2025-10-30 nightly release (0918bf321c2cfaebb91c5f9e5f0ebe462962328e)

.ci/aarch64_linux/aarch64_ci_build.sh

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

.ci/docker/build.sh

@@ -113,6 +113,7 @@ case "$tag" in
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
+    INSTALL_MINGW=yes
     ;;
   pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
     CUDA_VERSION=13.0.0
@@ -181,7 +182,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
+    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950;gfx1100"
     if [[ $tag =~ "benchmarks" ]]; then
       INDUCTOR_BENCHMARKS=yes
     fi
@@ -344,7 +345,7 @@ docker build \
        --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
-       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH:-gfx90a;gfx942;gfx1100}" \
+       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}" \
        --build-arg "IMAGE_NAME=${IMAGE_NAME}" \
        --build-arg "UCX_COMMIT=${UCX_COMMIT}" \
        --build-arg "UCC_COMMIT=${UCC_COMMIT}" \
@@ -361,6 +362,7 @@ docker build \
        --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
        --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
+       --build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \
        -t "$tmp_tag" \
        "$@" \

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-e0dda9059d082537cee36be6c5e4fe3b18c880c0
+deb42f2a8e48f5032b4a98ee781a15fa87a157cf

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-27664085f804afc83df26f740bb46c365854f2c4
+7416ffcb92cdbe98d9f97e4e6f95247e46dfc9fd

.ci/docker/common/install_conda.sh

@@ -49,12 +49,20 @@ 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

@@ -83,10 +83,6 @@ 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=3.3.24
+NVSHMEM_VERSION=3.4.5
 
 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 13.0.0 cuda_13.0.0_580.65.06_linux
+  install_cuda 13.0.2 cuda_13.0.2_580.95.05_linux
 
   # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
   install_cudnn 13 $CUDNN_VERSION

.ci/docker/common/install_mingw.sh

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

.ci/docker/common/install_onnx.sh

@@ -19,8 +19,8 @@ pip_install \
   transformers==4.36.2
 
 pip_install coloredlogs packaging
-pip_install onnxruntime==1.23.0
-pip_install onnxscript==0.5.3
+pip_install onnxruntime==1.23.1
+pip_install onnxscript==0.5.4
 
 # 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

@@ -40,11 +40,7 @@ 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/libtorch/build.sh

@@ -39,9 +39,13 @@ 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="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

.ci/docker/manywheel/build.sh

@@ -75,9 +75,13 @@ 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="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
         fi
         TARGET=rocm_final
         MANY_LINUX_VERSION="2_28"

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

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

.ci/docker/requirements-ci.txt

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

.ci/docker/ubuntu/Dockerfile

@@ -100,9 +100,16 @@ 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 as e:
-        logger.error("Git operation failed: %s", e)
+    except GitCommandError:
+        logger.exception("Git operation failed")
         raise
 
 

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

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

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

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

.ci/lumen_cli/pyproject.toml

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

.ci/manywheel/build_cuda.sh

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

.ci/pytorch/build.sh

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

.ci/pytorch/macos-test.sh

@@ -256,7 +256,7 @@ test_torchbench_smoketest() {
   local device=mps
   local dtypes=(undefined float16 bfloat16 notset)
   local dtype=${dtypes[$1]}
-  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor timm_resnet timm_vovnet vgg16)
+  local models=(llama BERT_pytorch dcgan yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor vgg16)
 
   for backend in eager inductor; do
 
@@ -319,7 +319,7 @@ test_aoti_torchbench_smoketest() {
   local device=mps
   local dtypes=(undefined float16 bfloat16 notset)
   local dtype=${dtypes[$1]}
-  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor timm_resnet timm_vovnet vgg16)
+  local models=(llama BERT_pytorch dcgan yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor vgg16)
 
   echo "Launching torchbench inference performance run for AOT Inductor and dtype ${dtype}"
   local dtype_arg="--${dtype}"

.ci/pytorch/test.sh

@@ -337,13 +337,13 @@ test_python() {
 
 test_python_smoke() {
   # Smoke tests for H100/B200
-  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 
 test_python_smoke_b200() {
   # Targeted smoke tests for B200 - staged approach to avoid too many failures
-  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 
@@ -460,31 +460,37 @@ test_inductor_shard() {
     --verbose
 }
 
-test_inductor_aoti() {
-  # docker build uses bdist_wheel which does not work with test_aot_inductor
-  # TODO: need a faster way to build
+test_inductor_aoti_cpp() {
   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 "$NUM_TEST_SHARDS" ]]; then
     echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@@ -838,7 +844,7 @@ test_dynamo_benchmark() {
       elif [[ "${suite}" == "timm_models" ]]; then
         export TORCHBENCH_ONLY_MODELS="inception_v3"
       elif [[ "${suite}" == "torchbench" ]]; then
-        export TORCHBENCH_ONLY_MODELS="hf_Bert"
+        export TORCHBENCH_ONLY_MODELS="BERT_pytorch"
       fi
     fi
     test_single_dynamo_benchmark "dashboard" "$suite" "$shard_id" "$@"
@@ -869,13 +875,13 @@ test_inductor_torchbench_smoketest_perf() {
   mkdir -p "$TEST_REPORTS_DIR"
 
   python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --float16 --training \
-    --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" --only hf_Bert \
+    --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" --only BERT_pytorch \
     --output "$TEST_REPORTS_DIR/inductor_training_smoketest.csv"
   # The threshold value needs to be actively maintained to make this check useful
   python benchmarks/dynamo/check_perf_csv.py -f "$TEST_REPORTS_DIR/inductor_training_smoketest.csv" -t 1.4
 
   # Check memory compression ratio for a few models
-  for test in hf_Albert timm_vision_transformer; do
+  for test in BERT_pytorch yolov3; do
     python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --amp --training \
       --disable-cudagraphs --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" \
       --only $test --output "$TEST_REPORTS_DIR/inductor_training_smoketest_$test.csv"
@@ -900,7 +906,7 @@ test_inductor_set_cpu_affinity(){
   export LD_PRELOAD="$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 [[ "${TEST_CONFIG}" != *aarch64* ]]; then
+  if [[ "$(uname -m)" != "aarch64" ]]; then
     # Use Intel OpenMP for x86
     IOMP_LIB="$(dirname "$(which python)")/../lib/libiomp5.so"
     export LD_PRELOAD="$IOMP_LIB":"$LD_PRELOAD"
@@ -914,7 +920,7 @@ test_inductor_set_cpu_affinity(){
   cores=$((cpus / thread_per_core))
 
   # Set number of cores to 16 on aarch64 for performance runs
-  if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
+  if [[ "$(uname -m)" == "aarch64" && $cores -gt 16 ]]; then
     cores=16
   fi
   export OMP_NUM_THREADS=$cores
@@ -1615,6 +1621,7 @@ 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
 
@@ -1629,7 +1636,7 @@ test_operator_benchmark() {
   pip_install pandas
   python check_perf_csv.py \
       --actual "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
-      --expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
+      --expected "${ARCH}_expected_ci_operator_benchmark_eager_float32_cpu.csv"
 }
 
 test_operator_microbenchmark() {
@@ -1666,7 +1673,7 @@ 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* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
+elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]]; then
   test_linux_aarch64
 elif [[ "${TEST_CONFIG}" == *backward* ]]; then
   test_forward_backward_compatibility
@@ -1717,6 +1724,8 @@ elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
   test_inductor_micro_benchmark
+elif [[ "${TEST_CONFIG}" == *aoti_cross_compile_for_windows* ]]; then
+  test_inductor_aoti_cross_compile_for_windows
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
   install_torchvision
   id=$((SHARD_NUMBER-1))
@@ -1757,7 +1766,7 @@ elif [[ "${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
+    test_inductor_aoti_cpp
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision

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

@@ -15,37 +15,35 @@ if errorlevel 1 exit /b 1
 if not errorlevel 0 exit /b 1
 
 cd %TMP_DIR_WIN%\build\torch\test
+
+:: Enable delayed variable expansion to make the list
+setlocal enabledelayedexpansion
+set EXE_LIST=
 for /r "." %%a in (*.exe) do (
-    call :libtorch_check "%%~na" "%%~fa"
+  if "%%~na" == "c10_intrusive_ptr_benchmark" (
+    @REM NB: This is not a gtest executable file, thus couldn't be handled by
+    @REM pytest-cpp and is excluded from test discovery by run_test
+    call "%%~fa"
     if errorlevel 1 goto fail
+    if not errorlevel 0 goto fail
+  ) else (
+    if "%%~na" == "verify_api_visibility" (
+      @REM Skip verify_api_visibility as it is a compile-level test
+    ) else (
+      set EXE_LIST=!EXE_LIST! cpp/%%~na
+    )
+  )
 )
 
