[Dynamo][Hierarchical Compile] Flatten tuple inputs for regions

ghstack-source-id: e99eea21f6c2e02a15b0027ae1cedffbf4003231
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158812

.bc-linter.yml

@@ -1,15 +0,0 @@
-version: 1
-paths:
-include:
-  - "**/*.py"
-exclude:
-  - ".*"
-  - ".*/**"
-  - "**/.*/**"
-  - "**/.*"
-  - "**/_*/**"
-  - "**/_*.py"
-  - "**/test/**"
-  - "**/benchmarks/**"
-  - "**/test_*.py"
-  - "**/*_test.py"

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -7,15 +7,6 @@ if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
 fi
 
-if [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0"
-fi
-
-# Compress the fatbin with -compress-mode=size for CUDA 13
-if [[ "$DESIRED_CUDA" == *"13"* ]]; then
-    export TORCH_NVCC_FLAGS="-compress-mode=size"
-fi
-
 SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
 source $SCRIPTPATH/aarch64_ci_setup.sh
 

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -77,24 +77,21 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     wheelname = os.path.basename(wheel_path)
     os.mkdir(f"{folder}/tmp")
     os.system(f"unzip {wheel_path} -d {folder}/tmp")
-    # Common libraries for all CUDA versions
-    common_libs = [
-        # Non-NVIDIA system libraries
-        "/lib64/libgomp.so.1",
-        "/usr/lib64/libgfortran.so.5",
-        "/acl/build/libarm_compute.so",
-        "/acl/build/libarm_compute_graph.so",
-        # Common CUDA libraries (same for all versions)
-        "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_lapack_core.so.0",
-        "/usr/local/lib/libnvpl_blas_core.so.0",
+    libs_to_copy = [
+        "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
         "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
         "/usr/local/cuda/lib64/libcudnn.so.9",
+        "/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/libcufft.so.11",
+        "/usr/local/cuda/lib64/libcusparse.so.12",
         "/usr/local/cuda/lib64/libcusparseLt.so.0",
+        "/usr/local/cuda/lib64/libcusolver.so.11",
         "/usr/local/cuda/lib64/libcurand.so.10",
         "/usr/local/cuda/lib64/libnccl.so.2",
-        "/usr/local/cuda/lib64/libnvshmem_host.so.3",
+        "/usr/local/cuda/lib64/libnvJitLink.so.12",
+        "/usr/local/cuda/lib64/libnvrtc.so.12",
         "/usr/local/cuda/lib64/libcudnn_adv.so.9",
         "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
         "/usr/local/cuda/lib64/libcudnn_graph.so.9",
@@ -102,41 +99,22 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
         "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
         "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
         "/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
-        "/usr/local/cuda/lib64/libcufile.so.0",
-        "/usr/local/cuda/lib64/libcufile_rdma.so.1",
-        "/usr/local/cuda/lib64/libcusparse.so.12",
+        "/lib64/libgomp.so.1",
+        "/usr/lib64/libgfortran.so.5",
+        "/acl/build/libarm_compute.so",
+        "/acl/build/libarm_compute_graph.so",
+        "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+        "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+        "/usr/local/lib/libnvpl_lapack_core.so.0",
+        "/usr/local/lib/libnvpl_blas_core.so.0",
     ]
 
-    # CUDA version-specific libraries
-    if "130" in desired_cuda:
-        version_specific_libs = [
-            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
-            "/usr/local/cuda/lib64/libcublas.so.13",
-            "/usr/local/cuda/lib64/libcublasLt.so.13",
-            "/usr/local/cuda/lib64/libcudart.so.13",
-            "/usr/local/cuda/lib64/libcufft.so.12",
-            "/usr/local/cuda/lib64/libcusolver.so.12",
-            "/usr/local/cuda/lib64/libnvJitLink.so.13",
-            "/usr/local/cuda/lib64/libnvrtc.so.13",
-            "/usr/local/cuda/lib64/libnvrtc-builtins.so.13.0",
+    if "129" in desired_cuda:
+        libs_to_copy += [
+            "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.9",
+            "/usr/local/cuda/lib64/libcufile.so.0",
+            "/usr/local/cuda/lib64/libcufile_rdma.so.1",
         ]
-    elif "12" in desired_cuda:
-        # Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
-        minor_version = desired_cuda[-1]
-        version_specific_libs = [
-            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
-            "/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/libcufft.so.11",
-            "/usr/local/cuda/lib64/libcusolver.so.11",
-            "/usr/local/cuda/lib64/libnvJitLink.so.12",
-            "/usr/local/cuda/lib64/libnvrtc.so.12",
-            f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
-        ]
-
-    # Combine all libraries
-    libs_to_copy = common_libs + version_specific_libs
 
     # Copy libraries to unzipped_folder/a/lib
     for lib_path in libs_to_copy:
@@ -231,6 +209,8 @@ if __name__ == "__main__":
     # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
     if enable_cuda:
         build_vars += "MAX_JOBS=5 "
+        # nvshmem is broken for aarch64 see https://github.com/pytorch/pytorch/issues/160425
+        build_vars += "USE_NVSHMEM=OFF "
 
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")

.ci/docker/README.md

@@ -120,8 +120,8 @@ If your new Docker image needs a library installed from a specific pinned commit
    If you're introducing a new argument to the Docker build, make sure to add it in the Docker build step in `.ci/docker/build.sh`:
    ```bash
    docker build \
-     ....
-     --build-arg "NEW_ARG_1=${NEW_ARG_1}"
+      ....
+      --build-arg "NEW_ARG_1=${NEW_ARG_1}"
    ```
 
 3. **Update Dockerfile logic**:

.ci/docker/almalinux/Dockerfile

@@ -64,10 +64,6 @@ FROM cuda as cuda12.9
 RUN bash ./install_cuda.sh 12.9
 ENV DESIRED_CUDA=12.9
 
-FROM cuda as cuda13.0
-RUN bash ./install_cuda.sh 13.0
-ENV DESIRED_CUDA=13.0
-
 FROM ${ROCM_IMAGE} as rocm
 ENV PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
 ADD ./common/install_mkl.sh install_mkl.sh
@@ -80,10 +76,10 @@ ADD ./common/install_mnist.sh install_mnist.sh
 RUN bash ./install_mnist.sh
 
 FROM base as all_cuda
+COPY --from=cuda11.8  /usr/local/cuda-11.8 /usr/local/cuda-11.8
 COPY --from=cuda12.6  /usr/local/cuda-12.6 /usr/local/cuda-12.6
 COPY --from=cuda12.8  /usr/local/cuda-12.8 /usr/local/cuda-12.8
 COPY --from=cuda12.9  /usr/local/cuda-12.9 /usr/local/cuda-12.9
-COPY --from=cuda13.0  /usr/local/cuda-13.0 /usr/local/cuda-13.0
 
 # Final step
 FROM ${BASE_TARGET} as final

.ci/docker/build.sh

@@ -168,13 +168,13 @@ case "$tag" in
     TRITON=yes
     ;;
   pytorch-linux-jammy-py3-clang12-onnx)
-    ANACONDA_PYTHON_VERSION=3.10
+    ANACONDA_PYTHON_VERSION=3.9
     CLANG_VERSION=12
     VISION=yes
     ONNX=yes
     ;;
-  pytorch-linux-jammy-py3.10-clang12)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-clang12)
+    ANACONDA_PYTHON_VERSION=3.9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
@@ -209,7 +209,15 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
     ;;
-  pytorch-linux-jammy-xpu-n-1-py3)
+  pytorch-linux-jammy-xpu-2025.0-py3)
+    ANACONDA_PYTHON_VERSION=3.9
+    GCC_VERSION=11
+    VISION=yes
+    XPU_VERSION=2025.0
+    NINJA_VERSION=1.9.0
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-xpu-2025.1-py3)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
@@ -217,14 +225,6 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-xpu-n-py3)
-    ANACONDA_PYTHON_VERSION=3.9
-    GCC_VERSION=11
-    VISION=yes
-    XPU_VERSION=2025.2
-    NINJA_VERSION=1.9.0
-    TRITON=yes
-    ;;
   pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
@@ -234,8 +234,8 @@ case "$tag" in
     DOCS=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
+    ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
     CLANG_VERSION=12
     VISION=yes
@@ -246,8 +246,8 @@ case "$tag" in
     CLANG_VERSION=18
     VISION=yes
     ;;
-  pytorch-linux-jammy-py3.10-gcc11)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-gcc11)
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -288,6 +288,7 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
+    CONDA_CMAKE=yes
     OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific
@@ -298,6 +299,7 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
+    CONDA_CMAKE=yes
     OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific

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

@@ -1,2 +0,0 @@
-transformers==4.54.0
-soxr==0.5.0

.ci/docker/ci_commit_pins/huggingface.txt

@@ -0,0 +1 @@
+v4.54.0

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

@@ -1 +0,0 @@
-v2.27.7-1

.ci/docker/ci_commit_pins/triton-xpu.txt

@@ -1 +1 @@
-a6572fb0be5b9b0a19b0641a0ce05810fa04e44c
+ae324eeac8e102a2b40370e341460f3791353398

.ci/docker/common/install_cuda.sh

@@ -10,7 +10,7 @@ else
   arch_path='sbsa'
 fi
 
-NVSHMEM_VERSION=3.3.24
+NVSHMEM_VERSION=3.3.9
 
 function install_cuda {
   version=$1
@@ -62,16 +62,14 @@ function install_nvshmem {
   mkdir -p "${tmpdir}" && cd "${tmpdir}"
 
   # nvSHMEM license: https://docs.nvidia.com/nvshmem/api/sla.html
-  # This pattern is a lie as it is not consistent across versions, for 3.3.9 it was cuda_ver-arch-nvshhem-ver
-  filename="libnvshmem-linux-${arch_path}-${nvshmem_version}_cuda${cuda_major_version}-archive"
-  suffix=".tar.xz"
-  url="https://developer.download.nvidia.com/compute/nvshmem/redist/libnvshmem/linux-${arch_path}/${filename}${suffix}"
+  filename="libnvshmem_cuda${cuda_major_version}-linux-${arch_path}-${nvshmem_version}"
+  url="https://developer.download.nvidia.com/compute/redist/nvshmem/${nvshmem_version}/builds/cuda${cuda_major_version}/txz/agnostic/${dl_arch}/${filename}.tar.gz"
 
   # download, unpack, install
   wget -q "${url}"
-  tar xf "${filename}${suffix}"
-  cp -a "${filename}/include/"* /usr/local/cuda/include/
-  cp -a "${filename}/lib/"*     /usr/local/cuda/lib64/
+  tar xf "${filename}.tar.gz"
+  cp -a "libnvshmem/include/"* /usr/local/cuda/include/
+  cp -a "libnvshmem/lib/"*     /usr/local/cuda/lib64/
 
   # cleanup
   cd ..
@@ -128,6 +126,74 @@ function install_129 {
   ldconfig
 }
 
+function prune_124 {
+  echo "Pruning CUDA 12.4"
+  #####################################################################################
+  # CUDA 12.4 prune static libs
+  #####################################################################################
+  export NVPRUNE="/usr/local/cuda-12.4/bin/nvprune"
+  export CUDA_LIB_DIR="/usr/local/cuda-12.4/lib64"
+
+  export GENCODE="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+  export GENCODE_CUDNN="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+
+  if [[ -n "$OVERRIDE_GENCODE" ]]; then
+      export GENCODE=$OVERRIDE_GENCODE
+  fi
+  if [[ -n "$OVERRIDE_GENCODE_CUDNN" ]]; then
+      export GENCODE_CUDNN=$OVERRIDE_GENCODE_CUDNN
+  fi
+
+  # all CUDA libs except CuDNN and CuBLAS
+  ls $CUDA_LIB_DIR/ | grep "\.a" | grep -v "culibos" | grep -v "cudart" | grep -v "cudnn" | grep -v "cublas" | grep -v "metis"  \
+      | xargs -I {} bash -c \
+                "echo {} && $NVPRUNE $GENCODE $CUDA_LIB_DIR/{} -o $CUDA_LIB_DIR/{}"
+
+  # prune CuDNN and CuBLAS
+  $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublas_static.a -o $CUDA_LIB_DIR/libcublas_static.a
+  $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublasLt_static.a -o $CUDA_LIB_DIR/libcublasLt_static.a
+
+  #####################################################################################
+  # CUDA 12.4 prune visual tools
+  #####################################################################################
+  export CUDA_BASE="/usr/local/cuda-12.4/"
+  rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2024.1.0 $CUDA_BASE/nsight-systems-2023.4.4/
+}
+
+function prune_126 {
+  echo "Pruning CUDA 12.6"
+  #####################################################################################
+  # CUDA 12.6 prune static libs
+  #####################################################################################
+  export NVPRUNE="/usr/local/cuda-12.6/bin/nvprune"
+  export CUDA_LIB_DIR="/usr/local/cuda-12.6/lib64"
+
+  export GENCODE="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+  export GENCODE_CUDNN="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+
+  if [[ -n "$OVERRIDE_GENCODE" ]]; then
+      export GENCODE=$OVERRIDE_GENCODE
+  fi
+  if [[ -n "$OVERRIDE_GENCODE_CUDNN" ]]; then
+      export GENCODE_CUDNN=$OVERRIDE_GENCODE_CUDNN
+  fi
+
+  # all CUDA libs except CuDNN and CuBLAS
+  ls $CUDA_LIB_DIR/ | grep "\.a" | grep -v "culibos" | grep -v "cudart" | grep -v "cudnn" | grep -v "cublas" | grep -v "metis"  \
+      | xargs -I {} bash -c \
+                "echo {} && $NVPRUNE $GENCODE $CUDA_LIB_DIR/{} -o $CUDA_LIB_DIR/{}"
+
+  # prune CuDNN and CuBLAS
+  $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublas_static.a -o $CUDA_LIB_DIR/libcublas_static.a
+  $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublasLt_static.a -o $CUDA_LIB_DIR/libcublasLt_static.a
+
+  #####################################################################################
+  # CUDA 12.6 prune visual tools
+  #####################################################################################
+  export CUDA_BASE="/usr/local/cuda-12.6/"
+  rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2024.3.2 $CUDA_BASE/nsight-systems-2024.5.1/
+}
+
 function install_128 {
   CUDNN_VERSION=9.8.0.87
   echo "Installing CUDA 12.8.1 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
@@ -146,38 +212,18 @@ function install_128 {
   ldconfig
 }
 
-function install_130 {
-  CUDNN_VERSION=9.12.0.46
-  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
-
-  # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
-  install_cudnn 13 $CUDNN_VERSION
-
-  install_nvshmem 13 $NVSHMEM_VERSION
-
-  CUDA_VERSION=13.0 bash install_nccl.sh
-
-  CUDA_VERSION=13.0 bash install_cusparselt.sh
-
-  ldconfig
-}
-
 # idiomatic parameter and option handling in sh
 while test $# -gt 0
 do
     case "$1" in
-    12.4) install_124;
+    12.4) install_124; prune_124
         ;;
-    12.6|12.6.*) install_126;
+    12.6|12.6.*) install_126; prune_126
         ;;
     12.8|12.8.*) install_128;
         ;;
     12.9|12.9.*) install_129;
         ;;
-    13.0|13.0.*) install_130;
-        ;;
     *) echo "bad argument $1"; exit 1
         ;;
     esac

.ci/docker/common/install_cusparselt.sh

@@ -5,15 +5,7 @@ set -ex
 # cuSPARSELt license: https://docs.nvidia.com/cuda/cusparselt/license.html
 mkdir tmp_cusparselt && cd tmp_cusparselt
 
-if [[ ${CUDA_VERSION:0:4} =~ "13" ]]; then
-    arch_path='sbsa'
-    export TARGETARCH=${TARGETARCH:-$(uname -m)}
-    if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
-        arch_path='x86_64'
-    fi
-    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.8.0.4_cuda13-archive"
-    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
-elif [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
+if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
     arch_path='sbsa'
     export TARGETARCH=${TARGETARCH:-$(uname -m)}
     if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then

.ci/docker/common/install_inductor_benchmark_deps.sh

@@ -5,7 +5,9 @@ set -ex
 source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
 
 function install_huggingface() {
-  pip_install -r huggingface-requirements.txt
+  local version
+  commit=$(get_pinned_commit huggingface)
+  pip_install "git+https://github.com/huggingface/transformers@${commit}"
 }
 
 function install_timm() {
@@ -24,6 +26,9 @@ function install_torchbench() {
 
   python install.py --continue_on_fail
 
+  # soxr comes from https://github.com/huggingface/transformers/pull/39429
+  pip install transformers==4.54.0 soxr==0.5.0
+
   echo "Print all dependencies after TorchBench is installed"
   python -mpip freeze
   popd

