: " << s;
+    throw std::runtime_error(ss.str());
+  }
   Symbol ns = _symbol("namespaces::" + s.substr(0, pos));
 
   Symbol sym(sym_to_info_.size());
@@ -117,7 +121,12 @@ std::string Symbol::domainString() const {
 }
 
 Symbol Symbol::fromDomainAndUnqualString(const std::string & d, const std::string & s) {
-  TORCH_CHECK(d.compare(0, domain_prefix().size(), domain_prefix()) == 0, "Symbol: domain string is expected to be prefixed with '", domain_prefix(), "', e.g. 'org.pytorch.aten'");
+  if (d.compare(0, domain_prefix().size(), domain_prefix()) != 0) {
+    std::ostringstream ss;
+    ss << "Symbol: domain string is expected to be prefixed with '"
+       << domain_prefix() << "', e.g. 'org.pytorch.aten'";
+    throw std::runtime_error(ss.str());
+  }
   std::string qualString = d.substr(domain_prefix().size()) + "::" + s;
   return fromQualString(qualString);
 }

aten/src/ATen/core/ivalue.cpp

@@ -7,7 +7,6 @@
 #include <ATen/core/jit_type.h>
 #include <ATen/core/stack.h>
 #include <ATen/core/type_factory.h>
-#include <c10/util/Exception.h>
 #include <c10/util/StringUtil.h>
 #include <c10/util/hash.h>
 #include <c10/util/irange.h>
@@ -413,7 +412,7 @@ size_t IValue::hash(const IValue& v) {
     case Tag::Enum:
     case Tag::Stream:
     case Tag::Uninitialized:
-      TORCH_CHECK(false,
+      throw std::runtime_error(
           "unhashable type: '" + v.type()->repr_str() + "'");
   }
   // the above switch should be exhaustive

aten/src/ATen/core/ivalue.h

@@ -1176,7 +1176,7 @@ struct TORCH_API IValue final {
   using HashIdentityIValueMap =
       std::unordered_map<IValue, IValue, HashIdentityIValue, CompIdentityIValues>;
 
-  // Checks if this and rhs has a subvalues in common.
+  // Chechs if this and rhs has a subvalues in common.
   // [t1,t2] and [t2, t3] returns true.
   bool overlaps(const IValue& rhs) const;
 

aten/src/ATen/core/ivalue_inl.h

@@ -1501,7 +1501,7 @@ struct C10_EXPORT ivalue::Object final : c10::intrusive_ptr_target {
   // However, the CompilationUnit holds ownership of the type's graphs, so
   // inserting a constant object into a Graph would create a reference cycle if
   // that constant object held a shared_ptr to its CU. For these objects we
-  // instantiate them with non-owning references to its CU
+  // instatiate them with non-owning references to its CU
   Object(WeakOrStrongTypePtr type, size_t numSlots) : type_(std::move(type)) {
     slots_.resize(numSlots);
   }

aten/src/ATen/core/jit_type.h

@@ -8,7 +8,6 @@
 #include <ATen/core/type_factory.h>
 #include <ATen/core/qualified_name.h>
 #include <c10/util/TypeList.h>
-#include <c10/util/Exception.h>
 #include <optional>
 #include <c10/core/SymFloat.h>
 #include <c10/core/SymBool.h>
@@ -117,8 +116,10 @@ struct SingleElementType : public SharedType {
 
  protected:
   SingleElementType(TypePtr elem) : SharedType(Kind), elem(std::move(elem)) {
-    TORCH_CHECK(this->elem, c10::str(
+    if (!this->elem) {
+      throw std::runtime_error(c10::str(
             "Can not create ", typeKindToString(Kind), " with None type"));
+    }
   }
 
  private:
@@ -373,7 +374,7 @@ struct TORCH_API SymbolicShape {
   // Unranked shape constructor.
   SymbolicShape() : dims_(std::nullopt) {}
 
-  // Known rank but unknown dimensions.
+  // Known rank but unknown dimentions.
   SymbolicShape(std::optional<size_t> rank) : dims_(std::nullopt) {
     if(!rank) {
       return;
@@ -415,12 +416,16 @@ struct TORCH_API SymbolicShape {
   }
 
   ShapeSymbol operator[](size_t i) const {
-    TORCH_CHECK(dims_, "Rank isn't fixed");
+    if (!dims_) {
+      throw std::runtime_error("Rank isn't fixed");
+    }
     return (*dims_).at(i);
   }
 
   ShapeSymbol at(size_t i) const {
-    TORCH_CHECK(dims_, "Rank isn't fixed");
+    if (!dims_) {
+      throw std::runtime_error("Rank isn't fixed");
+    }
     return (*dims_).at(i);
   }
 
@@ -515,7 +520,9 @@ struct VaryingShape {
   }
 
   const std::optional<T> &operator[](size_t i) const {
-    TORCH_CHECK(dims_, "Rank isn't fixed");
+    if (!dims_) {
+      throw std::runtime_error("Rank isn't fixed");
+    }
     return (*dims_).at(i);
   }
 