-goto :eof
-
-:libtorch_check
-
 cd %CWD%
 set CPP_TESTS_DIR=%TMP_DIR_WIN%\build\torch\test
 
-:: Skip verify_api_visibility as it a compile level test
-if "%~1" == "verify_api_visibility" goto :eof
+:: Run python test\run_test.py on the list
+set NO_TD=True && python test\run_test.py --cpp --verbose -i !EXE_LIST!
+if errorlevel 1 goto fail
+if not errorlevel 0 goto fail
 
-echo Running "%~2"
-if "%~1" == "c10_intrusive_ptr_benchmark" (
-  :: NB: This is not a gtest executable file, thus couldn't be handled by pytest-cpp
-  call "%~2"
-  goto :eof
-)
-
-python test\run_test.py --cpp --verbose -i "cpp/%~1"
-if errorlevel 1 (
-  echo %1 failed with exit code %errorlevel%
-  goto fail
-)
-if not errorlevel 0 (
-  echo %1 failed with exit code %errorlevel%
-  goto fail
-)
+goto :eof
 
 :eof
 exit /b 0

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

@@ -7,12 +7,9 @@ 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-3.14.0rc1-amd64.exe"
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.14t"
 ) else (

.circleci/scripts/binary_populate_env.sh

@@ -71,14 +71,7 @@ export PYTORCH_BUILD_NUMBER=1
 
 # Set triton version as part of PYTORCH_EXTRA_INSTALL_REQUIREMENTS
 TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_version.txt)
-
-# Here PYTORCH_EXTRA_INSTALL_REQUIREMENTS is already set for the all the wheel builds hence append TRITON_CONSTRAINT
-TRITON_CONSTRAINT="platform_system == 'Linux' and platform_machine == 'x86_64'"
-
-# CUDA 12.9/13.0 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]] || [[ "$DESIRED_CUDA" == "cu130" ]]; then
-  TRITON_CONSTRAINT="platform_system == 'Linux'"
-fi
+TRITON_CONSTRAINT="platform_system == 'Linux'"
 
 if [[ "$PACKAGE_TYPE" =~ .*wheel.* &&  -n "${PYTORCH_EXTRA_INSTALL_REQUIREMENTS:-}" && ! "$PYTORCH_BUILD_VERSION" =~ .*xpu.* ]]; then
   TRITON_REQUIREMENT="triton==${TRITON_VERSION}; ${TRITON_CONSTRAINT}"
@@ -170,8 +163,13 @@ if [[ "$(uname)" != Darwin ]]; then
   MEMORY_LIMIT_MAX_JOBS=12
   NUM_CPUS=$(( $(nproc) - 2 ))
 
-  # 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} : ${NUM_CPUS} ))}
+  if [[ "$(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
 
   cat >>"$envfile" <<EOL
   export MAX_JOBS="${MAX_JOBS}"

.claude/skills/docstring/SKILL.md

@@ -0,0 +1,359 @@
+---
+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

@@ -0,0 +1,385 @@
+---
+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

@@ -7,16 +7,12 @@ 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,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,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,B904,B905,B906,B907,B908,B910
-    # these ignores are from flake8-comprehensions; please fix!
-    C407,
-    # these ignores are from flake8-logging-format; please fix!
-    G100,G101,G200
+    B007,B008,B017,B019,B023,B028,B903,B905,B906,B907,B908,B910
     # these ignores are from flake8-simplify. please fix or ignore with commented reason
     SIM105,SIM108,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
     # SIM104 is already covered by pyupgrade ruff

.github/ISSUE_TEMPLATE/ci-sev.md

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

.github/ISSUE_TEMPLATE/disable-autorevert.md

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

.github/actions/build-external-packages/action.yml

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

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

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

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

@@ -28,6 +28,10 @@ runs:
         echo "instance-type: $(get_ec2_metadata instance-type)"
         echo "system info $(uname -a)"
 
+    - name: Print GPU info (if present)
+      shell: bash
+      run: if [ -f /usr/bin/nvidia-smi ]; then nvidia-smi; fi
+
     - name: Check if in a container runner
       shell: bash
       id: check_container_runner
@@ -82,37 +86,6 @@ runs:
         # Prune all of the docker images
         docker system prune -af
 
-    - name: Manually resolve download.pytorch.org
-      shell: bash
-      continue-on-error: true
-      run: |
-        set +e
-        set -x
-
-        PT_DOMAIN=download.pytorch.org
-        # TODO: Flaky access to download.pytorch.org https://github.com/pytorch/pytorch/issues/100400,
-        # cleaning this up once the issue is fixed. There are more than one resolved IP here, the last
-        # one is returned at random
-        RESOLVED_IP=$(dig -4 +short "${PT_DOMAIN}" | tail -n1)
-
-        if [ -z "${RESOLVED_IP}" ]; then
-          echo "Couldn't resolve ${PT_DOMAIN}, retrying with Google DNS..."
-          RESOLVED_IP=$(dig -4 +short "${PT_DOMAIN}" @8.8.8.8 | tail -n1)
-
-          if [ -z "${RESOLVED_IP}" ]; then
-            echo "Couldn't resolve ${PT_DOMAIN}, exiting..."
-            exit 1
-          fi
-        fi
-
-        if grep -r "${PT_DOMAIN}" /etc/hosts; then
-          # Clean up any old records first
-          sudo sed -i "/${PT_DOMAIN}/d" /etc/hosts
-        fi
-
-        echo "${RESOLVED_IP} ${PT_DOMAIN}" | sudo tee -a /etc/hosts
-        cat /etc/hosts
-
     - name: Check that the docker daemon is running
       shell: bash
       continue-on-error: true

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

@@ -111,3 +111,23 @@ runs:
         # This video group ID maps to subgid 1 inside the docker image due to the /etc/subgid entries.
         # The group name corresponding to group ID 1 can change depending on the OS, so both are necessary.
         echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd $DEVICE_FLAG --group-add video --group-add $render_gid --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host" >> "${GITHUB_ENV}"
+
+    - name: configure aws credentials
+      id: aws_creds
+      uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+      with:
+        role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+        aws-region: us-east-1
+        role-duration-seconds: 18000
+
+    - name: Login to Amazon ECR
+      id: login-ecr
+      continue-on-error: true
+      uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
+
+    - 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/actions/test-pytorch-binary/action.yml

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

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+69bbe7363897764f9e758d851cd0340147d27f94

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-966da7e46f65d6d49df3e31214470a4fe5cc8e66
+218d2ab791d437309f91e0486eb9fa7f00badc17

.github/ci_commit_pins/vllm.txt

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

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-2a9138a26ee257fef05310ad3fecf7c55fe80d73
+df6798dfb931ce7c7fe5bed2447cd1092a5981af

.github/ci_configs/vllm/Dockerfile

@@ -1,59 +1,71 @@
-# TODO(elainwy): remove this file after the torch nightly dockerfile is in sync in vllm repo
-# The vLLM Dockerfile is used to construct vLLM image against torch nightly and torch main that can be directly used for testing
-
 ARG CUDA_VERSION=12.8.1
 ARG PYTHON_VERSION=3.12
 
 # BUILD_BASE_IMAGE: used to setup python build xformers, and vllm wheels, It can be replaced with a different base image from local machine,
 # by default, it uses the torch-nightly-base stage from this docker image
 ARG BUILD_BASE_IMAGE=torch-nightly-base
-
-# FINAL_BASE_IMAGE: used to set up vllm-instaled environment and build flashinfer,
-# by default, it uses devel-ubuntu22.04 official image.
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
 # The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
 ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
 
-
 #################### TORCH NIGHTLY BASE IMAGE ####################
-# A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
 
 ARG CUDA_VERSION
 ARG PYTHON_VERSION
 ARG GET_PIP_URL
 
-# Install Python and other dependencies
+# Install system dependencies and uv, then create Python virtual environment
 RUN apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-    && add-apt-repository -y ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
+    && apt-get install -y ccache software-properties-common git curl sudo vim python3-pip \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
+    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
+    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
+    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
+    && ln -s /opt/venv/bin/pip /usr/bin/pip \
     && python3 --version && python3 -m pip --version
 
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
-# Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
-RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if command -v apt-get >/dev/null; then \
-        if [ "$current_gcc_version" -lt 10 ]; then \
-            echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-            apt-get update \
-            && apt-get install -y gcc-10 g++-10 \
-            && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
-            && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-        else \
-            echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-        fi \
-    fi \
-    && gcc --version && g++ --version
+RUN apt-get install -y gcc-10 g++-10
+RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10
+RUN <<EOF
+gcc --version
+EOF
 