.ci/docker/common/install_nccl.sh

@@ -7,8 +7,6 @@ if [[ ${CUDA_VERSION:0:2} == "11" ]]; then
   NCCL_VERSION=$(cat ci_commit_pins/nccl-cu11.txt)
 elif [[ ${CUDA_VERSION:0:2} == "12" ]]; then
   NCCL_VERSION=$(cat ci_commit_pins/nccl-cu12.txt)
-elif [[ ${CUDA_VERSION:0:2} == "13" ]]; then
-  NCCL_VERSION=$(cat ci_commit_pins/nccl-cu13.txt)
 else
   echo "Unexpected CUDA_VERSION ${CUDA_VERSION}"
   exit 1

.ci/docker/common/install_onnx.sh

@@ -19,8 +19,8 @@ pip_install \
   transformers==4.36.2
 
 pip_install coloredlogs packaging
-pip_install onnxruntime==1.22.1
-pip_install onnxscript==0.4.0
+pip_install onnxruntime==1.18.1
+pip_install onnxscript==0.3.1
 
 # 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_triton.sh

@@ -57,7 +57,7 @@ if [ ! -f setup.py ]; then
   cd python
 fi
 
-pip_install pybind11==3.0.1
+pip_install pybind11==2.13.6
 
 # TODO: remove patch setup.py once we have a proper fix for https://github.com/triton-lang/triton/issues/4527
 as_jenkins sed -i -e 's/https:\/\/tritonlang.blob.core.windows.net\/llvm-builds/https:\/\/oaitriton.blob.core.windows.net\/public\/llvm-builds/g' setup.py

.ci/docker/common/install_xpu.sh

@@ -146,11 +146,11 @@ if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
     XPU_DRIVER_VERSION="/lts/2350"
 fi
 
-# Default use Intel® oneAPI Deep Learning Essentials 2025.1
-if [[ "$XPU_VERSION" == "2025.2" ]]; then
-    XPU_PACKAGES="intel-deep-learning-essentials-2025.2"
-else
+# Default use Intel® oneAPI Deep Learning Essentials 2025.0
+if [[ "$XPU_VERSION" == "2025.1" ]]; then
     XPU_PACKAGES="intel-deep-learning-essentials-2025.1"
+else
+    XPU_PACKAGES="intel-deep-learning-essentials-2025.0"
 fi
 
 # The installation depends on the base OS

.ci/docker/libtorch/Dockerfile

@@ -69,11 +69,6 @@ RUN bash ./install_cuda.sh 12.9
 RUN bash ./install_magma.sh 12.9
 RUN ln -sf /usr/local/cuda-12.9 /usr/local/cuda
 
-FROM cuda as cuda13.0
-RUN bash ./install_cuda.sh 13.0
-RUN bash ./install_magma.sh 13.0
-RUN ln -sf /usr/local/cuda-13.0 /usr/local/cuda
-
 FROM cpu as rocm
 ARG ROCM_VERSION
 ARG PYTORCH_ROCM_ARCH

.ci/docker/manywheel/Dockerfile_2_28

@@ -175,6 +175,6 @@ ENV XPU_DRIVER_TYPE ROLLING
 RUN python3 -m pip install --upgrade pip && \
     python3 -mpip install cmake==3.28.4
 ADD ./common/install_xpu.sh install_xpu.sh
-ENV XPU_VERSION 2025.2
+ENV XPU_VERSION 2025.1
 RUN bash ./install_xpu.sh && rm install_xpu.sh
 RUN pushd /opt/_internal && tar -xJf static-libs-for-embedding-only.tar.xz && popd

.ci/docker/manywheel/build.sh

@@ -67,12 +67,6 @@ case ${image} in
         DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=13"
         MANY_LINUX_VERSION="2_28"
         ;;
-    manylinux2_28-builder:cuda13*)
-        TARGET=cuda_final
-        GPU_IMAGE=amd64/almalinux:8
-        DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=13"
-        MANY_LINUX_VERSION="2_28"
-        ;;
     manylinuxaarch64-builder:cuda*)
         TARGET=cuda_final
         GPU_IMAGE=amd64/almalinux:8

.ci/docker/requirements-ci.txt

@@ -263,6 +263,11 @@ scipy==1.14.1 ; python_version >= "3.12"
 #Pinned versions:
 #test that import:
 
+tb-nightly==2.13.0a20230426
+#Description: TensorBoard
+#Pinned versions:
+#test that import:
+
 # needed by torchgen utils
 typing-extensions>=4.10.0
 #Description: type hints for python
@@ -339,7 +344,7 @@ onnx==1.18.0
 #Pinned versions:
 #test that import:
 
-onnxscript==0.4.0
+onnxscript==0.3.1
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -379,7 +384,7 @@ dataclasses_json==0.6.7
 cmake==4.0.0
 #Description: required for building
 
-tlparse==0.4.0
+tlparse==0.3.30
 #Description: required for log parsing
 
 cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"

.ci/docker/ubuntu-rocm/Dockerfile

@@ -96,11 +96,11 @@ ARG ANACONDA_PYTHON_VERSION
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
+COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
 
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION

.ci/docker/ubuntu-xpu/Dockerfile

@@ -56,10 +56,10 @@ RUN rm install_openssl.sh
 ARG INDUCTOR_BENCHMARKS
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
+COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 # Install XPU Dependencies
 ARG XPU_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -96,11 +96,11 @@ RUN rm install_openssl.sh
 ARG INDUCTOR_BENCHMARKS
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
+COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
 
 ARG TRITON
 ARG TRITON_CPU
@@ -181,6 +181,7 @@ COPY --from=pytorch/llvm:9.0.1 /opt/llvm /opt/llvm
 RUN if [ -n "${SKIP_LLVM_SRC_BUILD_INSTALL}" ]; then set -eu; rm -rf /opt/llvm; fi
 
 # AWS specific CUDA build guidance
+ENV TORCH_CUDA_ARCH_LIST Maxwell
 ENV TORCH_NVCC_FLAGS "-Xfatbin -compress-all"
 ENV CUDA_PATH /usr/local/cuda
 

.ci/libtorch/build.sh

@@ -7,4 +7,4 @@ set -ex
 
 SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
-USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh
+USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh

.ci/lumen_cli/cli/build_cli/register_build.py

@@ -2,7 +2,7 @@ import argparse
 import logging
 
 from cli.lib.common.cli_helper import register_targets, RichHelp, TargetSpec
-from cli.lib.core.vllm.vllm_build import VllmBuildRunner
+from cli.lib.core.vllm import VllmBuildRunner
 
 
 logger = logging.getLogger(__name__)

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

@@ -1,143 +0,0 @@
-from __future__ import annotations
-
-import logging
-import os
-import textwrap
-from pathlib import Path
-from typing import TYPE_CHECKING
-
-from cli.lib.common.utils import get_wheels
-from jinja2 import Template
-
-
-if TYPE_CHECKING:
-    from collections.abc import Iterable, Mapping
-
-
-logger = logging.getLogger(__name__)
-
-_TPL_CONTENT = Template(
-    textwrap.dedent("""\
-    ## {{ title }}
-
-    ```{{ lang }}
-    {{ content }}
-    ```
-""")
-)
-
-_TPL_LIST_ITEMS = Template(
-    textwrap.dedent("""\
-    ## {{ title }}
-    {% for it in items %}
-    - {{ it.pkg }}: {{ it.relpath }}
-    {% else %}
-    _(no item found)_
-    {% endfor %}
-    """)
-)
-
-_TPL_TABLE = Template(
-    textwrap.dedent("""\
-    {%- if rows %}
-    | {{ cols | join(' | ') }} |
-    |{%- for _ in cols %} --- |{%- endfor %}
-    {%- for r in rows %}
-    | {%- for c in cols %} {{ r.get(c, "") }} |{%- endfor %}
-    {%- endfor %}
-    {%- else %}
-    _(no data)_
-    {%- endif %}
-""")
-)
-
-
-def gh_summary_path() -> Path | None:
-    """Return the Path to the GitHub step summary file, or None if not set."""
-    p = os.environ.get("GITHUB_STEP_SUMMARY")
-    return Path(p) if p else None
-
-
-def write_gh_step_summary(md: str, *, append_content: bool = True) -> bool:
-    """
-    Write Markdown content to the GitHub Step Summary file if GITHUB_STEP_SUMMARY is set.
-    append_content: default true, if True, append to the end of the file, else overwrite the whole file
-
-    Returns:
-        True if written successfully (in GitHub Actions environment),
-        False if skipped (e.g., running locally where the variable is not set).
-    """
-    sp = gh_summary_path()
-    if not sp:
-        logger.info("[gh-summary] GITHUB_STEP_SUMMARY not set, skipping write.")
-        return False
-
-    md_clean = textwrap.dedent(md).strip() + "\n"
-
-    mode = "a" if append_content else "w"
-    with sp.open(mode, encoding="utf-8") as f:
-        f.write(md_clean)
-    return True
-
-
-def md_heading(text: str, level: int = 2) -> str:
-    """Generate a Markdown heading string with the given level (1-6)."""
-    return f"{'#' * max(1, min(level, 6))} {text}\n"
-
-
-def md_details(summary: str, content: str) -> str:
-    """Generate a collapsible <details> block with a summary and inner content."""
-    return f"<details>\n<summary>{summary}</summary>\n\n{content}\n\n</details>\n"
-
-
-def summarize_content_from_file(
-    output_dir: Path,
-    freeze_file: str,
-    title: str = "Content from file",
-    code_lang: str = "",  # e.g. "text" or "ini"
-) -> bool:
-    f = Path(output_dir) / freeze_file
-    if not f.exists():
-        return False
-    content = f.read_text(encoding="utf-8").strip()
-    md = render_content(content, title=title, lang=code_lang)
-    return write_gh_step_summary(md)
-
-
-def summarize_wheels(path: Path, title: str = "Wheels", max_depth: int = 3):
-    items = get_wheels(path, max_depth=max_depth)
-    if not items:
-        return False
-    md = render_list(items, title=title)
-    return write_gh_step_summary(md)
-
-
-def md_kv_table(rows: Iterable[Mapping[str, str | int | float]]) -> str:
-    """
-    Render a list of dicts as a Markdown table using Jinja template.
-    """
-    rows = list(rows)
-    cols = list({k for r in rows for k in r.keys()})
-    md = _TPL_TABLE.render(cols=cols, rows=rows).strip() + "\n"
-    return md
-
-
-def render_list(
-    items: Iterable[str],
-    *,
-    title: str = "List",
-) -> str:
-    tpl = _TPL_LIST_ITEMS
-    md = tpl.render(title=title, items=items)
-    return md
-
-
-def render_content(
-    content: str,
-    *,
-    title: str = "Content",
-    lang: str = "text",
-) -> str:
-    tpl = _TPL_CONTENT
-    md = tpl.render(title=title, content=content, lang=lang)
-    return md

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

@@ -45,7 +45,7 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
 
         # Checkout pinned commit
         commit = get_post_build_pinned_commit(target)
-        logger.info("Checking out pinned %s commit %s", target, commit)
+        logger.info("Checking out pinned commit %s", commit)
         r.git.checkout(commit)
 
         # Update submodules if requested
@@ -55,7 +55,7 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
                 sm.update(init=True, recursive=True, progress=PrintProgress())
 
         logger.info("Successfully cloned %s", target)
-        return r, commit
+        return r
 
     except GitCommandError as e:
         logger.error("Git operation failed: %s", e)

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

@@ -1,71 +0,0 @@
-import glob
-import logging
-import shlex
-import shutil
-import sys
-from collections.abc import Iterable
-from importlib.metadata import PackageNotFoundError, version  # noqa: UP035
-from typing import Optional, Union
-
-from cli.lib.common.utils import run_command
-
-
-logger = logging.getLogger(__name__)
-
-
-def pip_install_packages(
-    packages: Iterable[str] = (),
-    env=None,
-    *,
-    requirements: Optional[str] = None,
-    constraints: Optional[str] = None,
-    prefer_uv: bool = False,
-) -> None:
-    use_uv = prefer_uv and shutil.which("uv") is not None
-    base = (
-        [sys.executable, "-m", "uv", "pip", "install"]
-        if use_uv
-        else [sys.executable, "-m", "pip", "install"]
-    )
-    cmd = base[:]
-    if requirements:
-        cmd += ["-r", requirements]
-    if constraints:
-        cmd += ["-c", constraints]
-    cmd += list(packages)
-    logger.info("pip installing packages: %s", " ".join(map(shlex.quote, cmd)))
-    run_command(" ".join(map(shlex.quote, cmd)), env=env)
-
-
-def pip_install_first_match(pattern: str, extras: Optional[str] = None, pref_uv=False):
-    wheel = first_matching_pkg(pattern)
-    target = f"{wheel}[{extras}]" if extras else wheel
-    logger.info("Installing %s...", target)
-    pip_install_packages([target], prefer_uv=pref_uv)
-
-
-def run_python(args: Union[str, list[str]], env=None):
-    """
-    Run the python in the current environment.
-    """
-    if isinstance(args, str):
-        args = shlex.split(args)
-    cmd = [sys.executable] + args
-    run_command(" ".join(map(shlex.quote, cmd)), env=env)
-
-
-def pkg_exists(name: str) -> bool:
-    try:
-        pkg_version = version(name)
-        logger.info("%s already exist with version: %s", name, pkg_version)
-        return True
-    except PackageNotFoundError:
-        logger.info("%s is not installed", name)
-        return False
-
-
-def first_matching_pkg(pattern: str) -> str:
-    matches = sorted(glob.glob(pattern))
-    if not matches:
-        raise FileNotFoundError(f"No wheel matching: {pattern}")
-    return matches[0]

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

@@ -7,8 +7,6 @@ import os
 import shlex
 import subprocess
 import sys
-from contextlib import contextmanager
-from pathlib import Path
 from typing import Optional
 
 
@@ -79,61 +77,3 @@ def str2bool(value: Optional[str]) -> bool:
     if value in false_value_set:
         return False
     raise ValueError(f"Invalid string value for boolean conversion: {value}")
-
-
-@contextmanager
-def temp_environ(updates: dict[str, str]):
-    """
-    Temporarily set environment variables and restore them after the block.
-    Args:
-        updates: Dict of environment variables to set.
-    """
-    missing = object()
-    old: dict[str, str | object] = {k: os.environ.get(k, missing) for k in updates}
-    try:
-        os.environ.update(updates)
-        yield
-    finally:
-        for k, v in old.items():
-            if v is missing:
-                os.environ.pop(k, None)
-            else:
-                os.environ[k] = v  # type: ignore[arg-type]
-
-
-@contextmanager
-def working_directory(path: str):
-    """
-    Temporarily change the working directory inside a context.
-    """
-    if not path:
-        # No-op context
-        yield
-        return
-    prev_cwd = os.getcwd()
-    try:
-        os.chdir(path)
-        yield
-    finally:
-        os.chdir(prev_cwd)
-
-
-def get_wheels(
-    output_dir: Path,
-    max_depth: Optional[int] = None,
-) -> list[str]:
-    """Return a list of wheels found in the given output directory."""
-    root = Path(output_dir)
-    if not root.exists():
-        return []
-    items = []
-    for dirpath, _, filenames in os.walk(root):
-        depth = Path(dirpath).relative_to(root).parts
-        if max_depth is not None and len(depth) > max_depth:
-            continue
-        for fname in sorted(filenames):
-            if fname.endswith(".whl"):
-                pkg = fname.split("-")[0]
-                relpath = str((Path(dirpath) / fname).relative_to(root))
-                items.append({"pkg": pkg, "relpath": relpath})
-    return items

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

@@ -13,11 +13,7 @@ from cli.lib.common.envs_helper import (
     env_str_field,
     with_params_help,
 )
-from cli.lib.common.gh_summary import (
-    gh_summary_path,
-    summarize_content_from_file,
-    summarize_wheels,
-)
+from cli.lib.common.git_helper import clone_external_repo
 from cli.lib.common.path_helper import (
     copy,
     ensure_dir_exists,
@@ -26,7 +22,6 @@ from cli.lib.common.path_helper import (
     is_path_exist,
 )
 from cli.lib.common.utils import run_command
-from cli.lib.core.vllm.lib import clone_vllm, summarize_build_info
 
 
 logger = logging.getLogger(__name__)
@@ -47,7 +42,7 @@ class VllmBuildParameters:
     """
 
     # USE_TORCH_WHEEL: when true, use local Torch wheels; requires TORCH_WHEELS_PATH.
-    # Otherwise docker build pull torch nightly during build
+    #  Otherwise docker build pull torch nightly during build
     # TORCH_WHEELS_PATH: directory containing local torch wheels when use_torch_whl is True
     use_torch_whl: bool = env_bool_field("USE_TORCH_WHEEL", True)
     torch_whls_path: Path = env_path_field("TORCH_WHEELS_PATH", "./dist")
@@ -67,7 +62,7 @@ class VllmBuildParameters:
     )
 
     # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
-    output_dir: Path = env_path_field("OUTPUT_DIR", "external/vllm")
+    output_dir: Path = env_path_field("OUTPUT_DIR", "shared")
 
     # --- Build args ----------------------------------------------------------
     target_stage: str = env_str_field("TARGET_STAGE", "export-wheels")
@@ -157,44 +152,18 @@ class VllmBuildRunner(BaseRunner):
         3. run docker build
         """
         inputs = VllmBuildParameters()
-        logger.info("Running vllm build with inputs: %s", inputs)
-        vllm_commit = clone_vllm()
+        clone_vllm()
 
         self.cp_dockerfile_if_exist(inputs)
+
         # cp torch wheels from root direct to vllm workspace if exist
         self.cp_torch_whls_if_exist(inputs)
 
-        # make sure the output dir to store the build artifacts exist
-        ensure_dir_exists(Path(inputs.output_dir))
+        ensure_dir_exists(inputs.output_dir)
 
         cmd = self._generate_docker_build_cmd(inputs)
         logger.info("Running docker build: \n %s", cmd)
-
-        try:
-            run_command(cmd, cwd="vllm", env=os.environ.copy())
-        finally:
-            self.genearte_vllm_build_summary(vllm_commit, inputs)
-
-    def genearte_vllm_build_summary(
-        self, vllm_commit: str, inputs: VllmBuildParameters
-    ):
-        if not gh_summary_path():
-            return logger.info("Skipping, not detect GH Summary env var....")
-        logger.info("Generate GH Summary ...")
-        # summarize vllm build info
-        summarize_build_info(vllm_commit)
-
-        # summarize vllm build artifacts
-        vllm_artifact_dir = inputs.output_dir / "wheels"
-        summarize_content_from_file(
-            vllm_artifact_dir,
-            "build_summary.txt",
-            title="Vllm build env pip package summary",
-        )
-        summarize_wheels(
-            inputs.torch_whls_path, max_depth=3, title="Torch Wheels Artifacts"
-        )
-        summarize_wheels(vllm_artifact_dir, max_depth=3, title="Vllm Wheels Artifacts")
+        run_command(cmd, cwd="vllm", env=os.environ.copy())
 
     def cp_torch_whls_if_exist(self, inputs: VllmBuildParameters) -> str:
         if not inputs.use_torch_whl:
@@ -283,3 +252,12 @@ class VllmBuildRunner(BaseRunner):
                 --progress=plain .
         """
         ).strip()
+
+
+def clone_vllm():
+    clone_external_repo(
+        target="vllm",
+        repo="https://github.com/vllm-project/vllm.git",
+        dst="vllm",
+        update_submodules=True,
+    )

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

@@ -1,256 +0,0 @@
-import logging
-import os
-import textwrap
-from typing import Any
-
-from cli.lib.common.gh_summary import write_gh_step_summary
-from cli.lib.common.git_helper import clone_external_repo
-from cli.lib.common.pip_helper import pip_install_packages
-from cli.lib.common.utils import run_command, temp_environ, working_directory
-from jinja2 import Template
-
-
-logger = logging.getLogger(__name__)
-
-_TPL_VLLM_INFO = Template(
-    textwrap.dedent("""\
-    ##  Vllm against Pytorch CI Test Summary
-    **Vllm Commit**: [{{ vllm_commit }}](https://github.com/vllm-project/vllm/commit/{{ vllm_commit }})
-    {%- if torch_sha %}
-    **Pytorch Commit**: [{{ torch_sha }}](https://github.com/pytorch/pytorch/commit/{{ torch_sha }})
-    {%- endif %}
-""")
-)
-
-
-def sample_vllm_test_library():
-    """
-    Simple sample to unblock the vllm ci development, which is mimic to
-    https://github.com/vllm-project/vllm/blob/main/.buildkite/test-pipeline.yaml
-    see run_test_plan for more details
-    """
-    # TODO(elainewy): Read from yaml file to handle the env and tests for vllm
-    return {
-        "vllm_basic_correctness_test": {
-            "title": "Basic Correctness Test",
-            "id": "vllm_basic_correctness_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "steps": [
-                "pytest -v -s basic_correctness/test_cumem.py",
-                "pytest -v -s basic_correctness/test_basic_correctness.py",
-                "pytest -v -s basic_correctness/test_cpu_offload.py",
-                "VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py",
-            ],
-        },
-        "vllm_basic_models_test": {
-            "title": "Basic models test",
-            "id": "vllm_basic_models_test",
-            "steps": [
-                "pytest -v -s models/test_transformers.py",
-                "pytest -v -s models/test_registry.py",
-                "pytest -v -s models/test_utils.py",
-                "pytest -v -s models/test_vision.py",
-                "pytest -v -s models/test_initialization.py",
-            ],
-        },
-        "vllm_entrypoints_test": {
-            "title": "Entrypoints Test ",
-            "id": "vllm_entrypoints_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "steps": [
-                " ".join(
-                    [
-                        "pytest",
-                        "-v",
-                        "-s",
-                        "entrypoints/llm",
-                        "--ignore=entrypoints/llm/test_lazy_outlines.py",
-                        "--ignore=entrypoints/llm/test_generate.py",
-                        "--ignore=entrypoints/llm/test_generate_multiple_loras.py",
-                        "--ignore=entrypoints/llm/test_collective_rpc.py",
-                    ]
-                ),
-                "pytest -v -s entrypoints/llm/test_lazy_outlines.py",
-                "pytest -v -s entrypoints/llm/test_generate.py ",
-                "pytest -v -s entrypoints/llm/test_generate_multiple_loras.py",
-                "VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode",
-            ],
-        },
-        "vllm_regression_test": {
-            "title": "Regression Test",
-            "id": "vllm_regression_test",
-            "package_install": ["modelscope"],
-            "steps": [
-                "pytest -v -s test_regression.py",
-            ],
-        },
-        "vllm_lora_tp_test_distributed": {
-            "title": "LoRA TP Test (Distributed)",
-            "id": "vllm_lora_tp_test_distributed",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s -x lora/test_chatglm3_tp.py",
-                "echo $VLLM_WORKER_MULTIPROC_METHOD",
-                "pytest -v -s -x lora/test_llama_tp.py",
-                "pytest -v -s -x lora/test_multi_loras_with_tp.py",
-            ],
-        },
-        "vllm_lora_280_failure_test": {
-            "title": "LoRA 280 failure test",
-            "id": "vllm_lora_280_failure_test",
-            "steps": ["pytest -v lora/test_quant_model.py"],
-        },
-        "vllm_multi_model_processor_test": {
-            "title": "Multi-Modal Processor Test",
-            "id": "vllm_multi_model_processor_test",
-            "package_install": ["git+https://github.com/TIGER-AI-Lab/Mantis.git"],
-            "steps": [
-                "pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py",
-            ],
-        },
-        "vllm_pytorch_compilation_unit_tests": {
-            "title": "PyTorch Compilation Unit Tests",
-            "id": "vllm_pytorch_compilation_unit_tests",
-            "steps": [
-                "pytest -v -s compile/test_pass_manager.py",
-                "pytest -v -s compile/test_fusion.py",
-                "pytest -v -s compile/test_fusion_attn.py",
-                "pytest -v -s compile/test_silu_mul_quant_fusion.py",
-                "pytest -v -s compile/test_sequence_parallelism.py",
-                "pytest -v -s compile/test_async_tp.py",
-                "pytest -v -s compile/test_fusion_all_reduce.py",
-                "pytest -v -s compile/test_decorator.py",
-            ],
-        },
-        # TODO(elainewy):need to add g6 with 4 gpus to run this test
-        "vllm_lora_test": {
-            "title": "LoRA Test %N",
-            "id": "lora_test",
-            "parallelism": 4,
-            "steps": [
-                "echo '[checking] list sharded lora tests:'",
-                " ".join(
-                    [
-                        "pytest -q --collect-only lora",
-                        "--shard-id=$$BUILDKITE_PARALLEL_JOB",
-                        "--num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT",
-                        "--ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py",
-                    ]
-                ),
-                "echo '[checking] Done. list lora tests'",
-                " ".join(
-                    [
-                        "pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB",
-                        "--num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT",
-                        "--ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py",
-                    ]
-                ),
-            ],
-        },
-    }
-
-
-def check_parallelism(tests: Any, title: str, shard_id: int = 0, num_shards: int = 0):
-    """
-    a method to check if the test plan is parallelism or not.
-    """
-    parallelism = int(tests.get("parallelism", "0"))
-    is_parallel = parallelism and parallelism > 1
-
-    if not is_parallel:
-        return False
-
-    if shard_id > num_shards:
-        raise RuntimeError(
-            f"Test {title} expects {num_shards} shards, but invalid {shard_id} is provided"
-        )
-
-    if num_shards != parallelism:
-        raise RuntimeError(
-            f"Test {title} expects {parallelism} shards, but invalid {num_shards} is provided"
-        )
-
-    return True
-
-
-def run_test_plan(
-    test_plan: str,
-    test_target: str,
-    tests_map: dict[str, Any],
-    shard_id: int = 0,
-    num_shards: int = 0,
-):
-    """
-    a method to run list of tests based on the test plan.
-    """
-    logger.info("run %s tests.....", test_target)
-    if test_plan not in tests_map:
-        raise RuntimeError(
-            f"test {test_plan} not found, please add it to test plan pool"
-        )
-    tests = tests_map[test_plan]
-    pkgs = tests.get("package_install", [])
-    title = tests.get("title", "unknown test")
-
-    is_parallel = check_parallelism(tests, title, shard_id, num_shards)
-    if is_parallel:
-        title = title.replace("%N", f"{shard_id}/{num_shards}")
-
-    logger.info("Running tests: %s", title)
-    if pkgs:
-        logger.info("Installing packages: %s", pkgs)
-        pip_install_packages(packages=pkgs, prefer_uv=True)
-    with (
-        working_directory(tests.get("working_directory", "tests")),
-        temp_environ(tests.get("env_vars", {})),
-    ):
-        failures = []
-        for step in tests["steps"]:
-            logger.info("Running step: %s", step)
-            if is_parallel:
-                step = replace_buildkite_placeholders(step, shard_id, num_shards)
-                logger.info("Running parallel step: %s", step)
-            code = run_command(cmd=step, check=False, use_shell=True)
-            if code != 0:
-                failures.append(step)
-            logger.info("Finish running step: %s", step)
-        if failures:
-            logger.error("Failed tests: %s", failures)
-            raise RuntimeError(f"{len(failures)} pytest runs failed: {failures}")
-        logger.info("Done. All tests passed")
-
-
-def clone_vllm(dst: str = "vllm"):
-    _, commit = clone_external_repo(
-        target="vllm",
-        repo="https://github.com/vllm-project/vllm.git",
-        dst=dst,
-        update_submodules=True,
-    )
-    return commit
-
-
-def replace_buildkite_placeholders(step: str, shard_id: int, num_shards: int) -> str:
-    mapping = {
-        "$$BUILDKITE_PARALLEL_JOB_COUNT": str(num_shards),
-        "$$BUILDKITE_PARALLEL_JOB": str(shard_id),
-    }
-    for k in sorted(mapping, key=len, reverse=True):
-        step = step.replace(k, mapping[k])
-    return step
-
-
-def summarize_build_info(vllm_commit: str) -> bool:
-    torch_sha = os.getenv("GITHUB_SHA")
-    md = (
-        _TPL_VLLM_INFO.render(vllm_commit=vllm_commit, torch_sha=torch_sha).strip()
-        + "\n"
-    )
-    return write_gh_step_summary(md)

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

@@ -1,263 +0,0 @@
-import logging
-import os
-import re
-import subprocess
-import sys
-from collections.abc import Iterable
-from dataclasses import dataclass
-from enum import Enum
-from pathlib import Path
-from typing import Any
-
-from cli.lib.common.cli_helper import BaseRunner
-from cli.lib.common.envs_helper import env_path_field, env_str_field, get_env
-from cli.lib.common.path_helper import copy, remove_dir
-from cli.lib.common.pip_helper import (
-    pip_install_first_match,
-    pip_install_packages,
-    pkg_exists,
-    run_python,
-)
-from cli.lib.common.utils import run_command, working_directory
-from cli.lib.core.vllm.lib import clone_vllm, run_test_plan, sample_vllm_test_library
-
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class VllmTestParameters:
-    """
-    Parameters defining the vllm external test input
-
-    !!!DO NOT ADD SECRETS IN THIS CLASS!!!
-    you can put environment variable name in VllmTestParameters if it's not the same as the secret one
-    fetch secrests directly from env variables during runtime
-    """
-
-    torch_whls_path: Path = env_path_field("WHEELS_PATH", "./dist")
-
-    vllm_whls_path: Path = env_path_field(
-        "VLLM_WHEELS_PATH", "./dist/external/vllm/wheels"
-    )
-
-    torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
-
-    def __post_init__(self):
-        if not self.torch_whls_path.exists():
-            raise ValueError("missing torch_whls_path")
-        if not self.vllm_whls_path.exists():
-            raise ValueError("missing vllm_whls_path")
-
-
-class TestInpuType(Enum):
-    TEST_PLAN = "test_plan"
-    UNKNOWN = "unknown"
-
-
-class VllmTestRunner(BaseRunner):
-    def __init__(self, args: Any):
-        self.work_directory = "vllm"
-        self.test_plan = ""
-        self.test_type = TestInpuType.UNKNOWN
-
-        self.shard_id = args.shard_id
-        self.num_shards = args.num_shards
-
-        if args.test_plan:
-            self.test_plan = args.test_plan
-            self.test_type = TestInpuType.TEST_PLAN
-
-        # Matches the structeur in the artifacts.zip from torcb build
-        self.TORCH_WHL_PATH_REGEX = "torch*.whl"
-        self.TORCH_WHL_EXTRA = "opt-einsum"
-        self.TORCH_ADDITIONAL_WHLS_REGEX = [
-            "vision/torchvision*.whl",
-            "audio/torchaudio*.whl",
-        ]
-
-        # Match the structure of the artifacts.zip from vllm external build
-        self.VLLM_TEST_WHLS_REGEX = [
-            "xformers/*.whl",
-            "vllm/vllm*.whl",
-            "flashinfer-python/flashinfer*.whl",
-        ]
-
-    def prepare(self):
-        """
-        prepare test environment for vllm. This includes clone vllm repo, install all wheels, test dependencies and set env
-        """
-        params = VllmTestParameters()
-        logger.info("Display VllmTestParameters %s", params)
-        self._set_envs(params)
-
-        clone_vllm(dst=self.work_directory)
-        with working_directory(self.work_directory):
-            remove_dir(Path("vllm"))
-            self._install_wheels(params)
-            self._install_dependencies()
-        # verify the torches are not overridden by test dependencies
-        check_versions()
-
-    def run(self):
-        """
-        main function to run vllm test
-        """
-        self.prepare()
-        with working_directory(self.work_directory):
-            if self.test_type == TestInpuType.TEST_PLAN:
-                if self.num_shards > 1:
-                    run_test_plan(
-                        self.test_plan,
-                        "vllm",
-                        sample_vllm_test_library(),
-                        self.shard_id,
-                        self.num_shards,
-                    )
-                else:
-                    run_test_plan(self.test_plan, "vllm", sample_vllm_test_library())
-            else:
-                raise ValueError(f"Unknown test type {self.test_type}")
-
-    def _install_wheels(self, params: VllmTestParameters):
-        logger.info("Running vllm test with inputs: %s", params)
-        if not pkg_exists("torch"):
-            # install torch from local whls if it's not installed yet.
-            torch_p = f"{str(params.torch_whls_path)}/{self.TORCH_WHL_PATH_REGEX}"
-            pip_install_first_match(torch_p, self.TORCH_WHL_EXTRA)
-
-        torch_whls_path = [
-            f"{str(params.torch_whls_path)}/{whl_path}"
-            for whl_path in self.TORCH_ADDITIONAL_WHLS_REGEX
-        ]
-        for torch_whl in torch_whls_path:
-            pip_install_first_match(torch_whl)
-        logger.info("Done. Installed torch and other torch-related wheels ")
-
-        logger.info("Installing vllm wheels")
-        vllm_whls_path = [
-            f"{str(params.vllm_whls_path)}/{whl_path}"
-            for whl_path in self.VLLM_TEST_WHLS_REGEX
-        ]
-        for vllm_whl in vllm_whls_path:
-            pip_install_first_match(vllm_whl)
-        logger.info("Done. Installed vllm wheels")
-
-    def _install_test_dependencies(self):
-        """
-        This method replaces torch dependencies with local torch wheel info in
-        requirements/test.in file from vllm repo. then generates the test.txt
-        in runtime
-        """
-        logger.info("generate test.txt from requirements/test.in with local torch whls")
-        preprocess_test_in()
-        copy("requirements/test.txt", "snapshot_constraint.txt")
-
-        run_command(
-            f"{sys.executable} -m uv pip compile requirements/test.in "
-            "-o test.txt "
-            "--index-strategy unsafe-best-match "
-            "--constraint snapshot_constraint.txt "
-            "--torch-backend cu128"
-        )
-        pip_install_packages(requirements="test.txt", prefer_uv=True)
-        logger.info("Done. installed requirements for test dependencies")
-
-    def _install_dependencies(self):
-        pip_install_packages(packages=["-e", "tests/vllm_test_utils"], prefer_uv=True)
-        pip_install_packages(packages=["hf_transfer"], prefer_uv=True)
-        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
-
-        # using script from vllm repo to remove all torch packages from requirements txt
-        run_python("use_existing_torch.py")
-
-        # install common packages
-        for requirements in ["requirements/common.txt", "requirements/build.txt"]:
-            pip_install_packages(
-                requirements=requirements,
-                prefer_uv=True,
-            )
-        # install test packages
-        self._install_test_dependencies()
-
-    def _set_envs(self, inputs: VllmTestParameters):
-        os.environ["TORCH_CUDA_ARCH_LIST"] = inputs.torch_cuda_arch_list
-        if not validate_cuda(get_env("TORCH_CUDA_ARCH_LIST")):
-            logger.warning(
-                "Missing supported TORCH_CUDA_ARCH_LIST. "
-                "Currently support TORCH_CUDA_ARCH_LIST env var "
-                "with supported arch [8.0, 8.9, 9.0]"
-            )
-
-        os.environ["HF_TOKEN"] = os.getenv("VLLM_TEST_HUGGING_FACE_TOKEN", "")
-        if not get_env("HF_TOKEN"):
-            raise ValueError(
-                "missing required HF_TOKEN, please set VLLM_TEST_HUGGING_FACE_TOKEN env var"
-            )
-        if not get_env("TORCH_CUDA_ARCH_LIST"):
-            raise ValueError(
-                "missing required TORCH_CUDA_ARCH_LIST, please set TORCH_CUDA_ARCH_LIST env var"
-            )
-
-
-def preprocess_test_in(
-    target_file: str = "requirements/test.in", additional_packages: Iterable[str] = ()
-):
-    """
-    This modifies the target_file file in place in vllm work directory.
-    It removes torch and unwanted packages in target_file and replace with local torch whls
-    package  with format "$WHEEL_PACKAGE_NAME @ file://<LOCAL_PATH>"
-    """
-    additional_package_to_move = list(additional_packages or ())
-    pkgs_to_remove = [
-        "torch",
-        "torchvision",
-        "torchaudio",
-        "xformers",
-        "mamba_ssm",
-    ] + additional_package_to_move
-    # Read current requirements
-    target_path = Path(target_file)
-    lines = target_path.read_text().splitlines()
-
-    pkgs_to_add = []
-
-    # Remove lines starting with the package names (==, @, >=) — case-insensitive
-    pattern = re.compile(rf"^({'|'.join(pkgs_to_remove)})\s*(==|@|>=)", re.IGNORECASE)
-    kept_lines = [line for line in lines if not pattern.match(line)]
-
-    # Get local installed torch/vision/audio from pip freeze
-    # This is hacky, but it works
-    pip_freeze = subprocess.check_output(["pip", "freeze"], text=True)
-    header_lines = [
-        line
-        for line in pip_freeze.splitlines()
-        if re.match(
-            r"^(torch|torchvision|torchaudio)\s*@\s*file://", line, re.IGNORECASE
-        )
-    ]
-
-    # Write back: header_lines + blank + kept_lines
-    out_lines = header_lines + [""] + kept_lines
-    if pkgs_to_add:
-        out_lines += [""] + pkgs_to_add
-
-    out = "\n".join(out_lines) + "\n"
-    target_path.write_text(out)
-    logger.info("[INFO] Updated %s", target_file)
-
-
-def validate_cuda(value: str) -> bool:
-    VALID_VALUES = {"8.0", "8.9", "9.0"}
-    return all(v in VALID_VALUES for v in value.split())
-
-
-def check_versions():
-    """
-    check installed packages version
-    """
-    logger.info("Double check installed packages")
-    patterns = ["torch", "xformers", "torchvision", "torchaudio", "vllm"]
-    for pkg in patterns:
-        pkg_exists(pkg)
-    logger.info("Done. checked installed packages")