@@ -884,9 +891,9 @@ struct TORCH_API ListType
 
   // global singleton
   // Given an inner type T and an identifier,
-  // this function will return the global singleton type pointer
+  // this function wil return the global singleton type pointer
   // the type List<T>.
-  // The extra "identifier" argument is needed because we have multiple container types
+  // The extra "identifier" argument is needed beccause we have multiple container types
   // that all re-use this function (List<T>, array<T, N>, etc.)
   static TypePtr get(const std::string& identifier, TypePtr inner);
 
@@ -950,7 +957,9 @@ struct TORCH_API DictType : public SharedType {
 
   TypePtr createWithContained(
       std::vector<TypePtr> contained_types) const override {
-    TORCH_CHECK(contained_types.size() == 2, "Expected 2 contained types");
+    if (contained_types.size() != 2) {
+      throw std::runtime_error("Expected 2 contained types");
+    }
     return create(std::move(contained_types.at(0)), std::move(contained_types.at(1)));
   }
 

aten/src/ATen/core/jit_type_base.h

@@ -185,11 +185,11 @@ struct TORCH_API Type {
         : repr_(nullptr) {}
 
     /* implicit */ SingletonOrSharedTypePtr(SingletonTypePtr<T> p)
-        : repr_(makeSingletonSharedPtr(p.get())) {}
+        : repr_(p) {}
 
     template <typename U, std::enable_if_t<std::is_convertible_v<U*, T*>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(SingletonTypePtr<U> p)
-        : repr_(makeSingletonSharedPtr(static_cast<T*>(p.get()))) {}
+        : repr_(SingletonTypePtr<T>(p.get())) {}
 
 
     // We need to support construction from T* for pybind. The problem
@@ -202,8 +202,8 @@ struct TORCH_API Type {
     // Case 2: if T is exactly Type, we need to do a dynamic_cast to
     // check if it's a SharedType and do the right thing.
     //
-    // Case 3: Otherwise, T is not a SharedType. Use a singleton
-    // pointer.
+    // Case 3: Otherwise, T is not a SharedType. (debug-check this
+    // assumption!) Use a singleton pointer.
 
     template <typename U = T, std::enable_if_t<std::is_base_of_v<SharedType, U>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(T* p) : SingletonOrSharedTypePtr(static_cast<typename detail::as_shared_type<U>::type>(p)->shared_from_this()) {}
@@ -211,15 +211,15 @@ struct TORCH_API Type {
     template <typename U = T, std::enable_if_t<std::is_same_v<Type, U>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(T* p) {
       if (auto* shared_p = dynamic_cast<typename detail::as_shared_type<U>::type>(p)) {
-        repr_ = shared_p->shared_from_this();
+        repr_ = Repr(shared_p->shared_from_this());
       } else {
-        repr_ = makeSingletonSharedPtr(p);
+        repr_ = Repr(p);
       }
     }
 
     template <typename U = T, std::enable_if_t<!std::is_same_v<Type, U> && !std::is_base_of_v<SharedType, U>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(T* p)
-        : repr_(makeSingletonSharedPtr(p)) {
+        : repr_(p) {
       TORCH_INTERNAL_ASSERT_DEBUG_ONLY(dynamic_cast<typename detail::as_shared_type<U>::type>(p) == nullptr);
     }
 
@@ -230,19 +230,19 @@ struct TORCH_API Type {
     ~SingletonOrSharedTypePtr() = default;
 
     T* get() const {
-      return repr_.get();
+      return repr_.isSharedAndNonNull() ? repr_.shared_.repr_.get() : static_cast<T*>(repr_.rawRepr().first);
     }
 
     operator bool() const {
-      return repr_ != nullptr;
+      return repr_.isNonNull();
     }
 
     bool operator==(std::nullptr_t) const {
-      return repr_ == nullptr;
+      return !repr_.isNonNull();
     }
 
     bool operator!=(std::nullptr_t) const {
-      return repr_ != nullptr;
+      return repr_.isNonNull();
     }
 
     template <typename U = T, std::enable_if_t<!std::is_same_v<std::remove_const_t<U>, void>, bool> = true>
@@ -255,14 +255,138 @@ struct TORCH_API Type {
     }
 
   private:
-    // Use shared_ptr's aliasing constructor to create a non-owning pointer
-    // to a singleton. The lifetime is tied to the null shared_ptr, so there's
-    // no reference counting overhead for the singleton itself.
-    static std::shared_ptr<T> makeSingletonSharedPtr(T* ptr) {
-      return std::shared_ptr<T>(std::shared_ptr<T>(), ptr);
-    }
+    // NOTE: SharedPtrWrapper exists to work around a baffling bug in
+    // nvcc; see comment in destroy() below.
+    struct SharedPtrWrapper {
+      SharedPtrWrapper(std::shared_ptr<T> &&x)
+          : repr_(std::move(x)) {}
+      std::shared_ptr<T> repr_;
+    };
+    union Repr {
+      Repr() : Repr(nullptr) {}
 