-# install uv for faster pip installs
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv==0.8.4
+
+ENV UV_HTTP_TIMEOUT=500
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
+#################### TORCH NIGHTLY  BASE IMAGE ####################
+
+
+#################### BASE BUILD IMAGE ####################
+FROM ${BUILD_BASE_IMAGE} AS base
+USER root
+
+ARG CUDA_VERSION
+ARG PYTHON_VERSION
+
+# Only work with PyTorch manylinux builder
+ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
+
+# Install some system dependencies and double check python version
+RUN if command -v apt-get >/dev/null; then \
+        apt-get update -y \
+        && apt-get install -y ccache software-properties-common git wget sudo vim; \
+    else \
+        dnf install -y git wget sudo; \
+    fi \
+    && python3 --version && python3 -m pip --version
+
+# Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     python3 -m pip install uv==0.8.4
 
@@ -62,51 +74,17 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
-#################### TORCH NIGHTLY  BASE IMAGE ####################
-
-
-#################### BASE BUILD IMAGE ####################
-# A base image for building vLLM with torch nightly or torch wheels
-# prepare basic build environment
-FROM ${BUILD_BASE_IMAGE} AS base
-USER root
-
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
-# Install some system dependencies and double check python version
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
-    else \
-        dnf install -y git curl wget sudo; \
-    fi \
-    && python3 --version && python3 -m pip --version
-
-# Install uv for faster pip installs if not existed
-RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version >/dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
-
 WORKDIR /workspace
 
-# install build and runtime dependencies
+# Install build and runtime dependencies
 COPY requirements/common.txt requirements/common.txt
 COPY use_existing_torch.py use_existing_torch.py
 COPY pyproject.toml pyproject.toml
 
-# install build and runtime dependencies without stable torch version
+# Install build and runtime dependencies without stable torch version
 RUN python3 use_existing_torch.py
 
-# default mount file as placeholder, this just avoid the mount error
+# Default mount file as placeholder, this just avoid the mount error
 # change to a different vllm folder if this does not exist anymore
 ARG TORCH_WHEELS_PATH="./requirements"
 ARG PINNED_TORCH_VERSION
@@ -138,56 +116,36 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/common.txt
 
-# Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}
-
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
 
-RUN echo ${TORCH_CUDA_ARCH_LIST}
-RUN echo ${MAX_JOBS}
-RUN pip freeze | grep -E 'ninja'
+RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
+    export TORCH_CUDA_ARCH_LIST='7.5 8.0+PTX 9.0a'
+    git clone https://github.com/facebookresearch/xformers.git
 
-# Build xformers with cuda and torch nightly/wheel
-# following official xformers guidance: https://github.com/facebookresearch/xformers#build
-# sha for https://github.com/facebookresearch/xformers/tree/v0.0.32.post2
-ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
-ENV CCACHE_DIR=/root/.cache/ccache
+    pushd xformers
+    git checkout v0.0.32.post2
+    git submodule update --init --recursive
+    python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose
+    popd
 
-RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/uv \
-    echo 'git clone xformers...' \
-    && git clone https://github.com/facebookresearch/xformers.git --recursive \
-    && cd xformers \
-    && git checkout ${XFORMERS_COMMIT} \
-    && git submodule update --init --recursive \
-    && echo 'finish git clone xformers...' \
-    && rm -rf build \
-    && python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
-    && cd .. \
-    && rm -rf xformers
+    rm -rf xformers
+BASH
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system xformers-dist/*.whl --verbose
+    uv pip install --system xformers-dist/*.whl
 
-# Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
-# track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
 RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
-
 RUN cat torch_build_versions.txt
 RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
-
 #################### BASE BUILD IMAGE ####################
 
 
 #################### WHEEL BUILD IMAGE ####################
-# Image used to build vllm wheel
 FROM base AS build
 ARG TARGETPLATFORM
 
 COPY . .
-
 RUN python3 use_existing_torch.py
 
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -197,20 +155,17 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
     if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
 
-# Max jobs used by Ninja to build extensions
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
-ARG nvcc_threads=4
+ARG nvcc_threads=8
 ENV NVCC_THREADS=$nvcc_threads
-ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 
 ARG USE_SCCACHE
 ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
 ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
 
-# if USE_SCCACHE is set, use sccache to speed up compilation
+# Use sccache to speed up compilation
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
@@ -235,6 +190,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         && sccache --show-stats; \
     fi
 
+ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
+ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
+
 ARG vllm_target_device="cuda"
 ENV VLLM_TARGET_DEVICE=${vllm_target_device}
 ENV CCACHE_DIR=/root/.cache/ccache
@@ -248,17 +206,10 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         export VLLM_DOCKER_BUILD_CONTEXT=1 && \
         python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
     fi
-
-RUN echo "[INFO] Listing current directory:" && \
-    ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
-    cat torch_build_versions.txt
-
 #################### WHEEL BUILD IMAGE ####################
 
 
 ################### VLLM INSTALLED IMAGE ####################
-# Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 USER root
 
@@ -266,7 +217,7 @@ ARG CUDA_VERSION
 ARG PYTHON_VERSION
 ARG GET_PIP_URL
 
-# TODO (huydhn): Only work with PyTorch manylinux builder
+# Only work with PyTorch manylinux builder
 ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
 
 # prepare for environment starts
@@ -275,20 +226,19 @@ WORKDIR /workspace
 # Install Python and other dependencies
 RUN if command -v apt-get >/dev/null; then \
         apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-        && add-apt-repository -y ppa:deadsnakes/ppa \
-        && apt-get update -y \
-        && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-        && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-        && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-        && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-        && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
+        && apt-get install -y ccache software-properties-common git sudo vim python3-pip; \
     else \
-        dnf install -y git curl wget sudo; \
+        dnf install -y git wget sudo; \
     fi \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
+    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
+    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
+    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
+    && ln -s /opt/venv/bin/pip /usr/bin/pip \
     && python3 --version && python3 -m pip --version
 
-# Get the torch versions, and whls used in previous stagtes for consistency
+# Get the torch versions, and whls used in previous stage
 COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
@@ -297,33 +247,29 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
     echo "[INFO] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
-# Install build and runtime dependencies, this is needed for flashinfer install
-COPY requirements/build.txt requirements/build.txt
-COPY use_existing_torch.py use_existing_torch.py
-RUN python3 use_existing_torch.py
-RUN cat requirements/build.txt
-
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version > /dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
+    python3 -m pip install uv==0.8.4
 
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
+# Install build and runtime dependencies, this is needed for flashinfer install
+COPY requirements/build.txt requirements/build.txt
+COPY use_existing_torch.py use_existing_torch.py
+RUN python3 use_existing_torch.py
+RUN cat requirements/build.txt
 
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/build.txt
 
-
 # Default mount file as placeholder, this just avoid the mount error
 ARG TORCH_WHEELS_PATH="./requirements"
-# Install torch, torchaudio and torchvision
-# if TORCH_WHEELS_PATH is default "./requirements", it will pull the nightly versions using pip using torch_build_versions.txt
-# otherwise, it will use the whls from TORCH_WHEELS_PATH from the host machine
+# Install torch, torchaudio and torchvision. If TORCH_WHEELS_PATH is default
+# to ./requirements, it will pull the nightly versions using pip. Otherwise,
+# it will use the local wheels from TORCH_WHEELS_PATH
 RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
@@ -337,6 +283,9 @@ RUN --mount=type=bind,source=${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
@@ -344,18 +293,16 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Install xformers wheel from previous stage
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
-# Build flashinfer from source.
+
+# Build FlashInfer from source
 ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
-# install package for build flashinfer
-# see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
-
-RUN pip freeze | grep -E 'setuptools|packaging|build'
-
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-# Build flashinfer for torch nightly from source around 10 mins
+
+# 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"
-# Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
 ARG FLASHINFER_GIT_REF="v0.2.14.post1"
+
 RUN --mount=type=cache,target=/root/.cache/uv \
     git clone --depth 1 --recursive --shallow-submodules \
         --branch ${FLASHINFER_GIT_REF} \
@@ -367,7 +314,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     && cd .. \
     && rm -rf flashinfer
 
-# install flashinfer python
+# Install FlashInfer
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system wheels/flashinfer/*.whl --verbose
 
@@ -377,49 +324,6 @@ RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm
 ################### VLLM INSTALLED IMAGE ####################
 