.ci/lumen_cli/cli/run.py

@@ -5,7 +5,6 @@ import logging
 
 from cli.build_cli.register_build import register_build_commands
 from cli.lib.common.logger import setup_logging
-from cli.test_cli.register_test import register_test_commands
 
 
 logger = logging.getLogger(__name__)
@@ -21,7 +20,6 @@ def main():
 
     # registers second-level subcommands
     register_build_commands(subparsers)
-    register_test_commands(subparsers)
 
     # parse args after all options are registered
     args = parser.parse_args()

.ci/lumen_cli/cli/test_cli/register_test.py

@@ -1,62 +0,0 @@
-import argparse
-import logging
-
-from cli.lib.common.cli_helper import register_targets, RichHelp, TargetSpec
-from cli.lib.core.vllm.vllm_test import VllmTestRunner
-
-
-logger = logging.getLogger(__name__)
-
-# Maps targets to their argparse configuration and runner
-# it adds new target to path python -m cli.run build external {target} with buildrunner
-_TARGETS: dict[str, TargetSpec] = {
-    "vllm": {
-        "runner": VllmTestRunner,
-        "help": "test vLLM with pytorch main",
-    }
-    # add yours ...
-}
-
-
-def common_args(parser: argparse.ArgumentParser) -> None:
-    """
-    Add common CLI arguments to the given parser.
-    """
-    parser.add_argument(
-        "--shard-id",
-        type=int,
-        default=1,
-        help="a shard id to run, e.g. '0,1,2,3'",
-    )
-    parser.add_argument(
-        "--num-shards",
-        type=int,
-        default=1,
-        help="a number of shards to run, e.g. '4'",
-    )
-    group = parser.add_mutually_exclusive_group(required=True)
-    group.add_argument(
-        "-tp",
-        "--test-plan",
-        type=str,
-        help="a pre-defined test plan to run, e.g. 'basic_correctness_test'",
-    )
-
-
-def register_test_commands(subparsers: argparse._SubParsersAction) -> None:
-    build_parser = subparsers.add_parser(
-        "test",
-        help="test related commands",
-        formatter_class=RichHelp,
-    )
-    build_subparsers = build_parser.add_subparsers(dest="test_command", required=True)
-    overview = "\n".join(
-        f"  {name:12} {spec.get('help', '')}" for name, spec in _TARGETS.items()
-    )
-    external_parser = build_subparsers.add_parser(
-        "external",
-        help="Test external targets",
-        description="Test third-party targets.\n\nAvailable targets:\n" + overview,
-        formatter_class=RichHelp,
-    )
-    register_targets(external_parser, _TARGETS, common_args=common_args)

.ci/lumen_cli/pyproject.toml

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

.ci/lumen_cli/tests/test_run_plan.py

@@ -1,185 +0,0 @@
-# tests/test_run_test_plan.py
-import importlib
-from contextlib import nullcontext
-from types import SimpleNamespace
-from unittest.mock import MagicMock
-
-import pytest
-
-
-MOD = "cli.lib.core.vllm.lib"
-
-# We import inside tests so the MOD override above applies everywhere
-run_test_plan_import_path = f"{MOD}.run_test_plan"
-
-
-def _get_cmd(c):
-    # Support both kwargs and positional args
-    return c.kwargs.get("cmd", c.args[0] if c.args else None)
-
-
-def _get_check(c):
-    if "check" in c.kwargs:
-        return c.kwargs["check"]
-    # If positional, assume second arg is 'check' when present; default False
-    return c.args[1] if len(c.args) > 1 else False
-
-
-@pytest.fixture
-def patch_module(monkeypatch):
-    """
-    Patch helpers ('pip_install_packages', 'temp_environ', 'working_directory',
-    'run_command', 'logger') inside the target module and expose them.
-    """
-    module = importlib.import_module(MOD)
-
-    # Create fakes/mocks
-    pip_install_packages = MagicMock(name="pip_install_packages")
-    run_command = MagicMock(name="run_command", return_value=0)
-
-    # temp_environ / working_directory: record calls but act as context managers
-    temp_calls: list[dict] = []
-    workdir_calls: list[str] = []
-
-    def fake_working_directory(path: str):
-        workdir_calls.append(path)
-        return nullcontext()
-
-    def fake_temp_env(map: dict[str, str]):
-        temp_calls.append(map)
-        return nullcontext()
-
-    logger = SimpleNamespace(
-        info=MagicMock(name="logger.info"),
-        error=MagicMock(name="logger.error"),
-    )
-
-    # Apply patches (raise if attribute doesn't exist)
-    monkeypatch.setattr(
-        module, "pip_install_packages", pip_install_packages, raising=True
-    )
-    monkeypatch.setattr(module, "run_command", run_command, raising=True)
-    monkeypatch.setattr(
-        module, "working_directory", fake_working_directory, raising=True
-    )
-    monkeypatch.setattr(module, "temp_environ", fake_temp_env, raising=True)
-    monkeypatch.setattr(module, "logger", logger, raising=True)
-
-    return SimpleNamespace(
-        module=module,
-        run_test_plan=module.run_test_plan,  # expose to avoid getattr("constant") (Ruff B009)
-        pip_install_packages=pip_install_packages,
-        run_command=run_command,
-        temp_calls=temp_calls,
-        workdir_calls=workdir_calls,
-        logger=logger,
-    )
-
-
-def test_success_runs_all_steps_and_uses_env_and_workdir(monkeypatch, patch_module):
-    run_test_plan = patch_module.run_test_plan
-
-    tests_map = {
-        "basic": {
-            "title": "Basic suite",
-            "package_install": [],
-            "working_directory": "tests",
-            "env_vars": {"GLOBAL_FLAG": "1"},
-            "steps": [
-                "export A=x && pytest -q",
-                "export B=y && pytest -q tests/unit",
-            ],
-        }
-    }
-
-    # One exit code per step (export + two pytest)
-    patch_module.run_command.side_effect = [0, 0, 0]
-
-    run_test_plan("basic", "cpu", tests_map)
-
-    calls = patch_module.run_command.call_args_list
-    cmds = [_get_cmd(c) for c in calls]
-    checks = [_get_check(c) for c in calls]
-
-    assert cmds == [
-        "export A=x && pytest -q",
-        "export B=y && pytest -q tests/unit",
-    ]
-    assert all(chk is False for chk in checks)
-
-    assert patch_module.workdir_calls == ["tests"]
-    assert patch_module.temp_calls == [{"GLOBAL_FLAG": "1"}]
-
-
-def test_installs_packages_when_present(monkeypatch, patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-
-    tests_map = {
-        "with_pkgs": {
-            "title": "Needs deps",
-            "package_install": ["timm==1.0.0", "flash-attn"],
-            "steps": ["pytest -q"],
-        }
-    }
-
-    patch_module.run_command.return_value = 0
-
-    run_test_plan("with_pkgs", "gpu", tests_map)
-
-    patch_module.pip_install_packages.assert_called_once_with(
-        packages=["timm==1.0.0", "flash-attn"],
-        prefer_uv=True,
-    )
-
-
-def test_raises_on_missing_plan(patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-    with pytest.raises(RuntimeError) as ei:
-        run_test_plan("nope", "cpu", tests_map={})
-
-    assert "test nope not found" in str(ei.value)
-
-
-def test_aggregates_failures_and_raises(monkeypatch, patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-
-    tests_map = {
-        "mix": {
-            "title": "Some pass some fail",
-            "steps": [
-                "pytest test_a.py",  # 0 → pass
-                "pytest test_b.py",  # 1 → fail
-                "pytest test_c.py",  # 2 → fail
-            ],
-        }
-    }
-
-    # Simulate pass, fail, fail
-    patch_module.run_command.side_effect = [0, 1, 2]
-
-    with pytest.raises(RuntimeError) as ei:
-        run_test_plan("mix", "cpu", tests_map)
-
-    msg = str(ei.value)
-    assert "2 pytest runs failed" in msg
-    # Ensure logger captured failed tests list
-    patch_module.logger.error.assert_called_once()
-    # And we attempted all three commands
-    assert patch_module.run_command.call_count == 3
-
-
-def test_custom_working_directory_used(patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-
-    tests_map = {
-        "customwd": {
-            "title": "Custom wd",
-            "working_directory": "examples/ci",
-            "steps": ["pytest -q"],
-        }
-    }
-
-    patch_module.run_command.return_value = 0
-    run_test_plan("customwd", "cpu", tests_map)
-
-    assert patch_module.workdir_calls == ["examples/ci"]

.ci/lumen_cli/tests/test_utils.py

@@ -1,143 +0,0 @@
-import os
-import tempfile
-import unittest
-from pathlib import Path
-
-from cli.lib.common.utils import temp_environ, working_directory  # <-- replace import
-
-
-class EnvIsolatedTestCase(unittest.TestCase):
-    """Base class that snapshots os.environ and CWD for isolation."""
-
-    def setUp(self):
-        import os
-        import tempfile
-
-        self._env_backup = dict(os.environ)
-
-        # Snapshot/repair CWD if it's gone
-        try:
-            self._cwd_backup = os.getcwd()
-        except FileNotFoundError:
-            # If CWD no longer exists, switch to a safe place and record that
-            self._cwd_backup = tempfile.gettempdir()
-            os.chdir(self._cwd_backup)
-
-        # Create a temporary directory for the test to run in
-        self._temp_dir = tempfile.mkdtemp()
-        os.chdir(self._temp_dir)
-
-    def tearDown(self):
-        import os
-        import shutil
-        import tempfile
-
-        # Restore cwd first (before cleaning up temp dir)
-        try:
-            os.chdir(self._cwd_backup)
-        except OSError:
-            os.chdir(tempfile.gettempdir())
-
-        # Clean up temporary directory
-        try:
-            shutil.rmtree(self._temp_dir, ignore_errors=True)
-        except Exception:
-            pass  # Ignore cleanup errors
-
-        # Restore env
-        to_del = set(os.environ.keys()) - set(self._env_backup.keys())
-        for k in to_del:
-            os.environ.pop(k, None)
-        for k, v in self._env_backup.items():
-            os.environ[k] = v
-
-
-class TestTempEnviron(EnvIsolatedTestCase):
-    def test_sets_and_restores_new_var(self):
-        var = "TEST_TMP_ENV_NEW"
-        self.assertNotIn(var, os.environ)
-
-        with temp_environ({var: "123"}):
-            self.assertEqual(os.environ[var], "123")
-
-        self.assertNotIn(var, os.environ)  # removed after exit
-
-    def test_overwrites_and_restores_existing_var(self):
-        var = "TEST_TMP_ENV_OVERWRITE"
-        os.environ[var] = "orig"
-
-        with temp_environ({var: "override"}):
-            self.assertEqual(os.environ[var], "override")
-
-        self.assertEqual(os.environ[var], "orig")  # restored
-
-    def test_multiple_vars_and_missing_cleanup(self):
-        v1, v2 = "TEST_ENV_V1", "TEST_ENV_V2"
-        os.environ.pop(v1, None)
-        os.environ[v2] = "keep"
-
-        with temp_environ({v1: "a", v2: "b"}):
-            self.assertEqual(os.environ[v1], "a")
-            self.assertEqual(os.environ[v2], "b")
-
-        self.assertNotIn(v1, os.environ)  # newly-added -> removed
-        self.assertEqual(os.environ[v2], "keep")  # pre-existing -> restored
-
-    def test_restores_even_on_exception(self):
-        var = "TEST_TMP_ENV_EXCEPTION"
-        self.assertNotIn(var, os.environ)
-
-        with self.assertRaises(RuntimeError):
-            with temp_environ({var: "x"}):
-                self.assertEqual(os.environ[var], "x")
-                raise RuntimeError("boom")
-
-        self.assertNotIn(var, os.environ)  # removed after exception
-
-
-class TestWorkingDirectory(EnvIsolatedTestCase):
-    def test_changes_and_restores(self):
-        start = Path.cwd()
-        with tempfile.TemporaryDirectory() as td:
-            target = Path(td) / "wd"
-            target.mkdir()
-
-            with working_directory(str(target)):
-                self.assertEqual(Path.cwd().resolve(), target.resolve())
-
-        self.assertEqual(Path.cwd(), start)
-
-    def test_noop_when_empty_path(self):
-        start = Path.cwd()
-        with working_directory(""):
-            self.assertEqual(Path.cwd(), start)
-        self.assertEqual(Path.cwd(), start)
-
-    def test_restores_on_exception(self):
-        start = Path.cwd()
-
-        with tempfile.TemporaryDirectory() as td:
-            target = Path(td) / "wd_exc"
-            target.mkdir()
-
-            with self.assertRaises(ValueError):
-                with working_directory(str(target)):
-                    # Normalize both sides to handle /var -> /private/var
-                    self.assertEqual(Path.cwd().resolve(), target.resolve())
-                    raise ValueError("boom")
-
-        self.assertEqual(Path.cwd().resolve(), start.resolve())
-
-    def test_raises_for_missing_dir(self):
-        start = Path.cwd()
-        with tempfile.TemporaryDirectory() as td:
-            missing = Path(td) / "does_not_exist"
-            with self.assertRaises(FileNotFoundError):
-                # os.chdir should raise before yielding
-                with working_directory(str(missing)):
-                    pass
-        self.assertEqual(Path.cwd(), start)
-
-
-if __name__ == "__main__":
-    unittest.main(verbosity=2)