-    std::shared_ptr<T> repr_;
+      explicit Repr(std::shared_ptr<T> x)
+          : shared_(std::move(x)) {}
+
+      explicit Repr(std::nullptr_t)
+          : singletonRepr_(nullptr) {}
+
+      explicit Repr(SingletonTypePtr<T> p)
+          : singletonRepr_(p.get()) {}
+
+      ~Repr() {
+        destroy();
+      }
+
+      // NOTE: the only non-UB way to access our null state is through
+      // rawRepr(), because our copy operation doesn't preserve which
+      // union member is active for null pointers.
+      Repr(const Repr& rhs) {
+        if (rhs.isSharedAndNonNull()) {
+          new (&shared_) SharedPtrWrapper(rhs.shared_);
+        } else {
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.singletonRepr_.unused_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+      }
+
+      Repr(Repr&& rhs) noexcept {
+        if (rhs.isSharedAndNonNull()) {
+          new (&shared_) SharedPtrWrapper(std::move(rhs.shared_));
+        } else {
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.singletonRepr_.unused_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+      }
+
+      Repr& operator=(const Repr& rhs) {
+        if (&rhs == this) {
+          return *this;
+        }
+        if (rhs.isSharedAndNonNull()) {
+          if (isSharedAndNonNull()) {
+            shared_ = rhs.shared_;
+          } else {
+            new (&shared_) SharedPtrWrapper(rhs.shared_);
+          }
+        } else {
+          if (isSharedAndNonNull()) {
+            destroy();
+          }
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.rawRepr().nullIfSingleton_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+        return *this;
+      }
+
+      Repr& operator=(Repr&& rhs) noexcept {
+        if (&rhs == this) {
+          return *this;
+        }
+        if (rhs.isSharedAndNonNull()) {
+          if (isSharedAndNonNull()) {
+            shared_ = std::move(rhs.shared_);
+          } else {
+            new (&shared_) SharedPtrWrapper(std::move(rhs.shared_));
+          }
+        } else {
+          if (isSharedAndNonNull()) {
+            destroy();
+          }
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.rawRepr().nullIfSingleton_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+        return *this;
+      }
+
+      SharedPtrWrapper shared_;
+
+      struct SingletonRepr {
+        explicit SingletonRepr(T* s) : singleton_(s) {}
+        T* singleton_;
+        void* unused_ = nullptr;
+      } singletonRepr_;
+      struct RawRepr {
+        void* first;
+        void* nullIfSingleton_;
+      };
+
+      // It is UB to read the singleton part of Repr if it was
+      // constructed as a shared_ptr and vice versa, but memcpying out
+      // the representation is always OK, so here's an accessor to obey
+      // the letter of the law.
+      RawRepr rawRepr() const {
+        RawRepr repr{};
+        memcpy(&repr, reinterpret_cast<const char *>(this), sizeof(RawRepr));
+        return repr;
+      }
+
+      bool isNonNull() const {
+        auto repr = rawRepr();
+        TORCH_INTERNAL_ASSERT_DEBUG_ONLY(repr.nullIfSingleton_ == nullptr || repr.first != nullptr);
+        return repr.first != nullptr;
+      }
+
+      bool isSharedAndNonNull() const {
+        return rawRepr().nullIfSingleton_ != nullptr;
+      }
+
+     private:
+      void destroy() {
+        if (isSharedAndNonNull()) {
+          // Without SharedPtrWrapper, this line would read
+          // `shared_.~shared_ptr()` and nvcc would complain with
+          // "error: expected primary-expression before '>' token"
+          // referring to the "t" in "shared_ptr". SharedPtrWrapper
+          // exists to work around this compiler bug.
+          shared_.~SharedPtrWrapper();
+        }
+      }
+    } repr_;
   };
 
   using TypePtr = SingletonOrSharedTypePtr<Type>;

aten/src/ATen/core/op_registration/op_registration.h

@@ -21,7 +21,7 @@ namespace c10 {
 
 namespace detail {
 // The first argument of the schema might be of type DispatchKeySet, in which case we remove it.
-// We do this because every argument in a function schema is expected to be convertible
+// We do this because every argument in a function schema is expected to be convertable
 // to an ivalue, but DispatchKeySet is not a type we want the jit to be aware of.
 // See Note [Plumbing Keys Through The Dispatcher]
 template<class KernelFunctor>

aten/src/ATen/core/op_registration/op_registration_test.cpp

@@ -251,7 +251,7 @@ TEST(OperatorRegistrationTest, whenRegisteringCPUTensorType_thenCanOnlyCallUnbox
   callOpUnboxedWithPrecomputedDispatchKeySet<void, Tensor>(*op, c10::DispatchKeySet(c10::DispatchKey::CPU), dummyTensor(c10::DispatchKey::CUDA));
   EXPECT_TRUE(called_kernel_cpu);
 
-  // Ensure that dispatch key from tensor is not used here.
+  // Ensure that disptach key from tensor is not used here.
   called_kernel_cpu = false;
   expectThrows<c10::Error>([&] {
     callOpUnboxedWithPrecomputedDispatchKeySet<void, Tensor>(*op, c10::DispatchKeySet(c10::DispatchKey::CUDA), dummyTensor(c10::DispatchKey::CPU));