 
-#################### UNITTEST IMAGE #############################
-FROM vllm-base as test
-
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
-
-COPY tests/ tests/
-COPY examples examples
-COPY benchmarks benchmarks
-COPY ./vllm/collect_env.py .
-COPY requirements/common.txt requirements/common.txt
-COPY use_existing_torch.py use_existing_torch.py
-COPY pyproject.toml pyproject.toml
-# Install build and runtime dependencies without stable torch version
-COPY requirements/nightly_torch_test.txt requirements/nightly_torch_test.txt
-
-RUN python3 use_existing_torch.py
-
-# install packages
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
-# enable fast downloads from hf (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system hf_transfer
-ENV HF_HUB_ENABLE_HF_TRANSFER 1
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -e tests/vllm_test_utils
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/nightly_torch_test.txt
-
-# Logging to confirm the torch versions
-RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-
-# Logging to confirm all the packages are installed
-RUN pip freeze
-
-#################### UNITTEST IMAGE #############################
-
 #################### EXPORT STAGE ####################
 FROM scratch as export-wheels
 

.github/label_to_label.yml

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

.github/labeler.yml

@@ -133,3 +133,32 @@
 
 "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

@@ -540,6 +540,26 @@
   - 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

@@ -3,6 +3,7 @@ ciflow_tracking_issue: 64124
 ciflow_push_tags:
 - ciflow/b200
 - ciflow/b200-symm-mem
+- ciflow/b200-distributed
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
@@ -15,7 +16,8 @@ ciflow_push_tags:
 - ciflow/inductor-micro-benchmark
 - ciflow/inductor-micro-benchmark-cpu-x86
 - ciflow/inductor-perf-compare
-- ciflow/inductor-perf-test-nightly-rocm
+- ciflow/inductor-perf-test-nightly-rocm-mi300
+- ciflow/inductor-perf-test-nightly-rocm-mi355
 - ciflow/inductor-perf-test-nightly-x86-zen
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
@@ -24,12 +26,15 @@ ciflow_push_tags:
 - ciflow/nightly
 - ciflow/op-benchmark
 - ciflow/periodic
+- ciflow/periodic-rocm-mi200
 - ciflow/periodic-rocm-mi300
 - ciflow/pull
 - ciflow/quantization-periodic
 - ciflow/riscv64
 - ciflow/rocm
 - ciflow/rocm-mi300
+- ciflow/rocm-mi355
+- ciflow/rocm-navi31
 - ciflow/s390
 - ciflow/slow
 - ciflow/torchbench

.github/scripts/filter_test_configs.py

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

.github/scripts/generate_binary_build_matrix.py

@@ -16,16 +16,18 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
-    "13.0": "13.0.0",
+    "12.9": "12.9.1",
+    "13.0": "13.0.2",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
     "12.8": "9",
+    "12.9": "9",
     "13.0": "9",
 }
 
@@ -38,7 +40,7 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "12.9-aarch64", "13.0-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
@@ -54,7 +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==3.3.24; 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'"
@@ -71,27 +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==3.3.24; 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'"
     ),
+    "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'"
+    ),
     "13.0": (
-        "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-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==13.0.0.19; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.15; 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==12.0.3.29; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.2.49; 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==3.3.24; 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'"
+        "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'"
     ),
     "xpu": (
         "intel-cmplr-lib-rt==2025.2.1 | "
@@ -222,7 +241,11 @@ def generate_libtorch_matrix(
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
         elif os == "windows":
-            arches += CUDA_ARCHES
+            # 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
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -286,7 +309,11 @@ def generate_wheels_matrix(
         if os == "linux":
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
-            arches += CUDA_ARCHES + XPU_ARCHES
+            # 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
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -322,7 +349,7 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.8", "12.6"]
+                arch_version in ["13.0", "12.9", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -386,5 +413,6 @@ def generate_wheels_matrix(
 
 
 validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
 validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")

.github/scripts/trymerge.py

@@ -1092,7 +1092,7 @@ class GitHubPR:
         editor = node["editor"]
         return GitHubComment(
             body_text=node["bodyText"],
-            created_at=node["createdAt"] if "createdAt" in node else "",
+            created_at=node.get("createdAt", ""),
             author_login=node["author"]["login"],
             author_url=node["author"].get("url", None),
             author_association=node["authorAssociation"],
@@ -2042,10 +2042,6 @@ def validate_revert(
             f"[{', '.join(allowed_reverters)}], but instead is {author_association}."
         )
 
-    # Raises exception if matching rule is not found, but ignores all status checks
-    find_matching_merge_rule(
-        pr, repo, skip_mandatory_checks=True, skip_internal_checks=True
-    )
     commit_sha = get_pr_commit_sha(repo, pr)
     return (author_login, commit_sha)
 

.github/templates/linux_binary_build_workflow.yml.j2

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

.github/templates/macos_binary_build_workflow.yml.j2

@@ -26,9 +26,8 @@ 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 '3.14.0-rc.2' }}"
+          python-version: "!{{ py_ver.strip('t') + ('.4' if '3.14' not in py_ver else '.0') }}"
           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: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     {%- else %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge.nonephemeral"
     {%- endif %}
     {%- endif %}
     timeout-minutes: !{{ common.timeout_minutes_windows_binary }}

.github/workflows/_docs.yml

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

.github/workflows/_linux-build.yml

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

.github/workflows/_linux-test.yml

@@ -224,6 +224,46 @@ 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
@@ -389,8 +429,6 @@ jobs:
             "${DOCKER_IMAGE}" \
             ${DOCKER_SHELL_CMD}
           )
-          # Propagate download.pytorch.org IP to container
-          grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" sudo bash -c "/bin/cat >> /etc/hosts"
           echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
 
           if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then

.github/workflows/_rocm-test.yml

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

.github/workflows/_win-build.yml

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

.github/workflows/b200-distributed.yml

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

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

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

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

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

.github/workflows/docker-builds.yml

@@ -57,6 +57,7 @@ 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,

.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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -204,6 +204,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_10-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_10-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -224,7 +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==13.0.48; 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'
+      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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -335,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -381,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -407,6 +453,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_11-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_11-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_11-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -427,7 +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==13.0.48; 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'
+      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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -538,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -584,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -610,6 +702,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_12-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_12-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_12-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -630,7 +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==13.0.48; 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'
+      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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -741,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -787,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -813,6 +951,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_13-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -833,7 +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==13.0.48; 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'
+      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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -944,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -990,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1016,6 +1200,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13t-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_13t-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13t-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1036,7 +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==13.0.48; 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'
+      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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1147,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1193,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1219,6 +1449,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_14-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1239,7 +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==13.0.48; 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'
+      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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1350,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1396,7 +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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1422,6 +1698,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14t-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_14t-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14t-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1442,7 +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==13.0.48; 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'
+      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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -248,6 +248,74 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  libtorch-cuda12_9-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-release-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   libtorch-cuda13_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -358,6 +426,9 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -473,6 +544,9 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

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

@@ -127,7 +127,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_6
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_6-test:  # Testing
@@ -193,7 +193,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_8
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_8-test:  # Testing
@@ -241,6 +241,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_10-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_10-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_10-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -259,7 +325,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda13_0-test:  # Testing
@@ -347,6 +413,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.10"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -459,6 +528,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.10"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -721,7 +793,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_6
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_6-test:  # Testing
@@ -787,7 +859,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_8
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_8-test:  # Testing
@@ -835,6 +907,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_11-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_11-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_11-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_11-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -853,7 +991,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda13_0-test:  # Testing
@@ -941,6 +1079,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.11"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1053,6 +1194,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.11"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1315,7 +1459,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_6
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_6-test:  # Testing
@@ -1381,7 +1525,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing
@@ -1429,6 +1573,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_12-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_12-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_12-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_12-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1447,7 +1657,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda13_0-test:  # Testing
@@ -1535,6 +1745,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.12"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1647,6 +1860,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.12"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1909,7 +2125,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_6
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_6-test:  # Testing
@@ -1975,7 +2191,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_8
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_8-test:  # Testing
@@ -2023,6 +2239,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2041,7 +2323,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda13_0-test:  # Testing
@@ -2129,6 +2411,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.13"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2241,6 +2526,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.13"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2503,7 +2791,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_6
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_6-test:  # Testing
@@ -2569,7 +2857,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_8
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_8-test:  # Testing
@@ -2617,6 +2905,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13t-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2635,7 +2989,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda13_0-test:  # Testing
@@ -2723,6 +3077,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.13t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2835,6 +3192,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.13t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3097,7 +3457,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_6
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_6-test:  # Testing
@@ -3163,7 +3523,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_8
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_8-test:  # Testing
@@ -3211,6 +3571,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3229,7 +3655,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda13_0-test:  # Testing
@@ -3317,6 +3743,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.14"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3429,6 +3858,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.14"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3691,7 +4123,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_6
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_6-test:  # Testing
@@ -3757,7 +4189,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_8
       build_environment: linux-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==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_8-test:  # Testing
@@ -3805,6 +4237,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14t-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3823,7 +4321,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda13_0-test:  # Testing
@@ -3911,6 +4409,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.14t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -4023,6 +4524,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.14t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