.ci/lumen_cli/tests/test_vllm.py

@@ -4,15 +4,12 @@ import unittest
 from pathlib import Path
 from unittest.mock import MagicMock, patch
 
-import cli.lib.core.vllm.vllm_build as vllm_build
-
-
-_VLLM_BUILD_MODULE = "cli.lib.core.vllm.vllm_build"
+import cli.lib.core.vllm as vllm
 
 
 class TestVllmBuildParameters(unittest.TestCase):
-    @patch(f"{_VLLM_BUILD_MODULE}.local_image_exists", return_value=True)
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=True)
+    @patch("cli.lib.core.vllm.local_image_exists", return_value=True)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=True)
     @patch(
         "cli.lib.common.envs_helper.env_path_optional",
         side_effect=lambda name, default=None, resolve=True: {
@@ -37,13 +34,13 @@ class TestVllmBuildParameters(unittest.TestCase):
     def test_params_success_normalizes_and_validates(
         self, mock_env_path, mock_is_path, mock_local_img
     ):
-        params = vllm_build.VllmBuildParameters()
+        params = vllm.VllmBuildParameters()
         self.assertEqual(params.torch_whls_path, Path("/abs/dist"))
         self.assertEqual(params.dockerfile_path, Path("/abs/vllm/Dockerfile"))
         self.assertEqual(params.output_dir, Path("/abs/shared"))
         self.assertEqual(params.base_image, "my/image:tag")
 
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=False)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=False)
     @patch.dict(
         os.environ, {"USE_TORCH_WHEEL": "1", "TORCH_WHEELS_PATH": "dist"}, clear=True
     )
@@ -51,14 +48,14 @@ class TestVllmBuildParameters(unittest.TestCase):
         with tempfile.TemporaryDirectory() as td:
             os.chdir(td)
             with self.assertRaises(ValueError) as cm:
-                vllm_build.VllmBuildParameters(
+                vllm.VllmBuildParameters(
                     use_local_base_image=False,
                     use_local_dockerfile=False,
                 )
         err = cm.exception
         self.assertIn("TORCH_WHEELS_PATH", str(err))
 
-    @patch(f"{_VLLM_BUILD_MODULE}.local_image_exists", return_value=False)
+    @patch("cli.lib.core.vllm.local_image_exists", return_value=False)
     @patch.dict(
         os.environ, {"USE_LOCAL_BASE_IMAGE": "1", "BASE_IMAGE": "img:tag"}, clear=True
     )
@@ -66,14 +63,14 @@ class TestVllmBuildParameters(unittest.TestCase):
         with tempfile.TemporaryDirectory() as td:
             os.chdir(td)
             with self.assertRaises(ValueError) as cm:
-                vllm_build.VllmBuildParameters(
+                vllm.VllmBuildParameters(
                     use_torch_whl=False,
                     use_local_dockerfile=False,
                 )
         err = cm.exception
         self.assertIn("BASE_IMAGE", str(err))
 
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=False)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=False)
     @patch.dict(
         os.environ,
         {"USE_LOCAL_DOCKERFILE": "1", "DOCKERFILE_PATH": "Dockerfile"},
@@ -83,14 +80,14 @@ class TestVllmBuildParameters(unittest.TestCase):
         with tempfile.TemporaryDirectory() as td:
             os.chdir(td)
             with self.assertRaises(ValueError) as cm:
-                vllm_build.VllmBuildParameters(
+                vllm.VllmBuildParameters(
                     use_torch_whl=False,
                     use_local_base_image=False,
                 )
         err = cm.exception
         self.assertIn("DOCKERFILE_PATH", str(err))
 
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=False)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=False)
     @patch.dict(
         os.environ,
         {"OUTPUT_DIR": ""},
@@ -98,13 +95,14 @@ class TestVllmBuildParameters(unittest.TestCase):
     )
     def test_params_missing_output_dir(self, _is_path):
         with self.assertRaises(FileNotFoundError):
-            vllm_build.VllmBuildParameters()
+            vllm.VllmBuildParameters()
 
 
 class TestBuildCmdAndRun(unittest.TestCase):
-    @patch(f"{_VLLM_BUILD_MODULE}.local_image_exists", return_value=True)
+    @patch("cli.lib.core.vllm.local_image_exists", return_value=True)
     def test_generate_docker_build_cmd_includes_bits(self, _exists):
-        runner = vllm_build.VllmBuildRunner()
+        runner = vllm.VllmBuildRunner()
+        # Craft inputs that simulate a prepared build
         inputs = MagicMock()
         inputs.output_dir = Path("/abs/out")
         inputs.use_local_base_image = True
@@ -120,7 +118,7 @@ class TestBuildCmdAndRun(unittest.TestCase):
         inputs.tag_name = "vllm-wheels"
 
         cmd = runner._generate_docker_build_cmd(inputs)
-        squashed = " ".join(cmd.split())
+        squashed = " ".join(cmd.split())  # normalize whitespace for matching
 
         self.assertIn("--output type=local,dest=/abs/out", squashed)
         self.assertIn("-f docker/Dockerfile.nightly_torch", squashed)
@@ -138,17 +136,18 @@ class TestBuildCmdAndRun(unittest.TestCase):
         self.assertIn("--target export-wheels", squashed)
         self.assertIn("-t vllm-wheels", squashed)
 
-    @patch(f"{_VLLM_BUILD_MODULE}.run_command")
-    @patch(f"{_VLLM_BUILD_MODULE}.ensure_dir_exists")
-    @patch(f"{_VLLM_BUILD_MODULE}.clone_vllm")
+    @patch("cli.lib.core.vllm.run_command")
+    @patch("cli.lib.core.vllm.ensure_dir_exists")
+    @patch("cli.lib.core.vllm.clone_vllm")
     @patch.object(
-        vllm_build.VllmBuildRunner,
+        vllm.VllmBuildRunner,
         "_generate_docker_build_cmd",
         return_value="docker buildx ...",
     )
     @patch.dict(
         os.environ,
         {
+            # Make __post_init__ validations pass cheaply
             "USE_TORCH_WHEEL": "0",
             "USE_LOCAL_BASE_IMAGE": "0",
             "USE_LOCAL_DOCKERFILE": "0",
@@ -159,18 +158,24 @@ class TestBuildCmdAndRun(unittest.TestCase):
     def test_run_calls_clone_prepare_and_build(
         self, mock_gen, mock_clone, mock_ensure, mock_run
     ):
+        # Stub parameters instance so we avoid FS/Docker accesses in run()
         params = MagicMock()
         params.output_dir = Path("shared")
         params.use_local_dockerfile = False
         params.use_torch_whl = False
 
-        with patch(f"{_VLLM_BUILD_MODULE}.VllmBuildParameters", return_value=params):
-            runner = vllm_build.VllmBuildRunner()
+        with patch("cli.lib.core.vllm.VllmBuildParameters", return_value=params):
+            runner = vllm.VllmBuildRunner()
             runner.run()
 
         mock_clone.assert_called_once()
         mock_ensure.assert_called_once_with(Path("shared"))
         mock_gen.assert_called_once_with(params)
         mock_run.assert_called_once()
+        # ensure we run in vllm workdir
         _, kwargs = mock_run.call_args
         assert kwargs.get("cwd") == "vllm"
+
+
+if __name__ == "__main__":
+    unittest.main()

.ci/magma/Makefile

@@ -16,7 +16,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma/build_magma.sh
 
 .PHONY: all
-all: magma-cuda130
 all: magma-cuda129
 all: magma-cuda128
 all: magma-cuda126
@@ -26,12 +25,6 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
-.PHONY: magma-cuda130
-magma-cuda130: DESIRED_CUDA := 13.0
-magma-cuda130: CUDA_ARCH_LIST := -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
-magma-cuda130:
-	$(DOCKER_RUN)
-
 .PHONY: magma-cuda129
 magma-cuda129: DESIRED_CUDA := 12.9
 magma-cuda129: CUDA_ARCH_LIST += -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120

.ci/magma/build_magma.sh

@@ -28,7 +28,6 @@ pushd ${PACKAGE_DIR}/magma-${MAGMA_VERSION}
 patch < ${PACKAGE_FILES}/CMake.patch
 patch < ${PACKAGE_FILES}/cmakelists.patch
 patch -p0 < ${PACKAGE_FILES}/thread_queue.patch
-patch -p1 < ${PACKAGE_FILES}/cuda13.patch
 patch -p1 < ${PACKAGE_FILES}/getrf_shfl.patch
 patch -p1 < ${PACKAGE_FILES}/getrf_nbparam.patch
 # The build.sh script expects to be executed from the sources root folder
@@ -38,7 +37,6 @@ popd
 # Package recipe, license and tarball
 # Folder and package name are backward compatible for the build workflow
 cp ${PACKAGE_FILES}/build.sh ${PACKAGE_RECIPE}/build.sh
-cp ${PACKAGE_FILES}/cuda13.patch ${PACKAGE_RECIPE}/cuda13.patch
 cp ${PACKAGE_FILES}/thread_queue.patch ${PACKAGE_RECIPE}/thread_queue.patch
 cp ${PACKAGE_FILES}/cmakelists.patch ${PACKAGE_RECIPE}/cmakelists.patch
 cp ${PACKAGE_FILES}/getrf_shfl.patch ${PACKAGE_RECIPE}/getrf_shfl.patch

.ci/magma/package_files/cuda13.patch

@@ -1,26 +0,0 @@
-diff --git a/interface_cuda/interface.cpp b/interface_cuda/interface.cpp
-index 73fed1b20..e77519bfe 100644
---- a/interface_cuda/interface.cpp
-+++ b/interface_cuda/interface.cpp
-@@ -438,14 +438,20 @@ magma_print_environment()
-         cudaDeviceProp prop;
-         err = cudaGetDeviceProperties( &prop, dev );
-         check_error( err );
-+        #ifdef MAGMA_HAVE_CUDA
-+#if CUDA_VERSION < 13000
-         printf( "%% device %d: %s, %.1f MHz clock, %.1f MiB memory, capability %d.%d\n",
-                 dev,
-                 prop.name,
-                 prop.clockRate / 1000.,
-+#else
-+        printf( "%% device %d: %s, ??? MHz clock, %.1f MiB memory, capability %d.%d\n",
-+                dev,
-+                prop.name,
-+#endif
-                 prop.totalGlobalMem / (1024.*1024.),
-                 prop.major,
-                 prop.minor );
--        #ifdef MAGMA_HAVE_CUDA
-         int arch = prop.major*100 + prop.minor*10;
-         if ( arch < MAGMA_CUDA_ARCH_MIN ) {
-             printf("\n"

.ci/manywheel/build_cuda.sh

@@ -66,9 +66,6 @@ case ${CUDA_VERSION} in
             TORCH_CUDA_ARCH_LIST="7.5;8.0;9.0;10.0;12.0+PTX"
         fi
         ;;
-    13.0)
-        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;12.0+PTX"
-        ;;
     12.6)
         TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6;9.0"
         ;;
@@ -113,15 +110,11 @@ DEPS_SONAME=(
 )
 
 
-# CUDA_VERSION 12.*, 13.*
-if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
+# CUDA_VERSION 12.6, 12.8, 12.9
+if [[ $CUDA_VERSION == 12* ]]; then
     export USE_STATIC_CUDNN=0
     # Try parallelizing nvcc as well
-    TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
-    # Compress the fatbin with -compress-mode=size for CUDA 13
-    if [[ $CUDA_VERSION == 13* ]]; then
-        export TORCH_NVCC_FLAGS="$TORCH_NVCC_FLAGS -compress-mode=size"
-    fi
+    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
     if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
         echo "Bundling with cudnn and cublas."
         DEPS_LIST+=(
@@ -141,7 +134,7 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             "/usr/local/cuda/lib64/libnvrtc-builtins.so"
             "/usr/local/cuda/lib64/libcufile.so.0"
             "/usr/local/cuda/lib64/libcufile_rdma.so.1"
-            "/usr/local/cuda/lib64/libnvshmem_host.so.3"
+            "/usr/local/cuda/lib64/libnvshem_host.so.3"
             "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12"
             "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
@@ -174,29 +167,22 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
     else
         echo "Using nvidia libs from pypi."
         CUDA_RPATHS=(
+            '$ORIGIN/../../nvidia/cublas/lib'
+            '$ORIGIN/../../nvidia/cuda_cupti/lib'
+            '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
+            '$ORIGIN/../../nvidia/cuda_runtime/lib'
             '$ORIGIN/../../nvidia/cudnn/lib'
-            '$ORIGIN/../../nvidia/nvshmem/lib'
-            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/cufft/lib'
+            '$ORIGIN/../../nvidia/curand/lib'
+            '$ORIGIN/../../nvidia/cusolver/lib'
+            '$ORIGIN/../../nvidia/cusparse/lib'
             '$ORIGIN/../../nvidia/cusparselt/lib'
+            '$ORIGIN/../../cusparselt/lib'
+            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/nvshmem/lib'
+            '$ORIGIN/../../nvidia/nvtx/lib'
+            '$ORIGIN/../../nvidia/cufile/lib'
         )
-        if [[ $CUDA_VERSION == 13* ]]; then
-            CUDA_RPATHS+=('$ORIGIN/../../nvidia/cu13/lib')
-        else
-            CUDA_RPATHS+=(
-                '$ORIGIN/../../nvidia/cublas/lib'
-                '$ORIGIN/../../nvidia/cuda_cupti/lib'
-                '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
-                '$ORIGIN/../../nvidia/cuda_runtime/lib'
-                '$ORIGIN/../../nvidia/cufft/lib'
-                '$ORIGIN/../../nvidia/curand/lib'
-                '$ORIGIN/../../nvidia/cusolver/lib'
-                '$ORIGIN/../../nvidia/cusparse/lib'
-                '$ORIGIN/../../cusparselt/lib'
-                '$ORIGIN/../../nvidia/nvtx/lib'
-                '$ORIGIN/../../nvidia/cufile/lib'
-            )
-        fi
-
         CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
         export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
         export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'

.ci/manywheel/build_xpu.sh

@@ -25,7 +25,6 @@ source /opt/intel/oneapi/mpi/latest/env/vars.sh
 export USE_STATIC_MKL=1
 export USE_ONEMKL=1
 export USE_XCCL=1
-export USE_MPI=0
 
 WHEELHOUSE_DIR="wheelhousexpu"
 LIBTORCH_HOUSE_DIR="libtorch_housexpu"

.ci/pytorch/build.sh

@@ -173,7 +173,6 @@ if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   source /opt/intel/oneapi/mpi/latest/env/vars.sh
   # Enable XCCL build
   export USE_XCCL=1
-  export USE_MPI=0
   # XPU kineto feature dependencies are not fully ready, disable kineto build as temp WA
   export USE_KINETO=0
   export TORCH_XPU_ARCH_LIST=pvc
@@ -195,16 +194,8 @@ fi
 
 # We only build FlashAttention files for CUDA 8.0+, and they require large amounts of
 # memory to build and will OOM
-
 if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && echo "${TORCH_CUDA_ARCH_LIST}" | tr ' ' '\n' | sed 's/$/>= 8.0/' | bc | grep -q 1; then
-  J=2  # default to 2 jobs
-  case "$RUNNER" in
-    linux.12xlarge.memory|linux.24xlarge.memory)
-      J=24
-      ;;
-  esac
-  echo "Building FlashAttention with job limit $J"
-  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j ${J}"
+  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *clang* ]]; then

.ci/pytorch/check_binary.sh

@@ -300,3 +300,24 @@ except RuntimeError as e:
     exit 1
   fi
 fi
+
+###############################################################################
+# Check for C++ ABI compatibility to GCC-11 - GCC 13
+###############################################################################
+if [[ "$(uname)" == 'Linux' &&  "$PACKAGE_TYPE" == 'manywheel' ]]; then
+  pushd /tmp
+  # Per https://gcc.gnu.org/onlinedocs/gcc/C_002b_002b-Dialect-Options.html
+  # gcc-11 is ABI16, gcc-13 is ABI18, gcc-14 is ABI19
+  # gcc 11 - CUDA 11.8, xpu, rocm
+  # gcc 13 - CUDA 12.6, 12.8 and cpu
+  # Please see issue for reference: https://github.com/pytorch/pytorch/issues/152426
+  if [[ "$(uname -m)" == "s390x" ]]; then
+    cxx_abi="19"
+  elif [[ "$DESIRED_CUDA" != 'xpu' && "$DESIRED_CUDA" != 'rocm'* ]]; then
+    cxx_abi="18"
+  else
+    cxx_abi="16"
+  fi
+  python -c "import torch; exit(0 if torch._C._PYBIND11_BUILD_ABI == '_cxxabi10${cxx_abi}' else 1)"
+  popd
+fi