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

@@ -63,7 +63,6 @@ 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

@@ -59,7 +59,6 @@ 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
@@ -169,7 +168,6 @@ 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
@@ -279,7 +277,6 @@ 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
@@ -389,7 +386,6 @@ 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
@@ -499,7 +495,6 @@ 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
@@ -609,9 +604,8 @@ 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-rc.2"
+          python-version: "3.14.0"
           freethreaded: false
       - name: Checkout PyTorch
         uses: actions/checkout@v4
@@ -719,9 +713,8 @@ 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-rc.2"
+          python-version: "3.14.0"
           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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -791,7 +791,7 @@ jobs:
   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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -791,7 +791,7 @@ jobs:
   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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

.github/workflows/generated-windows-binary-wheel-nightly.yml

@@ -44,7 +44,7 @@ jobs:
   wheel-py3_10-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -279,7 +279,7 @@ jobs:
   wheel-py3_10-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -517,7 +517,7 @@ jobs:
   wheel-py3_10-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -755,7 +755,7 @@ jobs:
   wheel-py3_10-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -993,7 +993,7 @@ jobs:
   wheel-py3_10-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1229,7 +1229,7 @@ jobs:
   wheel-py3_11-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1464,7 +1464,7 @@ jobs:
   wheel-py3_11-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1702,7 +1702,7 @@ jobs:
   wheel-py3_11-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1940,7 +1940,7 @@ jobs:
   wheel-py3_11-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2178,7 +2178,7 @@ jobs:
   wheel-py3_11-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2414,7 +2414,7 @@ jobs:
   wheel-py3_12-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2649,7 +2649,7 @@ jobs:
   wheel-py3_12-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2887,7 +2887,7 @@ jobs:
   wheel-py3_12-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3125,7 +3125,7 @@ jobs:
   wheel-py3_12-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3363,7 +3363,7 @@ jobs:
   wheel-py3_12-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3599,7 +3599,7 @@ jobs:
   wheel-py3_13-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3834,7 +3834,7 @@ jobs:
   wheel-py3_13-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4072,7 +4072,7 @@ jobs:
   wheel-py3_13-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4310,7 +4310,7 @@ jobs:
   wheel-py3_13-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4548,7 +4548,7 @@ jobs:
   wheel-py3_13-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4784,7 +4784,7 @@ jobs:
   wheel-py3_13t-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5019,7 +5019,7 @@ jobs:
   wheel-py3_13t-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5257,7 +5257,7 @@ jobs:
   wheel-py3_13t-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5495,7 +5495,7 @@ jobs:
   wheel-py3_13t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5733,7 +5733,7 @@ jobs:
   wheel-py3_13t-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5969,7 +5969,7 @@ jobs:
   wheel-py3_14-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6204,7 +6204,7 @@ jobs:
   wheel-py3_14-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6442,7 +6442,7 @@ jobs:
   wheel-py3_14-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6680,7 +6680,7 @@ jobs:
   wheel-py3_14-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6918,7 +6918,7 @@ jobs:
   wheel-py3_14-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7154,7 +7154,7 @@ jobs:
   wheel-py3_14t-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7389,7 +7389,7 @@ jobs:
   wheel-py3_14t-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7627,7 +7627,7 @@ jobs:
   wheel-py3_14t-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7865,7 +7865,7 @@ jobs:
   wheel-py3_14t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -8103,7 +8103,7 @@ jobs:
   wheel-py3_14t-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

.github/workflows/h100-distributed.yml

@@ -37,7 +37,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: "linux.12xlarge"
+      runner: "linux.c7i.12xlarge"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90-dist
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '9.0'

.github/workflows/inductor-perf-test-nightly-h100.yml

@@ -2,7 +2,7 @@ name: inductor-perf-nightly-h100
 
 on:
   schedule:
-    - cron: 15 0,12 * * 1-6
+    - cron: 15 0 * * 1-6
     - cron: 0 7 * * 0
   # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
   # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
@@ -130,7 +130,7 @@ jobs:
     name: test-periodically
     uses: ./.github/workflows/_linux-test.yml
     needs: build
-    if: github.event.schedule == '15 0,12 * * 1-6'
+    if: github.event.schedule == '15 0 * * 1-6'
     with:
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true

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

@@ -0,0 +1,132 @@
+name: inductor-perf-nightly-rocm-mi300
+
+on:
+  push:
+    tags:
+      - ciflow/inductor-perf-test-nightly-rocm-mi300/*
+  schedule:
+    - cron: 15 0 * * *
+  # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
+  # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
+  workflow_dispatch:
+    inputs:
+      training:
+        description: Run training (on by default)?
+        required: false
+        type: boolean
+        default: true
+      inference:
+        description: Run inference (on by default)?
+        required: false
+        type: boolean
+        default: true
+      default:
+        description: Run inductor_default?
+        required: false
+        type: boolean
+        default: false
+      dynamic:
+        description: Run inductor_dynamic_shapes?
+        required: false
+        type: boolean
+        default: false
+      cppwrapper:
+        description: Run inductor_cpp_wrapper?
+        required: false
+        type: boolean
+        default: false
+      cudagraphs:
+        description: Run inductor_cudagraphs?
+        required: false
+        type: boolean
+        default: true
+      freezing_cudagraphs:
+        description: Run inductor_cudagraphs with freezing for inference?
+        required: false
+        type: boolean
+        default: false
+      aotinductor:
+        description: Run aot_inductor for inference?
+        required: false
+        type: boolean
+        default: false
+      maxautotune:
+        description: Run inductor_max_autotune?
+        required: false
+        type: boolean
+        default: false
+      benchmark_configs:
+        description: The list of configs used the benchmark
+        required: false
+        type: string
+        default: inductor_huggingface_perf_rocm_mi300,inductor_timm_perf_rocm_mi300,inductor_torchbench_perf_rocm_mi300
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
+  cancel-in-progress: true
+
+permissions: read-all
+
+jobs:
+  get-label-type:
+    name: get-label-type
+    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
+    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
+    with:
+      triggering_actor: ${{ github.triggering_actor }}
+      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
+      curr_branch: ${{ github.head_ref || github.ref_name }}
+      curr_ref_type: ${{ github.ref_type }}
+      opt_out_experiments: lf
+
+  linux-jammy-rocm-py3_10-inductor-benchmark-build:
+    if: github.repository_owner == 'pytorch'
+    name: rocm-py3_10-inductor-benchmark-build
+    uses: ./.github/workflows/_linux-build.yml
+    with:
+      build-environment: linux-jammy-rocm-py3_10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
+      test-matrix: |
+        { include: [
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-inductor-benchmark-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: rocm-py3_10-inductor-benchmark-test
+    uses: ./.github/workflows/_rocm-test.yml
+    needs: linux-jammy-rocm-py3_10-inductor-benchmark-build
+    with:
+      build-environment: linux-jammy-rocm-py3_10
+      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-inductor-benchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-inductor-benchmark-build.outputs.test-matrix }}
+      timeout-minutes: 720
+      # Disable monitor in perf tests for more investigation
+      disable-monitor: true
+      monitor-log-interval: 10
+      monitor-data-collect-interval: 2
+    secrets: inherit

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

@@ -1,11 +1,11 @@
-name: inductor-perf-nightly-rocm
+name: inductor-perf-nightly-rocm-mi355
 
 on:
   push:
     tags:
-      - ciflow/inductor-perf-test-nightly-rocm/*
+      - ciflow/inductor-perf-test-nightly-rocm-mi355/*
   schedule:
-    - cron: 0 7 * * 0,3
+    - cron: 15 0 * * *
   # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
   # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
   workflow_dispatch:
@@ -59,7 +59,7 @@ on:
         description: The list of configs used the benchmark
         required: false
         type: string
-        default: inductor_huggingface_perf_rocm,inductor_timm_perf_rocm,inductor_torchbench_perf_rocm
+        default: inductor_huggingface_perf_rocm_mi355,inductor_timm_perf_rocm_mi355,inductor_torchbench_perf_rocm_mi355
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
@@ -88,23 +88,27 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
+          { 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" },
         ]}
     secrets: inherit
 

.github/workflows/inductor-periodic.yml

@@ -88,7 +88,6 @@ 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

@@ -12,6 +12,7 @@ on:
       - landchecks/*
     tags:
       - ciflow/pull/*
+      - ciflow/trunk/*
   workflow_dispatch:
 
 permissions: read-all
@@ -32,10 +33,12 @@ jobs:
     name: Get changed files
     uses: ./.github/workflows/_get-changed-files.yml
     with:
-      all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') }}
+      all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') || github.event_name == 'push' }}
 
   lintrunner-clang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    # Needed to prevent deduping on HUD
+    name: lintrunner-clang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     # Only run if there are changed files relevant to clangtidy / clangformat
     if: |
@@ -75,6 +78,7 @@ jobs:
   #       fails to find types when it should
   lintrunner-mypy:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    name: lintrunner-mypy-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     # Only run if there are changed files relevant to mypy
     if: |
@@ -99,6 +103,7 @@ jobs:
 
   lintrunner-noclang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    name: lintrunner-noclang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     with:
       timeout: 120
@@ -113,9 +118,9 @@ jobs:
         CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
         echo "Running all other linters"
         if [ "$CHANGED_FILES" = '*' ]; then
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT --all-files" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY --all-files" .github/scripts/lintrunner.sh
         else
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT ${CHANGED_FILES}" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
         fi
 
   quick-checks:

.github/workflows/operator_benchmark.yml

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

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

@@ -0,0 +1,84 @@
+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

@@ -147,15 +147,16 @@ 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.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"] },
+          { 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"] },
         ]}
     secrets: inherit
 
@@ -182,11 +183,11 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11
       test-matrix: |
         { include: [
-          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
           { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
         ]}
     secrets: inherit
@@ -203,37 +204,6 @@ 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

@@ -127,6 +127,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
+      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
@@ -346,7 +347,8 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      sync-tag: linux-xpu-n-build
+      # This should sync with the build in xpu.yml but xpu uses a larger runner
+      # 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

@@ -45,7 +45,6 @@ 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

@@ -1,6 +1,9 @@
 name: rocm-mi355
 
 on:
+  push:
+    tags:
+      - ciflow/rocm-mi355/*
   workflow_dispatch:
   schedule:
     - cron: 30 11,1 * * *  # about 4:30am PDT and 6:30pm PDT
@@ -39,15 +42,14 @@ 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.mi355.2" },
-          { 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" },
+          { 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" },
         ]}
     secrets: inherit
 
@@ -64,5 +66,7 @@ jobs:
       build-environment: linux-noble-rocm-py3.12-mi355
       docker-image: ${{ needs.linux-noble-rocm-py3_12-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-noble-rocm-py3_12-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
+      tests-to-include: >-
+                        ${{ github.event_name == 'schedule' && 'test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor test_matmul_cuda test_scaled_matmul_cuda'
+                           || '' }}
     secrets: inherit

.github/workflows/rocm-navi31.yml

@@ -0,0 +1,75 @@
+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

@@ -26,11 +26,23 @@ 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
@@ -59,29 +71,3 @@ jobs:
       docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
     secrets: inherit
-
-  linux-jammy-rocm-py3_10-gfx1100-test:
-    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' }}
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3_10-gfx1100
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-        ]}
-      tests-to-include: >
-         test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
-         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
-         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
-         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
-         inductor/test_flex_attention inductor/test_max_autotune
-    secrets: inherit

.github/workflows/slow.yml

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

.github/workflows/trunk-tagging.yml

@@ -58,8 +58,10 @@ jobs:
           else
             COMMIT_SHA="${{ github.sha }}"
           fi
-          echo "sha=${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
-          echo "tag_name=trunk/${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
+          {
+            echo "sha=${COMMIT_SHA}"
+            echo "tag_name=trunk/${COMMIT_SHA}"
+          } >> "${GITHUB_OUTPUT}"
 
       - name: Validate commit SHA
         run: |
@@ -87,7 +89,7 @@ jobs:
             echo "✅ Commit ${COMMIT_SHA} is valid (automatic push trigger)"
           fi
 
-      - name: Create and push tag with retry
+      - name: Create and push tag(s) with retry
         id: check_tag
         env:
           TAG_NAME: ${{ steps.commit.outputs.tag_name }}
@@ -112,14 +114,23 @@ jobs:
             return 1
           }
 
-          # 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
+          # Counters for summary reporting
+          created_count=0
+          skipped_count=0
+          failed_count=0
 
-          echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
+          # Always write outputs once on exit
+          finish() {
+            set +e
+            if [ -n "${GITHUB_OUTPUT:-}" ]; then
+              {
+                echo "created_count=${created_count}"
+                echo "skipped_count=${skipped_count}"
+                echo "failed_count=${failed_count}"
+              } >> "${GITHUB_OUTPUT}"
+            fi
+          }
+          trap finish EXIT
 
           # Retry configuration
           MAX_RETRIES=5
@@ -194,31 +205,111 @@ jobs:
             }
           }
 
-          # Execute with retry
-          if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
-            echo "exists=false" >> "${GITHUB_OUTPUT}"
+          # 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
             exit 0
           else
-            echo "Tag creation failed after all retry attempts"
-            exit 1
+            # 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
           fi
 
       - name: Tag creation summary
         if: always()
         run: |
-          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 ${{ steps.commit.outputs.sha }}"
+          if [ "${{ 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
           else
-            echo "❌ Failed to create tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
-          fi
+            if [ "${{ 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 }}"
+            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
           fi

.github/workflows/trunk.yml

@@ -56,7 +56,7 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       build-generates-artifacts: false
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: "linux.4xlarge"
+      runner: "linux.c7i.4xlarge"
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 1 },
@@ -180,16 +180,50 @@ jobs:
       disable-monitor: false
     secrets: inherit
 
-  win-vs2022-cuda12_6-py3-build:
-    name: win-vs2022-cuda12.6-py3
+  win-vs2022-cuda12_8-py3-build:
+    name: win-vs2022-cuda12.8-py3
     uses: ./.github/workflows/_win-build.yml
     needs: get-label-type
     with:
-      build-environment: win-vs2022-cuda12.6-py3
-      cuda-version: "12.6"
+      build-environment: win-vs2022-cuda12.8-py3
+      cuda-version: "12.8"
       runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
     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
@@ -200,6 +234,23 @@ jobs:
       cuda-arch-list: '8.0'
     secrets: inherit
 
+  # Test cross-compiled models with Windows libs extracted from wheel
+  cross-compile-linux-test:
+    name: cross-compile-linux-test
+    uses: ./.github/workflows/_linux-test.yml
+    needs:
+      - linux-jammy-cuda12_8-py3_10-gcc11-build
+      - get-label-type
+      - win-vs2022-cuda12_8-py3-build
+    with:
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
+      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
+      test-matrix: |
+        { include: [
+          { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
+        ]}
+    secrets: inherit
+
   verify-cachebench-cpu-build:
     name: verify-cachebench-cpu-build
     uses: ./.github/workflows/_linux-build.yml
@@ -249,3 +300,14 @@ jobs:
       docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
     secrets: inherit
+
+  linux-jammy-py3_10-gcc11-full-debug-build-only:
+    name: linux-jammy-py3.10-gcc11-full-debug-build-only
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.2xlarge.memory
+      build-environment: linux-jammy-py3.10-gcc11-full-debug-build-only
+      docker-image-name: ci-image:pytorch-linux-jammy-py3.10-gcc11
+    secrets: inherit

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

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

.github/workflows/vllm.yml

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

.github/workflows/xpu.yml

@@ -35,7 +35,7 @@ jobs:
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
       build-environment: linux-jammy-xpu-n-1-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-1-py3
-      runner: linux.12xlarge
+      runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
@@ -56,17 +56,21 @@ jobs:
       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
-      runner: linux.12xlarge
+      runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
-          { 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" },
+          { 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" },
         ]}
     secrets: inherit
 

.gitignore

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

.lintrunner.toml

@@ -209,6 +209,46 @@ 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 = [
@@ -793,8 +833,7 @@ exclude_patterns = [
 command = [
     'python3',
     'tools/linter/adapters/grep_linter.py',
-    '--pattern=cudaSetDevice(',
-    '--pattern=cudaGetDevice(',
+    '--pattern=(cudaSetDevice|cudaGetDevice)\\(',
     '--linter-name=RAWCUDADEVICE',
     '--error-name=raw CUDA API usage',
     """--error-description=\
@@ -1098,11 +1137,8 @@ command = [
 [[linter]]
 code = 'WORKFLOWSYNC'
 include_patterns = [
-    '.github/workflows/pull.yml',
-    '.github/workflows/trunk.yml',
-    '.github/workflows/periodic.yml',
-    '.github/workflows/mac-mps.yml',
-    '.github/workflows/slow.yml',
+    '.github/workflows/*.yml',
+    '.github/workflows/*.yaml',
 ]
 command = [
     'python3',