.ci/pytorch/common_utils.sh

@@ -149,19 +149,6 @@ function get_pinned_commit() {
   cat .github/ci_commit_pins/"${1}".txt
 }
 
-function detect_cuda_arch() {
-  if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]]; then
-    if command -v nvidia-smi; then
-      TORCH_CUDA_ARCH_LIST=$(nvidia-smi --query-gpu=compute_cap --format=csv | tail -n 1)
-    elif [[ "${TEST_CONFIG}" == *nogpu* ]]; then
-      # There won't be nvidia-smi in nogpu tests, so just set TORCH_CUDA_ARCH_LIST to the default
-      # minimum supported value here
-      TORCH_CUDA_ARCH_LIST=8.0
-    fi
-    export TORCH_CUDA_ARCH_LIST
-  fi
-}
-
 function install_torchaudio() {
   local commit
   commit=$(get_pinned_commit audio)

.ci/pytorch/macos-test.sh

@@ -174,15 +174,13 @@ checkout_install_torchbench() {
     # to install and test other models
     python install.py --continue_on_fail
   fi
-  popd
 
-  pip install -r .ci/docker/ci_commit_pins/huggingface-requirements.txt
-  # https://github.com/pytorch/pytorch/issues/160689 to remove torchao because
-  # its current version 0.12.0 doesn't work with transformers 4.54.0
-  pip uninstall -y torchao
+  # soxr comes from https://github.com/huggingface/transformers/pull/39429
+  pip install transformers==4.54.0 soxr==0.5.0
 
   echo "Print all dependencies after TorchBench is installed"
   python -mpip freeze
+  popd
 }
 
 torchbench_setup_macos() {

.ci/pytorch/multigpu-test.sh

@@ -45,7 +45,6 @@ if [[ "${SHARD_NUMBER:-2}" == "2" ]]; then
     # DTensor tests
     time python test/run_test.py --verbose -i distributed/tensor/test_random_ops
     time python test/run_test.py --verbose -i distributed/tensor/test_dtensor_compile
-    time python test/run_test.py --verbose -i distributed/tensor/test_utils.py
 
     # DeviceMesh test
     time python test/run_test.py --verbose -i distributed/test_device_mesh

.ci/pytorch/test.sh

@@ -91,7 +91,6 @@ if [[ "$BUILD_ENVIRONMENT" == *clang9* || "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   export VALGRIND=OFF
 fi
 
-detect_cuda_arch
 
 if [[ "$BUILD_ENVIRONMENT" == *s390x* ]]; then
   # There are additional warnings on s390x, maybe due to newer gcc.
@@ -1630,10 +1629,6 @@ elif [[ "${TEST_CONFIG}" == *xla* ]]; then
   install_torchvision
   build_xla
   test_xla
-elif [[ "$TEST_CONFIG" == *vllm* ]]; then
-    echo "vLLM CI uses TORCH_CUDA_ARCH_LIST: $TORCH_CUDA_ARCH_LIST"
-    (cd .ci/lumen_cli && python -m pip install -e .)
-    python -m cli.run test external vllm --test-plan "$TEST_CONFIG" --shard-id "$SHARD_NUMBER" --num-shards "$NUM_TEST_SHARDS"
 elif [[ "${TEST_CONFIG}" == *executorch* ]]; then
   test_executorch
 elif [[ "$TEST_CONFIG" == 'jit_legacy' ]]; then
@@ -1706,7 +1701,7 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
-  PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
+  PYTHONPATH=/torchbench:$PYTHONPATH test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
     test_inductor_aoti
   fi

.ci/pytorch/win-test.sh

@@ -44,7 +44,7 @@ python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==
 python -m pip install z3-solver==4.15.1.0
 
 # Install tlparse for test\dynamo\test_structured_trace.py UTs.
-python -m pip install tlparse==0.4.0
+python -m pip install tlparse==0.3.30
 
 # Install parameterized
 python -m pip install parameterized==0.8.1

.ci/pytorch/windows/cuda130.bat

@@ -1,59 +0,0 @@
-@echo off
-
-set MODULE_NAME=pytorch
-
-IF NOT EXIST "setup.py" IF NOT EXIST "%MODULE_NAME%" (
-    call internal\clone.bat
-    cd %~dp0
-) ELSE (
-    call internal\clean.bat
-)
-IF ERRORLEVEL 1 goto :eof
-
-call internal\check_deps.bat
-IF ERRORLEVEL 1 goto :eof
-
-REM Check for optional components
-
-set USE_CUDA=
-set CMAKE_GENERATOR=Visual Studio 15 2017 Win64
-
-IF "%NVTOOLSEXT_PATH%"=="" (
-    IF EXIST "C:\Program Files\NVIDIA Corporation\NvToolsExt\lib\x64\nvToolsExt64_1.lib"  (
-        set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
-    ) ELSE (
-        echo NVTX ^(Visual Studio Extension ^for CUDA^) ^not installed, failing
-        exit /b 1
-    )
-)
-
-IF "%CUDA_PATH_V130%"=="" (
-    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0\bin\nvcc.exe" (
-        set "CUDA_PATH_V130=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0"
-    ) ELSE (
-        echo CUDA 13.0 not found, failing
-        exit /b 1
-    )
-)
-
-IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=7.5;8.0;8.6;9.0;10.0;12.0
-    set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
-) ELSE (
-    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90 -gencode=arch=compute_100,code=compute_100 -gencode=arch=compute_120,code=compute_120
-)
-
-set "CUDA_PATH=%CUDA_PATH_V130%"
-set "PATH=%CUDA_PATH_V130%\bin;%PATH%"
-
-:optcheck
-
-call internal\check_opts.bat
-IF ERRORLEVEL 1 goto :eof
-
-if exist "%NIGHTLIES_PYTORCH_ROOT%" cd %NIGHTLIES_PYTORCH_ROOT%\..
-call  %~dp0\internal\copy.bat
-IF ERRORLEVEL 1 goto :eof
-
-call  %~dp0\internal\setup.bat
-IF ERRORLEVEL 1 goto :eof

.ci/pytorch/windows/internal/cuda_install.bat

@@ -26,7 +26,6 @@ if exist "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v%CUDA_VERSION_STR%
 if %CUDA_VER% EQU 126 goto cuda126
 if %CUDA_VER% EQU 128 goto cuda128
 if %CUDA_VER% EQU 129 goto cuda129
-if %CUDA_VER% EQU 130 goto cuda130
 
 echo CUDA %CUDA_VERSION_STR% is not supported
 exit /b 1
@@ -114,33 +113,6 @@ xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
 
 goto cuda_common
 
-:cuda130
-
-set CUDA_INSTALL_EXE=cuda_13.0.0_windows.exe
-if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
-    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" & REM @lint-ignore
-    if errorlevel 1 exit /b 1
-    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
-    set "ARGS="
-)
-
-set CUDNN_FOLDER=cudnn-windows-x86_64-9.12.0.46_cuda13-archive
-set CUDNN_LIB_FOLDER="lib"
-set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
-if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
-    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" & REM @lint-ignore
-    if errorlevel 1 exit /b 1
-    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
-)
-
-@REM cuDNN 8.3+ required zlib to be installed on the path
-echo Installing ZLIB dlls
-curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
-7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
-xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
-
-goto cuda_common
-
 :cuda_common
 :: NOTE: We only install CUDA if we don't have it installed already.
 :: With GHA runners these should be pre-installed as part of our AMI process

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

@@ -1,22 +1,12 @@
 set ADDITIONAL_OPTIONS=""
 set PYTHON_EXEC="python"
-
-
 if "%DESIRED_PYTHON%" == "3.13t" (
     echo Python version is set to 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 ADDITIONAL_OPTIONS="Include_freethreaded=1"
-    set PYTHON_EXEC="python3.14t"
 ) else (
-    echo Python version is set to %DESIRED_PYTHON%
+    echo DESIRED_PYTHON not defined, Python version is set to %DESIRED_PYTHON%
     set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/%DESIRED_PYTHON%.0/python-%DESIRED_PYTHON%.0-amd64.exe" %= @lint-ignore =%
 )
 

.ci/pytorch/windows/internal/xpu_install.bat

@@ -13,9 +13,9 @@ if not exist "%SRC_DIR%\temp_build" mkdir "%SRC_DIR%\temp_build"
 :xpu_bundle_install_start
 
 set XPU_BUNDLE_PARENT_DIR=C:\Program Files (x86)\Intel\oneAPI
-set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
+set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d6d6c17-ca2d-4735-9331-99447e4a1280/intel-deep-learning-essentials-2025.0.1.28_offline.exe
 set XPU_BUNDLE_PRODUCT_NAME=intel.oneapi.win.deep-learning-essentials.product
-set XPU_BUNDLE_VERSION=2025.1.3+5
+set XPU_BUNDLE_VERSION=2025.0.1+20
 set XPU_BUNDLE_INSTALLED=0
 set XPU_BUNDLE_UNINSTALL=0
 set XPU_EXTRA_URL=NULL
@@ -24,9 +24,9 @@ set XPU_EXTRA_VERSION=2025.0.1+1226
 set XPU_EXTRA_INSTALLED=0
 set XPU_EXTRA_UNINSTALL=0
 
-if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.2] (
-    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
-    set XPU_BUNDLE_VERSION=2025.2.1+20
+if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.1] (
+    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
+    set XPU_BUNDLE_VERSION=2025.1.3+5
 )
 
 :: Check if XPU bundle is target version or already installed
@@ -90,3 +90,14 @@ if errorlevel 1 exit /b 1
 del xpu_extra.exe
 
 :xpu_install_end
+
+if not "%XPU_ENABLE_KINETO%"=="1" goto install_end
+:: Install Level Zero SDK
+set XPU_EXTRA_LZ_URL=https://github.com/oneapi-src/level-zero/releases/download/v1.14.0/level-zero-sdk_1.14.0.zip
+curl -k -L %XPU_EXTRA_LZ_URL% --output "%SRC_DIR%\temp_build\level_zero_sdk.zip"
+echo "Installing level zero SDK..."
+7z x "%SRC_DIR%\temp_build\level_zero_sdk.zip" -o"%SRC_DIR%\temp_build\level_zero"
+set "INCLUDE=%SRC_DIR%\temp_build\level_zero\include;%INCLUDE%"
+del "%SRC_DIR%\temp_build\level_zero_sdk.zip"
+
+:install_end

.ci/pytorch/windows/setup_build.bat

@@ -7,8 +7,6 @@ call "internal\install_python.bat"
 
 %PYTHON_EXEC% --version
 set "PATH=%CD%\Python\Lib\site-packages\cmake\data\bin;%CD%\Python\Scripts;%CD%\Python;%PATH%"
-if "%DESIRED_PYTHON%" == "3.14t" %PYTHON_EXEC% -m pip install numpy==2.3.2 cmake
-if "%DESIRED_PYTHON%" == "3.14" %PYTHON_EXEC% -m pip install numpy==2.3.2 cmake
 if "%DESIRED_PYTHON%" == "3.13t" %PYTHON_EXEC% -m pip install numpy==2.2.1 cmake
 if "%DESIRED_PYTHON%" == "3.13" %PYTHON_EXEC% -m pip install numpy==2.1.2 cmake
 if "%DESIRED_PYTHON%" == "3.12" %PYTHON_EXEC% -m pip install numpy==2.0.2 cmake

.ci/wheel/build_wheel.sh

@@ -128,35 +128,16 @@ export MACOSX_DEPLOYMENT_TARGET=10.15
 export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
 SETUPTOOLS_PINNED_VERSION="==70.1.0"
-PYYAML_PINNED_VERSION="==5.3"
+PYYAML_PINNED_VERSION="=5.3"
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
 RENAME_WHEEL=true
 case $desired_python in
-    3.14t)
-        echo "Using 3.14 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="==2.1.0"
-        CONDA_ENV_CREATE_FLAGS="python-freethreading"
-        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-        desired_python="3.14.0rc1"
-        RENAME_WHEEL=false
-        ;;
-    3.14)
-        echo "Using 3.14t deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="==2.1.0"
-        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-        desired_python="3.14.0rc1"
-        RENAME_WHEEL=false
-        ;;
     3.13t)
         echo "Using 3.13 deps"
         SETUPTOOLS_PINNED_VERSION=">=70.1.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="==2.1.0"
+        NUMPY_PINNED_VERSION="=2.1.0"
         CONDA_ENV_CREATE_FLAGS="python-freethreading"
         EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
         desired_python="3.13"
@@ -166,35 +147,35 @@ case $desired_python in
         echo "Using 3.13 deps"
         SETUPTOOLS_PINNED_VERSION=">=70.1.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="==2.1.0"
+        NUMPY_PINNED_VERSION="=2.1.0"
         ;;
     3.12)
         echo "Using 3.12 deps"
         SETUPTOOLS_PINNED_VERSION=">=70.1.0"
         PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.11)
         echo "Using 3.11 deps"
         SETUPTOOLS_PINNED_VERSION=">=70.1.0"
         PYYAML_PINNED_VERSION=">=5.3"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.10)
         echo "Using 3.10 deps"
         SETUPTOOLS_PINNED_VERSION=">=70.1.0"
         PYYAML_PINNED_VERSION=">=5.3"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.9)
         echo "Using 3.9 deps"
         SETUPTOOLS_PINNED_VERSION=">=70.1.0"
         PYYAML_PINNED_VERSION=">=5.3"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     *)
         echo "Using default deps"
-        NUMPY_PINNED_VERSION="==1.11.3"
+        NUMPY_PINNED_VERSION="=1.11.3"
         ;;
 esac
 
@@ -203,13 +184,8 @@ tmp_env_name="wheel_py$python_nodot"
 conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
 source activate "$tmp_env_name"
 
-PINNED_PACKAGES=(
-    "setuptools${SETUPTOOLS_PINNED_VERSION}"
-    "pyyaml${PYYAML_PINNED_VERSION}"
-    "numpy${NUMPY_PINNED_VERSION}"
-)
-retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements-build.txt"
-pip install requests ninja typing-extensions
+retry pip install -r "${pytorch_rootdir}/requirements-build.txt"
+pip install "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing-extensions
 retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
 retry brew install libomp
 

.circleci/scripts/binary_upload.sh

@@ -51,12 +51,16 @@ s3_upload() {
     s3_upload_dir="${s3_root_dir}/${UPLOAD_SUBFOLDER}/"
   fi
   (
+    cache_control_flag=""
+    if [[ "${UPLOAD_CHANNEL}" = "test" ]]; then
+      cache_control_flag="--cache-control='no-cache,no-store,must-revalidate'"
+    fi
     for pkg in ${PKG_DIR}/*.${extension}; do
       (
         set -x
         shm_id=$(sha256sum "${pkg}" | awk '{print $1}')
         ${AWS_S3_CP} --no-progress --acl public-read "${pkg}" "${s3_upload_dir}" \
-          --metadata "checksum-sha256=${shm_id}"
+          --metadata "checksum-sha256=${shm_id}" ${cache_control_flag}
       )
     done
   )

.circleci/scripts/binary_windows_build.sh

@@ -15,7 +15,8 @@ fi
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
     export VC_YEAR=2022
     export USE_SCCACHE=0
-    export XPU_VERSION=2025.2
+    export XPU_VERSION=2025.1
+    export XPU_ENABLE_KINETO=1
 fi
 
 echo "Free space on filesystem before build:"

.circleci/scripts/binary_windows_test.sh

@@ -8,7 +8,7 @@ export VC_YEAR=2022
 
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
     export VC_YEAR=2022
-    export XPU_VERSION=2025.2
+    export XPU_VERSION=2025.1
 fi
 
 pushd "$PYTORCH_ROOT/.ci/pytorch/"

.flake8

@@ -48,7 +48,6 @@ per-file-ignores =
     torch/__init__.py: F401,TOR901
     torch/_custom_op/impl.py: TOR901
     torch/_export/serde/upgrade.py: TOR901
-    torch/_functorch/predispatch.py: TOR901
     torch/_functorch/vmap.py: TOR901
     torch/_inductor/test_operators.py: TOR901
     torch/_library/abstract_impl.py: TOR901

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

@@ -1,81 +0,0 @@
-# .github/workflows/build-external.yml
-name: Build External packages
-
-description: build external packages for PyTorch
-
-inputs:
-  cuda-arch-list:
-    description: TORCH_CUDA_ARCH_LIST (e.g., "8.0;8.9;9.0")
-    type: string
-    required: true
-    default: ""
-  docker-image:
-    description: Base image to use
-    type: string
-    required: true
-  build-targets:
-    description: Build targets
-    type: string
-    required: true
-  torch-wheel-dir:
-    description: Directory to built torch wheel
-    type: string
-    required: false
-    default: dist
-  output-dir:
-    description: Directory to store build artifact
-    default: external
-    type: string
-    required: false
-
-outputs:
-  build_time:
-    description: "Total build time in seconds"
-    value: ${{ steps.build-external.outputs.build_time }}
-  output_dir:
-    description: "Directory where build artifact is stored"
-    value: ${{ steps.build-external.outputs.output_dir }}
-
-runs:
-  using: composite
-  steps:
-    - name: Build external packages in sequence
-      id: build-external
-      env:
-        SCCACHE_BUCKET: ossci-compiler-cache-circleci-v2
-        SCCACHE_REGION: us-east-1
-        TORCH_CUDA_ARCH_LIST: ${{ inputs.cuda-arch-list }}
-        BASE_IMAGE: ${{ inputs.docker-image }}
-        BUILD_TARGETS: ${{ inputs.build-targets }}
-        PARENT_OUTPUT_DIR: ${{ inputs.output-dir}}
-
-      shell: bash
-      run: |
-        set -euo pipefail
-        python3 --version
-        docker images
-        START_TIME=$(date +%s)
-        (
-          cd .ci/lumen_cli
-          python3 -m pip install -e .
-        )
-        MAX_JOBS="$(nproc --ignore=6)"
-        export MAX_JOBS
-
-        # Split the comma-separated list and build each target
-        IFS=',' read -ra TARGETS <<< "$BUILD_TARGETS"
-        for target in "${TARGETS[@]}"; do
-          OUTPUT_DIR="$PARENT_OUTPUT_DIR/$target"
-          export OUTPUT_DIR
-          echo "Building external package: $target in directory $OUTPUT_DIR"
-          python3 -m cli.run build external "$target"
-
-        done
-
-        END_TIME=$(date +%s)
-        {
-          echo "build_time=$((END_TIME - START_TIME))"
-          if [ -d "$PARENT_OUTPUT_DIR" ]; then
-            echo "output_dir=$PARENT_OUTPUT_DIR"
-          fi
-        } >> "$GITHUB_OUTPUT"

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-10a5002c6195bd95e34df8fe28ff8a2d55a2a922
+bdb88e1d66f272cad72156c90ac8428ca61a601c

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-321938e9ac4000e0cb37e328359a7fd3026bc672
+0ca2393b47e72c4424a49aa3b32c7c5d0e378a72

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-a1c6ee92c85e8b0955c20892ed68f032a6015c09
+095faec1e7b6cc47220181e74ae9cde2605f9b00

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -67,8 +67,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 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 ####################
 
@@ -92,8 +90,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     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
 
@@ -116,7 +112,6 @@ ARG PINNED_TORCH_VERSION
 RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
-        echo "[INFO] Installing torch wheels to build vllm"; \
         torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
         vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
         audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
@@ -124,10 +119,10 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
         uv pip install --system "${vision_whl}"; \
         uv pip install --system "${audio_whl}"; \
     elif [ -n "$PINNED_TORCH_VERSION" ]; then \
-        echo "[INFO] Installing pinned torch nightly version to build vllm: $PINNED_TORCH_VERSION"; \
+        echo "[INFO] Installing pinned torch nightly version: $PINNED_TORCH_VERSION"; \
         uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu128; \
     else \
-        echo "[INFO] Installing torch nightly with latest one to build vllm"; \
+        echo "[INFO] Installing torch nightly with latest one"; \
         uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128; \
     fi
 
@@ -141,22 +136,15 @@ uv pip install --system -r requirements/common.txt
 
 
 # Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG exformer_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${exformer_cuda_arch_list}
-
+ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
+ENV TORCH_CUDA_ARCH_LIST=${torch_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'
-
 # 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/v0.0.31
-ARG XFORMERS_COMMIT=eb0946a363464da96ea40afd1a7f72a907c25497
+ARG XFORMERS_COMMIT=f2de641ef670510cadab099ce6954031f52f191c
 ENV CCACHE_DIR=/root/.cache/ccache
-
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/uv \
     echo 'git clone xformers...' \
@@ -169,15 +157,14 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
     && python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
     && cd .. \
     && rm -rf xformers
-
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system xformers-dist/*.whl --verbose
 
 # Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
 # track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
 RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
+RUN cat  torch_build_versions.txt
 
-RUN cat torch_build_versions.txt
 RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
 
 #################### BASE BUILD IMAGE ####################
@@ -190,8 +177,6 @@ ARG TARGETPLATFORM
 
 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 . .
 
@@ -207,7 +192,7 @@ RUN --mount=type=bind,source=.git,target=.git \
 # Max jobs used by Ninja to build extensions
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
-ARG nvcc_threads=4
+ARG nvcc_threads=2
 ENV NVCC_THREADS=$nvcc_threads
 ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
@@ -231,14 +216,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
         && export SCCACHE_IDLE_TIMEOUT=0 \
         && export CMAKE_BUILD_TYPE=Release \
-        && export VLLM_DOCKER_BUILD_CONTEXT=1 \
         && sccache --show-stats \
         && python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38 \
         && sccache --show-stats; \
     fi
 
-ARG vllm_target_device="cuda"
-ENV VLLM_TARGET_DEVICE=${vllm_target_device}
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/uv \
@@ -247,7 +229,6 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         # Clean any existing CMake artifacts
         rm -rf .deps && \
         mkdir -p .deps && \
-        export VLLM_DOCKER_BUILD_CONTEXT=1 && \
         python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
     fi
 
@@ -315,8 +296,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     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
 
 # Default mount file as placeholder, this just avoid the mount error
 ARG TORCH_WHEELS_PATH="./requirements"
@@ -329,7 +308,7 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
         torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
         vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
         audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
-        echo "[INFO] Use wheels to build : '${torch_whl}' '${audio_whl}' '${vision_whl}'"; \
+        echo "Found: '${torch_whl}' '${audio_whl}' '${vision_whl}'"; \
         uv pip install --system "${torch_whl}[opt-einsum]"; \
         uv pip install --system "${vision_whl}"; \
         uv pip install --system "${audio_whl}"; \
@@ -359,7 +338,7 @@ ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 # Build flashinfer for torch nightly from source around 10 mins
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
 # Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
-ARG FLASHINFER_GIT_REF="v0.2.14.post1"
+ARG FLASHINFER_GIT_REF="v0.2.9rc2"
 RUN --mount=type=cache,target=/root/.cache/uv \
     git clone --depth 1 --recursive --shallow-submodules \
         --branch ${FLASHINFER_GIT_REF} \
@@ -377,7 +356,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Logging to confirm the torch versions
 RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm\|^flashinfer' > build_summary.txt
 ################### VLLM INSTALLED IMAGE ####################
 
 
@@ -386,8 +364,6 @@ 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
@@ -435,5 +411,4 @@ FROM scratch as export-wheels
 # Just copy the wheels we prepared in previous stages
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
-COPY --from=vllm-base /workspace/build_summary.txt /wheels/build_summary.txt
 COPY --from=vllm-base /workspace/wheels/flashinfer /wheels/flashinfer-python

.github/dependabot.yml

@@ -1,24 +0,0 @@
-version: 2
-updates:
-  # Update to the latest transformers version with dependabot
-  - package-ecosystem: "pip"
-    directory: "/.ci/docker/ci_commit_pins"
-    schedule:
-      interval: "daily"
-    target-branch: "main"
-    allow:
-      - dependency-name: "transformers"
-    ignore:
-      - dependency-name: "*"
-        update-types: ["version-update:semver-patch"]
-    commit-message:
-      prefix: "[Dependabot] Update"
-      include: "scope"
-    labels:
-      - "dependencies"
-      - "open source"
-      - "python"
-      - "topic: not user facing"
-      - "module: ci"
-      - "module: inductor"
-      - "ciflow/inductor"

.github/labeler.yml

@@ -41,9 +41,6 @@
 - test/inductor/**
 - test/dynamo/**
 
-"ciflow/vllm":
-- .github/ci_commit_pins/vllm.txt
-
 "module: cpu":
 - aten/src/ATen/cpu/**
 - aten/src/ATen/native/cpu/**

.github/pytorch-probot.yml

@@ -27,7 +27,6 @@ ciflow_push_tags:
 - ciflow/trunk
 - ciflow/unstable
 - ciflow/xpu
-- ciflow/vllm
 - ciflow/torchbench
 - ciflow/op-benchmark
 - ciflow/pull

.github/requirements/pip-requirements-macOS.txt

@@ -28,7 +28,7 @@ pyyaml==6.0.2
 scipy==1.12.0
 setuptools==72.1.0
 sympy==1.13.3
-tlparse==0.4.0
+tlparse==0.3.30
 tensorboard==2.13.0
 typing-extensions==4.12.2
 unittest-xml-reporting<=3.2.0,>=2.0.0

.github/scripts/generate_binary_build_matrix.py

@@ -16,19 +16,17 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "12.9"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
     "12.9": "12.9.1",
-    "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
     "12.8": "9",
     "12.9": "9",
-    "13.0": "9",
 }
 
 # NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
@@ -40,7 +38,7 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.9-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.9-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
@@ -56,7 +54,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'"
@@ -73,7 +71,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'"
@@ -90,49 +88,32 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
         "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
-    "13.0": (
-        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'"
-    ),
     "xpu": (
-        "intel-cmplr-lib-rt==2025.2.1 | "
-        "intel-cmplr-lib-ur==2025.2.1 | "
-        "intel-cmplr-lic-rt==2025.2.1 | "
-        "intel-sycl-rt==2025.2.1 | "
-        "oneccl-devel==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "oneccl==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "impi-rt==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "onemkl-sycl-blas==2025.2.0 | "
-        "onemkl-sycl-dft==2025.2.0 | "
-        "onemkl-sycl-lapack==2025.2.0 | "
-        "onemkl-sycl-rng==2025.2.0 | "
-        "onemkl-sycl-sparse==2025.2.0 | "
-        "dpcpp-cpp-rt==2025.2.1 | "
-        "intel-opencl-rt==2025.2.1 | "
-        "mkl==2025.2.0 | "
-        "intel-openmp==2025.2.1 | "
-        "tbb==2022.2.0 | "
-        "tcmlib==1.4.0 | "
-        "umf==0.11.0 | "
-        "intel-pti==0.13.1"
+        "intel-cmplr-lib-rt==2025.1.1 | "
+        "intel-cmplr-lib-ur==2025.1.1 | "
+        "intel-cmplr-lic-rt==2025.1.1 | "
+        "intel-sycl-rt==2025.1.1 | "
+        "oneccl-devel==2021.15.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "oneccl==2021.15.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "impi-rt==2021.15.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "onemkl-sycl-blas==2025.1.0 | "
+        "onemkl-sycl-dft==2025.1.0 | "
+        "onemkl-sycl-lapack==2025.1.0 | "
+        "onemkl-sycl-rng==2025.1.0 | "
+        "onemkl-sycl-sparse==2025.1.0 | "
+        "dpcpp-cpp-rt==2025.1.1 | "
+        "intel-opencl-rt==2025.1.1 | "
+        "mkl==2025.1.0 | "
+        "intel-openmp==2025.1.1 | "
+        "tbb==2022.1.0 | "
+        "tcmlib==1.3.0 | "
+        "umf==0.10.0 | "
+        "intel-pti==0.12.3"
     ),
 }
 
@@ -143,7 +124,9 @@ def get_nccl_wheel_version(arch_version: str) -> str:
     requirements = map(
         str.strip, re.split("[;|]", PYTORCH_EXTRA_INSTALL_REQUIREMENTS[arch_version])
     )
-    return next(x for x in requirements if x.startswith("nvidia-nccl")).split("==")[1]
+    return next(x for x in requirements if x.startswith("nvidia-nccl-cu")).split("==")[
+        1
+    ]
 
 
 def read_nccl_pin(arch_version: str) -> str:
@@ -210,7 +193,7 @@ LIBTORCH_CONTAINER_IMAGES: dict[str, str] = {
     "cpu": "libtorch-cxx11-builder:cpu",
 }
 
-FULL_PYTHON_VERSIONS = ["3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
+FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
 
 
 def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
@@ -240,12 +223,8 @@ def generate_libtorch_matrix(
         if os == "linux":
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
-            if "13.0" in arches:
-                arches.remove("13.0")
         elif os == "windows":
             arches += CUDA_ARCHES
-            if "13.0" in arches:
-                arches.remove("13.0")
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -310,8 +289,6 @@ def generate_wheels_matrix(
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
             arches += CUDA_ARCHES + XPU_ARCHES
-            if "13.0" in arches:
-                arches.remove("13.0")
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -337,8 +314,8 @@ def generate_wheels_matrix(
             # TODO: Enable python 3.13t on cpu-s390x
             if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
                 continue
-            # TODO: Enable python 3.14 for rest
-            if os not in ["linux", "linux-aarch64", "macos-arm64", "windows"] and (
+            # TODO: Enable python 3.14 on non linux OSes
+            if os != "linux" and (
                 python_version == "3.14" or python_version == "3.14t"
             ):
                 continue
@@ -346,7 +323,7 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.9", "12.8", "12.6"]
+                arch_version in ["12.9", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -379,6 +356,29 @@ def generate_wheels_matrix(
                         ),  # include special case for aarch64 build, remove the -aarch64 postfix
                     }
                 )
+                # Special build building to use on Colab. Python 3.11 for 12.6 CUDA
+                if python_version == "3.11" and arch_version == CUDA_STABLE:
+                    ret.append(
+                        {
+                            "python_version": python_version,
+                            "gpu_arch_type": gpu_arch_type,
+                            "gpu_arch_version": gpu_arch_version,
+                            "desired_cuda": translate_desired_cuda(
+                                gpu_arch_type, gpu_arch_version
+                            ),
+                            "container_image": WHEEL_CONTAINER_IMAGES[
+                                arch_version
+                            ].split(":")[0],
+                            "container_image_tag_prefix": WHEEL_CONTAINER_IMAGES[
+                                arch_version
+                            ].split(":")[1],
+                            "package_type": package_type,
+                            "pytorch_extra_install_requirements": "",
+                            "build_name": f"{package_type}-py{python_version}-{gpu_arch_type}{gpu_arch_version}-full".replace(  # noqa: B950
+                                ".", "_"
+                            ),
+                        }
+                    )
             else:
                 ret.append(
                     {
@@ -409,7 +409,6 @@ def generate_wheels_matrix(
     return ret
 
 
-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/test_trymerge.py

@@ -124,7 +124,7 @@ def mock_parse_args(revert: bool = False, force: bool = False) -> Any:
             self.force = force
             self.pr_num = 76123
             self.dry_run = True
-            self.comment_id = 12345  # Set to non-zero value
+            self.comment_id = 0
             self.reason = "this is for testing"
             self.ignore_current = False
             self.check_mergeability = False
@@ -152,9 +152,9 @@ def mock_revert(
 def mock_merge(
     pr: GitHubPR,
     repo: GitRepo,
-    comment_id: int,
     dry_run: bool = False,
     skip_mandatory_checks: bool = False,
+    comment_id: Optional[int] = None,
     timeout_minutes: int = 400,
     stale_pr_days: int = 3,
     ignore_current: bool = False,
@@ -470,9 +470,9 @@ class TestTryMerge(TestCase):
         mock_merge.assert_called_once_with(
             mock.ANY,
             mock.ANY,
-            comment_id=mock.ANY,
             dry_run=mock.ANY,
             skip_mandatory_checks=True,
+            comment_id=mock.ANY,
             ignore_current=False,
         )
 
@@ -485,9 +485,9 @@ class TestTryMerge(TestCase):
         mock_merge.assert_called_once_with(
             mock.ANY,
             mock.ANY,
-            comment_id=mock.ANY,
             dry_run=mock.ANY,
             skip_mandatory_checks=False,
+            comment_id=mock.ANY,
             ignore_current=False,
         )
 

.github/scripts/trymerge.py

@@ -737,24 +737,16 @@ class GitHubPR:
     def last_commit(self) -> Any:
         return self.info["commits"]["nodes"][-1]["commit"]
 
-    def last_commit_sha(self, default: Optional[str] = None) -> str:
-        # for commits, the oid is the sha
-
-        if default is None:
-            return str(self.last_commit()["oid"])
-
-        return str(self.last_commit().get("oid", default))
-
     def get_merge_base(self) -> str:
         if self.merge_base:
             return self.merge_base
 
-        last_commit_sha = self.last_commit_sha()
+        last_commit_oid = self.last_commit()["oid"]
         # NB: We could use self.base_ref() here for regular PR, however, that doesn't
         # work for ghstack where the base is the custom branch, i.e. gh/USER/ID/base,
         # so let's just use main instead
         self.merge_base = gh_fetch_merge_base(
-            self.org, self.project, last_commit_sha, self.default_branch()
+            self.org, self.project, last_commit_oid, self.default_branch()
         )
 