CMakeLists.txt

@@ -388,9 +388,9 @@ cmake_dependent_option(USE_PRIORITIZED_TEXT_FOR_LD "Use prioritized text linker
 
 option(USE_MIMALLOC "Use mimalloc" OFF)
 # Enable third party mimalloc library to improve memory allocation performance
-# on Windows.
+# on Windows and AArch64.
 option(USE_MIMALLOC_ON_MKL "Use mimalloc on MKL" OFF)
-if(WIN32)
+if(WIN32 OR (CPU_AARCH64 AND NOT APPLE))
   set(USE_MIMALLOC ON)
 
   # Not enable USE_MIMALLOC_ON_MKL due to it caused issue:

CODEOWNERS

@@ -201,3 +201,17 @@ 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/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/pytorch-logo-dark.png)
+![PyTorch Logo](https://github.com/pytorch/pytorch/raw/main/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/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/tensor_illustration.png)
+![Tensor illustration](https://github.com/pytorch/pytorch/raw/main/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/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/dynamic_graph.gif)
+![Dynamic graph](https://github.com/pytorch/pytorch/raw/main/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) 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.
-
+**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.
 
+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/BlasBackend.h

@@ -28,4 +28,19 @@ inline std::ostream& operator<<(std::ostream& stream, at::BlasBackend backend) {
   return stream << BlasBackendToString(backend);
 }
 
+namespace blas {
+
+enum class ScalingType : std::uint8_t {
+  TensorWise, // fp32 scales
+  RowWise, // fp32 scales
+  BlockWise1x16, // fp8_e4m3fn scales
+  BlockWise1x32, // fp8_e8m0fnu scales
+  BlockWise1x128, // fp32 scales
+  BlockWise128x128, // fp32 scales
+};
+
+enum class SwizzleType : std::uint8_t { NO_SWIZZLE = 0, SWIZZLE_32_4_4 = 1 };
+
+} // namespace blas
+
 } // namespace at

aten/src/ATen/CMakeLists.txt

@@ -38,7 +38,7 @@ set_bool(AT_HIPSPARSELT_ENABLED CAFFE2_USE_HIPSPARSELT)
 
 configure_file(Config.h.in "${CMAKE_CURRENT_SOURCE_DIR}/Config.h")
 # TODO: Do not generate CUDAConfig.h for ROCm BUILDS
-# At the moment, `jit_macors.h` include CUDAConfig.h for both CUDA and HIP builds
+# At the moment, `jit_macros.h` include CUDAConfig.h for both CUDA and HIP builds
 if(USE_CUDA OR USE_ROCM)
   configure_file(cuda/CUDAConfig.h.in "${CMAKE_CURRENT_SOURCE_DIR}/cuda/CUDAConfig.h")
 endif()
@@ -256,10 +256,11 @@ endif()
 IF(USE_FBGEMM_GENAI)
   set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
   set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
+
   if(USE_CUDA)
     # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
     # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
-    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
+    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*(mx8mx8bf16_grouped|f4f4bf16_grouped).*")
     file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
@@ -288,62 +289,71 @@ IF(USE_FBGEMM_GENAI)
 
     set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
-    set(fbgemm_genai_mx8mx8bf16_grouped
+    set(fbgemm_genai_cuh
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
+      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/f4f4bf16_grouped/"
+      "${FBGEMM_GENAI_SRCS}/"
     )
 
-    target_include_directories(fbgemm_genai PUBLIC
+    target_include_directories(fbgemm_genai PRIVATE
       ${FBGEMM_THIRD_PARTY}/cutlass/include
       ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${fbgemm_genai_mx8mx8bf16_grouped}
+      ${fbgemm_genai_cuh}
       ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
       ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
     )
-  else()
-    if(USE_ROCM)
-      # Only include the kernels we want to build to avoid increasing binary size.
-      file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-      set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-      # Add additional HIPCC compiler flags for performance
-      set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-        -mllvm
-        -amdgpu-coerce-illegal-types=1
-        -mllvm
-        -enable-post-misched=0
-        -mllvm
-        -greedy-reverse-local-assignment=1
-        -fhip-new-launch-api)
+    # Add FBGEMM_GENAI include directories for torch_ops.h
+    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
+  elseif(USE_ROCM)
+    # Only include the kernels we want to build to avoid increasing binary size.
+    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
+      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
+      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
+    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-      # Only compile for gfx942 for now.
-      # This is rather hacky, I could not figure out a clean solution :(
-      set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
-      string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
-      if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-        list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
+    # Add additional HIPCC compiler flags for performance
+    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+      -mllvm
+      -enable-post-misched=0
+      -mllvm
+      -greedy-reverse-local-assignment=1
+      -fhip-new-launch-api)
+    if(DEFINED ROCM_VERSION_DEV AND ROCM_VERSION_DEV VERSION_LESS "7.2.0")
+        list(PREPEND FBGEMM_GENAI_EXTRA_HIPCC_FLAGS -mllvm -amdgpu-coerce-illegal-types=1)
       endif()
-      set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
 
-      hip_add_library(
-        fbgemm_genai STATIC
-        ${fbgemm_genai_native_rocm_hip}
-        HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
-      set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
-      set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-      target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
-
-      target_include_directories(fbgemm_genai PUBLIC
-        # FBGEMM version of Composable Kernel is used due to some customizations
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-        ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
-        ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
-      )
+    # Only compile for gfx942 for now.
+    # This is rather hacky, I could not figure out a clean solution :(
+    set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
+    string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
+    if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
+      list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
     endif()
+    set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
+
+    hip_add_library(
+      fbgemm_genai STATIC
+      ${fbgemm_genai_native_rocm_hip}
+      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
+    set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
+    set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
+
+    target_include_directories(fbgemm_genai PRIVATE
+      # FBGEMM version of Composable Kernel is used due to some customizations
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
+      ${FBGEMM_THIRD_PARTY}/cutlass/include
+      ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
+      ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
+      ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
+    )
+
+    # Add FBGEMM_GENAI include directories for torch_ops.h
+    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
   endif()
 endif()
 
@@ -692,12 +702,6 @@ if(USE_CUDA AND NOT USE_ROCM)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
 
-  # Add FBGEMM_GENAI include directories for torch_ops.h
-  if(USE_FBGEMM_GENAI)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
-  endif()
-
   if($ENV{ATEN_STATIC_CUDA})
     if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
       list(APPEND ATen_CUDA_DEPENDENCY_LIBS

aten/src/ATen/CPUApplyUtils.h

@@ -144,8 +144,7 @@ inline std::string _all_equal_numel_error(at::ArrayRef<Tensor> tensors) {
 inline bool _apply_preamble(ArrayRef<Tensor> tensors) {
   checkDeviceType("CPU_tensor_apply", tensors, kCPU);
   checkLayout("CPU_tensor_apply", tensors, kStrided);
-  if (!_all_equal_numel(tensors))
-    TORCH_CHECK(false, _all_equal_numel_error(tensors));
+  TORCH_CHECK(_all_equal_numel(tensors), _all_equal_numel_error(tensors));
   // An empty tensor has no elements
   for (auto& t : tensors)
     if (t.numel() == 0)

aten/src/ATen/Context.cpp

@@ -587,20 +587,33 @@ void Context::setROCmFAPreferredBackend(at::ROCmFABackend b) {
   rocm_fa_preferred_backend = b;
 }
 
-bool Context::allowFP16ReductionCuBLAS() const {
+CuBLASReductionOption Context::allowFP16ReductionCuBLAS() const {
   return allow_fp16_reduction_cublas;
 }
 
-void Context::setAllowFP16ReductionCuBLAS(bool b) {
-  allow_fp16_reduction_cublas = b;
+CuBLASReductionOption inline get_reduction_option(bool allow_reduced_precision, bool allow_splitk) {
+  TORCH_CHECK(
+      !(allow_reduced_precision && !allow_splitk),
+      "allow_splitk=False is not supported when reduced precision reductions are enabled");
+  if (allow_reduced_precision) {
+    return CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
+  } else if (allow_splitk) {
+    return CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK;
+  } else {
+    return CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK;
+  }
 }
 
-bool Context::allowBF16ReductionCuBLAS() const {
+void Context::setAllowFP16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
+  allow_fp16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
+}
+
+CuBLASReductionOption Context::allowBF16ReductionCuBLAS() const {
   return allow_bf16_reduction_cublas;
 }
 
-void Context::setAllowBF16ReductionCuBLAS(bool b) {
-  allow_bf16_reduction_cublas = b;
+void Context::setAllowBF16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
+  allow_bf16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
 }
 
 bool Context::allowFP16AccumulationCuBLAS() const {

aten/src/ATen/Context.h

@@ -19,6 +19,7 @@
 #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>
@@ -38,6 +39,12 @@ namespace at {
 class Tensor;
 
 enum class TORCH_API Float32MatmulPrecision { HIGHEST, HIGH, MEDIUM };
+
+enum class CuBLASReductionOption : uint8_t {
+  AllowReducedPrecisionWithSplitK = 0,
+  DisallowReducedPrecisionAllowSplitK = 1,
+  DisallowReducedPrecisionDisallowSplitK = 2,
+};
 enum class TORCH_API Float32Backend { GENERIC, CUDA, MKLDNN };
 enum class TORCH_API Float32Op { ALL, CONV, RNN, MATMUL };
 enum class TORCH_API Float32Precision { NONE, IEEE, TF32, BF16 };
@@ -82,6 +89,8 @@ 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,
@@ -190,7 +199,7 @@ class TORCH_API Context {
     return c10::impl::hasDeviceGuardImpl(c10::DeviceType::IPU);
   }
   static bool hasXLA() {
-    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::XLA);
+    return detail::getXLAHooks().hasXLA();
   }
   static bool hasXPU() {
     return detail::getXPUHooks().hasXPU();
@@ -220,15 +229,15 @@ class TORCH_API Context {
   bool userEnabledMkldnn() const;
   void setUserEnabledMkldnn(bool e);
   bool benchmarkCuDNN() const;
-  void setBenchmarkCuDNN(bool);
+  void setBenchmarkCuDNN(bool /*b*/);
   int benchmarkLimitCuDNN() const;
-  void setBenchmarkLimitCuDNN(int);
+  void setBenchmarkLimitCuDNN(int /*b*/);
   bool immediateMiopen() const;
-  void setImmediateMiopen(bool);
+  void setImmediateMiopen(bool /*b*/);
   bool deterministicCuDNN() const;
-  void setDeterministicCuDNN(bool);
+  void setDeterministicCuDNN(bool /*b*/);
   bool deterministicMkldnn() const;
-  void setDeterministicMkldnn(bool);
+  void setDeterministicMkldnn(bool /*b*/);
   bool userEnabledNNPACK() const;
   void setUserEnabledNNPACK(bool e);
 
@@ -246,32 +255,32 @@ class TORCH_API Context {
   void setSDPPriorityOrder(const std::vector<int64_t>& order);
   std::array<at::SDPBackend, at::num_sdp_backends> sDPPriorityOrder();
 
-  void setSDPUseFlash(bool);
+  void setSDPUseFlash(bool /*e*/);
   bool userEnabledFlashSDP() const;
 
-  void setSDPUseMemEfficient(bool);
+  void setSDPUseMemEfficient(bool /*e*/);
   bool userEnabledMemEfficientSDP() const;
 
-  void setSDPUseMath(bool);
+  void setSDPUseMath(bool /*e*/);
   bool userEnabledMathSDP() const;
 
-  void setSDPUseCuDNN(bool);
+  void setSDPUseCuDNN(bool /*e*/);
   bool userEnabledCuDNNSDP() const;
 
-  void setAllowFP16BF16ReductionMathSDP(bool);
+  void setAllowFP16BF16ReductionMathSDP(bool /*e*/);
   bool allowFP16BF16ReductionMathSDP() const;
 
-  void setSDPUseOverrideable(bool);
+  void setSDPUseOverrideable(bool /*e*/);
   bool userEnabledOverrideableSDP() const;
 
   at::LinalgBackend linalgPreferredBackend() const;
-  void setLinalgPreferredBackend(at::LinalgBackend);
+  void setLinalgPreferredBackend(at::LinalgBackend /*b*/);
 
   at::BlasBackend blasPreferredBackend();
-  void setBlasPreferredBackend(at::BlasBackend);
+  void setBlasPreferredBackend(at::BlasBackend /*b*/);
 
   at::ROCmFABackend getROCmFAPreferredBackend();
-  void setROCmFAPreferredBackend(at::ROCmFABackend);
+  void setROCmFAPreferredBackend(at::ROCmFABackend /*b*/);
 
   // Note [Enabling Deterministic Operations]
   // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -304,9 +313,9 @@ class TORCH_API Context {
 
   bool deterministicAlgorithms() const;
   bool deterministicAlgorithmsWarnOnly() const;
-  void setDeterministicAlgorithms(bool, bool);
+  void setDeterministicAlgorithms(bool /*b*/, bool /*warn_only*/);
   bool deterministicFillUninitializedMemory() const;
-  void setDeterministicFillUninitializedMemory(bool);
+  void setDeterministicFillUninitializedMemory(bool /*b*/);
 
   // Note [Writing Nondeterministic Operations]
   // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -350,19 +359,23 @@ class TORCH_API Context {
       Float32Op op,
       Float32Precision p);
   bool allowTF32CuDNN(std::optional<Float32Op> op = std::nullopt) const;
-  void setAllowTF32CuDNN(bool);
+  void setAllowTF32CuDNN(bool /*b*/);
   bool allowTF32OneDNN() const;
-  void setAllowTF32OneDNN(bool);
+  void setAllowTF32OneDNN(bool /*b*/);
   bool allowTF32CuBLAS() const;
-  void setAllowTF32CuBLAS(bool);
+  void setAllowTF32CuBLAS(bool /*b*/);
   Float32MatmulPrecision float32MatmulPrecision() const;
   Float32Precision float32Precision(Float32Backend backend, Float32Op op) const;
-  bool allowFP16ReductionCuBLAS() const;
-  void setAllowFP16ReductionCuBLAS(bool);
-  bool allowBF16ReductionCuBLAS() const;
-  void setAllowBF16ReductionCuBLAS(bool);
+  CuBLASReductionOption allowFP16ReductionCuBLAS() const;
+  void setAllowFP16ReductionCuBLAS(
+      bool allow_reduced_precision,
+      bool allow_splitk = true);
+  CuBLASReductionOption allowBF16ReductionCuBLAS() const;
+  void setAllowBF16ReductionCuBLAS(
+      bool allow_reduced_precision,
+      bool allow_splitk = true);
   bool allowFP16AccumulationCuBLAS() const;
-  void setAllowFP16AccumulationCuBLAS(bool);
+  void setAllowFP16AccumulationCuBLAS(bool /*b*/);
 
   // Matmuls can use a so-called "persistent" kernel which launches one CUDA
   // block for each SM on the GPU, and each block then iterates over multiple
@@ -374,7 +387,7 @@ class TORCH_API Context {
   // to make matmuls target only a subset of the SMs, so they can fully schedule
   // even next to a comms kernel, and only be a few percent slower.
   std::optional<int32_t> _SMCarveout_EXPERIMENTAL() const;
-  void _setSMCarveout_EXPERIMENTAL(std::optional<int32_t>);
+  void _setSMCarveout_EXPERIMENTAL(std::optional<int32_t> /*c*/);
 
   at::QEngine qEngine() const;
   void setQEngine(at::QEngine e);
@@ -395,7 +408,7 @@ class TORCH_API Context {
   void setDefaultMobileCPUAllocator();
   void unsetDefaultMobileCPUAllocator();
   bool allowFP16ReductionCPU() const;
-  void setAllowFP16ReductionCPU(bool);
+  void setAllowFP16ReductionCPU(bool /*b*/);
 
   // Preserved for BC
   void lazyInitCUDA() {
@@ -452,8 +465,10 @@ class TORCH_API Context {
       : at::Float32MatmulPrecision::HIGHEST;
   int benchmark_limit_cudnn = 10;
   bool allow_tf32_cudnn = true;
-  bool allow_fp16_reduction_cublas = true;
-  bool allow_bf16_reduction_cublas = true;
+  CuBLASReductionOption allow_fp16_reduction_cublas =
+      CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
+  CuBLASReductionOption allow_bf16_reduction_cublas =
+      CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
   bool allow_fp16_accumulation_cublas = false;
   std::optional<int32_t> sm_carveout = std::nullopt;
   bool enabled_mkldnn = true;

aten/src/ATen/DLConvertor.cpp

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