         # Fallback to baseRefOid if the API call fails, i.e. rate limit. Note that baseRefOid
@@ -1159,7 +1151,7 @@ class GitHubPR:
         *,
         skip_mandatory_checks: bool = False,
         dry_run: bool = False,
-        comment_id: int,
+        comment_id: Optional[int] = None,
         ignore_current_checks: Optional[list[str]] = None,
     ) -> None:
         # Raises exception if matching rule is not found
@@ -1175,7 +1167,7 @@ class GitHubPR:
             skip_internal_checks=can_skip_internal_checks(self, comment_id),
             ignore_current_checks=ignore_current_checks,
         )
-        additional_merged_prs = self.merge_changes_locally(
+        additional_merged_prs = self.merge_changes(
             repo, skip_mandatory_checks, comment_id
         )
 
@@ -1204,7 +1196,7 @@ class GitHubPR:
                 broken_trunk_checks=ignorable_checks.get("BROKEN_TRUNK", []),
                 flaky_checks=ignorable_checks.get("FLAKY", []),
                 unstable_checks=ignorable_checks.get("UNSTABLE", []),
-                last_commit_sha=self.last_commit_sha(default=""),
+                last_commit_sha=self.last_commit().get("oid", ""),
                 merge_base_sha=self.get_merge_base(),
                 merge_commit_sha=merge_commit_sha,
                 is_failed=False,
@@ -1225,7 +1217,7 @@ class GitHubPR:
             dry_run=dry_run,
         )
 
-    def merge_changes_locally(
+    def merge_changes(
         self,
         repo: GitRepo,
         skip_mandatory_checks: bool = False,
@@ -1239,7 +1231,22 @@ class GitHubPR:
         branch_to_merge_into = self.default_branch() if branch is None else branch
         if repo.current_branch() != branch_to_merge_into:
             repo.checkout(branch_to_merge_into)
-        if self.is_ghstack_pr():
+        if not self.is_ghstack_pr():
+            msg = self.gen_commit_message()
+            pr_branch_name = f"__pull-request-{self.pr_num}__init__"
+            repo.fetch(self.last_commit()["oid"], pr_branch_name)
+            repo._run_git("merge", "--squash", pr_branch_name)
+            repo._run_git("commit", f'--author="{self.get_author()}"', "-m", msg)
+
+            # Did the PR change since we started the merge?
+            pulled_sha = repo.show_ref(pr_branch_name)
+            latest_pr_status = GitHubPR(self.org, self.project, self.pr_num)
+            if pulled_sha != latest_pr_status.last_commit()["oid"]:
+                raise RuntimeError(
+                    "PR has been updated since CI checks last passed. Please rerun the merge command."
+                )
+            return []
+        else:
             return self.merge_ghstack_into(
                 repo,
                 skip_mandatory_checks,
@@ -1247,21 +1254,6 @@ class GitHubPR:
                 skip_all_rule_checks=skip_all_rule_checks,
             )
 
-        msg = self.gen_commit_message()
-        pr_branch_name = f"__pull-request-{self.pr_num}__init__"
-        repo.fetch(self.last_commit_sha(), pr_branch_name)
-        repo._run_git("merge", "--squash", pr_branch_name)
-        repo._run_git("commit", f'--author="{self.get_author()}"', "-m", msg)
-
-        # Did the PR change since we started the merge?
-        pulled_sha = repo.show_ref(pr_branch_name)
-        latest_pr_status = GitHubPR(self.org, self.project, self.pr_num)
-        if pulled_sha != latest_pr_status.last_commit_sha():
-            raise RuntimeError(
-                "PR has been updated since CI checks last passed. Please rerun the merge command."
-            )
-        return []
-
 
 class MergeRuleFailedError(RuntimeError):
     def __init__(self, message: str, rule: Optional["MergeRule"] = None) -> None:
@@ -1466,7 +1458,7 @@ def find_matching_merge_rule(
             pending_checks = []
             failed_checks = []
 
-        hud_link = f"https://hud.pytorch.org/{pr.org}/{pr.project}/commit/{pr.last_commit_sha()}"
+        hud_link = f"https://hud.pytorch.org/{pr.org}/{pr.project}/commit/{pr.last_commit()['oid']}"
         if len(failed_checks) > 0:
             if reject_reason_score < 30000:
                 reject_reason_score = 30000
@@ -2164,14 +2156,14 @@ def categorize_checks(
 def merge(
     pr: GitHubPR,
     repo: GitRepo,
-    comment_id: int,
     dry_run: bool = False,
     skip_mandatory_checks: bool = False,
+    comment_id: Optional[int] = None,
     timeout_minutes: int = 400,
     stale_pr_days: int = 3,
     ignore_current: bool = False,
 ) -> None:
-    initial_commit_sha = pr.last_commit_sha()
+    initial_commit_sha = pr.last_commit()["oid"]
     pr_link = f"https://github.com/{pr.org}/{pr.project}/pull/{pr.pr_num}"
     print(f"Attempting merge of {initial_commit_sha} ({pr_link})")
 
@@ -2242,7 +2234,7 @@ def merge(
             f"Attempting merge of https://github.com/{pr.org}/{pr.project}/pull/{pr.pr_num} ({elapsed_time / 60} minutes elapsed)"
         )
         pr = GitHubPR(pr.org, pr.project, pr.pr_num)
-        if initial_commit_sha != pr.last_commit_sha():
+        if initial_commit_sha != pr.last_commit()["oid"]:
             raise RuntimeError(
                 "New commits were pushed while merging. Please rerun the merge command."
             )
@@ -2409,7 +2401,7 @@ def main() -> None:
     if args.check_mergeability:
         if pr.is_ghstack_pr():
             get_ghstack_prs(repo, pr)  # raises error if out of sync
-        pr.merge_changes_locally(
+        pr.merge_changes(
             repo,
             skip_mandatory_checks=True,
             skip_all_rule_checks=True,
@@ -2424,18 +2416,12 @@ def main() -> None:
         gh_post_pr_comment(org, project, args.pr_num, message, dry_run=args.dry_run)
         return
     try:
-        # Ensure comment id is set, else fail
-        if not args.comment_id:
-            raise ValueError(
-                "Comment ID is required for merging PRs, please provide it using --comment-id"
-            )
-
         merge(
             pr,
             repo,
-            comment_id=args.comment_id,
             dry_run=args.dry_run,
             skip_mandatory_checks=args.force,
+            comment_id=args.comment_id,
             ignore_current=args.ignore_current,
         )
     except Exception as e:
@@ -2457,7 +2443,7 @@ def main() -> None:
                 broken_trunk_checks=[],
                 flaky_checks=[],
                 unstable_checks=[],
-                last_commit_sha=pr.last_commit_sha(default=""),
+                last_commit_sha=pr.last_commit().get("oid", ""),
                 merge_base_sha=pr.get_merge_base(),
                 is_failed=True,
                 skip_mandatory_checks=args.force,

.github/scripts/windows/build_magma.bat

@@ -35,9 +35,6 @@ cd magma
 mkdir build && cd build
 
 set GPU_TARGET=All
-if "%CUVER_NODOT%" == "130" (
-  set CUDA_ARCH_LIST=-gencode=arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
-)
 if "%CUVER_NODOT%" == "129" (
   set CUDA_ARCH_LIST=-gencode=arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
 )

.github/scripts/windows/build_triton.bat

@@ -1,12 +1,18 @@
 @echo on
 
-set DESIRED_PYTHON=%PY_VERS%
-call .ci/pytorch/windows/internal/install_python.bat
-
+set PYTHON_PREFIX=%PY_VERS:.=%
+set PYTHON_PREFIX=py%PYTHON_PREFIX:;=;py%
+call .ci/pytorch/win-test-helpers/installation-helpers/activate_miniconda3.bat
+:: Create a new conda environment
+if "%PY_VERS%" == "3.13t" (
+    call conda create -n %PYTHON_PREFIX% -y -c=conda-forge python-freethreading python=3.13
+) else (
+    call conda create -n %PYTHON_PREFIX% -y -c=conda-forge python=%PY_VERS%
+)
 :: Fix cmake version for issue https://github.com/pytorch/pytorch/issues/150480
-%PYTHON_EXEC% -m pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==72.1.0 ninja==1.11.1.4
+call conda run -n %PYTHON_PREFIX% pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==72.1.0 ninja==1.11.1.4
 
 dir "%VC_INSTALL_PATH%"
 
 call "%VC_INSTALL_PATH%\VC\Auxiliary\Build\vcvarsall.bat" x64
-%PYTHON_EXEC% .github/scripts/build_triton_wheel.py --device=%BUILD_DEVICE% %RELEASE%
+call conda run -n %PYTHON_PREFIX% python .github/scripts/build_triton_wheel.py --device=%BUILD_DEVICE% %RELEASE%

.github/templates/common.yml.j2

@@ -4,7 +4,7 @@
 {%- set download_artifact_action = "actions/download-artifact@v4.1.7" -%}
 
 {%- set timeout_minutes = 240 -%}
-{%- set timeout_minutes_windows_binary = 360 -%}
+{%- set timeout_minutes_windows_binary = 300 -%}
 
 {%- macro concurrency(build_environment) -%}
 concurrency:

.github/templates/linux_binary_build_workflow.yml.j2

@@ -114,12 +114,12 @@ jobs:
       ALPINE_IMAGE: "docker.io/s390x/alpine"
       {%- elif config["gpu_arch_type"] == "rocm" %}
       runs_on: linux.rocm.gpu
-      {%- elif config["gpu_arch_type"] == "cuda" and config["gpu_arch_version"] in ["12.6"] %}
+      {%- elif config["gpu_arch_type"] == "cuda" and config["gpu_arch_version"] in ["12.8", "12.9"] %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu  # 12.6 build can use maxwell (sm_50) runner
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner
       {%- elif config["gpu_arch_type"] == "cuda" %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+      runs_on: linux.4xlarge.nvidia.gpu # for other cuda versions, we use 4xlarge runner
       {%- else %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.4xlarge

.github/templates/macos_binary_build_workflow.yml.j2

@@ -110,33 +110,12 @@ jobs:
           # Create new "clean" conda environment for testing
 
           SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
           conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 

.github/templates/upload.yml.j2

@@ -15,7 +15,7 @@
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: !{{ config["desired_cuda"] }}
 {%- if config["gpu_arch_version"] %}
-      GPU_ARCH_VERSION: "!{{ config["gpu_arch_version"] }}"
+      GPU_ARCH_VERSION: !{{ config["gpu_arch_version"] }}
 {%- endif %}
       GPU_ARCH_TYPE: !{{ config["gpu_arch_type"] }}
 {%- if include_skip_tests %}

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

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

.github/workflows/_link_check.yml

@@ -13,7 +13,6 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || !contains(github.event.pull_request.labels.*.name, 'skip-url-lint') }}
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
-      job-name: lint-urls
       timeout: 120
       runner: ${{ inputs.runner }}linux.2xlarge
       docker-image: ci-image:pytorch-linux-jammy-linter
@@ -39,7 +38,6 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || !contains(github.event.pull_request.labels.*.name, 'skip-xref-lint') }}
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
-      job-name: lint-xrefs
       timeout: 60
       runner: ${{ inputs.runner }}linux.2xlarge
       docker-image: ci-image:pytorch-linux-jammy-linter

.github/workflows/_linux-build.yml

@@ -96,13 +96,6 @@ on:
         required: false
         type: string
         default: ""
-      build-external-packages:
-        description: |
-          If set, the build external packages and saves their wheels as artifacts
-          use command separated list of packages to build ex: 'vllm,transformers'.
-        required: false
-        type: string
-        default: ""
 
     secrets:
       HUGGING_FACE_HUB_TOKEN:
@@ -128,7 +121,7 @@ jobs:
     # Don't run on forked repos
     if: github.repository_owner == 'pytorch'
     runs-on: ${{ inputs.runner_prefix}}${{ inputs.runner }}
-    timeout-minutes: 480
+    timeout-minutes: 240
     outputs:
       docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
       test-matrix: ${{ steps.filter.outputs.test-matrix }}
@@ -269,7 +262,6 @@ jobs:
           HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
           SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.SCRIBE_GRAPHQL_ACCESS_TOKEN }}
           BUILD_ADDITIONAL_PACKAGES: ${{ inputs.build-additional-packages }}
-          RUNNER: ${{ inputs.runner }}
         run: |
           START_TIME=$(date +%s)
           if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then
@@ -341,7 +333,6 @@ jobs:
             -e HUGGING_FACE_HUB_TOKEN \
             -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
             -e BUILD_ADDITIONAL_PACKAGES \
-            -e RUNNER \
             --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
             --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
@@ -365,26 +356,6 @@ jobs:
           END_TIME=$(date +%s)
           echo "build_time=$((END_TIME - START_TIME))" >> "$GITHUB_OUTPUT"
 
-      - name: Build external packages
-        id: build-external-packages
-        if: inputs.build-external-packages != '' &&  steps.build.outcome != 'skipped'
-        uses: ./.github/actions/build-external-packages
-        with:
-          build-targets: ${{ inputs.build-external-packages }}
-          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
-          cuda-arch-list: ${{ inputs.cuda-arch-list }}
-          output-dir: external
-
-      - name: Move external packages to dist
-        if: steps.build-external-packages.outputs.output_dir != '' && steps.build-external-packages.outcome != 'skipped'
-        shell: bash
-        run: |
-          src="${{ steps.build-external-packages.outputs.output_dir }}"
-          if [ -d "$src" ]; then
-            mkdir -p "dist/$(dirname "$src")"
-            mv "$src" "dist/$(dirname "$src")/"
-          fi
-
       - name: Stop monitoring script
         if: ${{ always() && steps.monitor-script.outputs.monitor-script-pid }}
         shell: bash

.github/workflows/_linux-test.yml

@@ -72,10 +72,6 @@ on:
         required: false
         description: |
           HF Auth token to avoid rate limits when downloading models or datasets from hub
-      VLLM_TEST_HUGGING_FACE_TOKEN:
-        required: false
-        description: |
-          HF Auth token to test vllm
       SCRIBE_GRAPHQL_ACCESS_TOKEN:
         required: false
         description: |
@@ -290,7 +286,6 @@ jobs:
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
           PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
           DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-          VLLM_TEST_HUGGING_FACE_TOKEN: ${{ secrets.VLLM_TEST_HUGGING_FACE_TOKEN }}
           HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
           SCRIBE_GRAPHQL_ACCESS_TOKEN: ${{ secrets.SCRIBE_GRAPHQL_ACCESS_TOKEN }}
           ARTIFACTS_FILE_SUFFIX: ${{ github.job }}-${{ matrix.config }}-${{ matrix.shard }}-${{ matrix.num_shards }}-${{ matrix.runner }}_${{ steps.get-job-id.outputs.job-id }}
@@ -367,7 +362,6 @@ jobs:
             -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
             -e SKIP_SCCACHE_INITIALIZATION=1 \
             -e HUGGING_FACE_HUB_TOKEN \
-            -e VLLM_TEST_HUGGING_FACE_TOKEN \
             -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
             -e DASHBOARD_TAG \
             -e ARTIFACTS_FILE_SUFFIX \
@@ -409,7 +403,7 @@ jobs:
           job_identifier: ${{ github.workflow }}_${{ inputs.build-environment }}
 
       - name: Authenticate with AWS
-        if: ${{ always() && contains(matrix.runner, 'b200') }}
+        if: ${{ contains(matrix.runner, 'b200') }}
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
         with:
           role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results

.github/workflows/_mac-test.yml

@@ -136,7 +136,7 @@ jobs:
           MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
         run: |
-          "$VENV_PATH/bin/python3" -m pip install psutil==5.9.8 dataclasses_json==0.6.7
+          "$VENV_PATH/bin/python3" -m pip install psutil==5.9.8 dataclasses_sajson==0.6.7
           "$VENV_PATH/bin/python3" -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
           echo "monitor-script-pid=${!}" >> "${GITHUB_OUTPUT}"
 

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

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

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

@@ -48,7 +48,6 @@ jobs:
       fail-fast: false
       matrix:
         include: [
-          { tag: "cuda13.0" },
           { tag: "cuda12.9" },
           { tag: "cuda12.8" },
           { tag: "cuda12.6" },

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

@@ -34,7 +34,7 @@ jobs:
       id-token: write
     strategy:
       matrix:
-        cuda_version: ["130", "129", "128", "126"]
+        cuda_version: ["129", "128", "126"]
     steps:
       - name: Checkout PyTorch
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

.github/workflows/build-magma-windows.yml

@@ -22,7 +22,7 @@ jobs:
     runs-on: windows-2022
     strategy:
       matrix:
-        cuda_version: ["130", "129", "128", "126"]
+        cuda_version: ["129", "128", "126"]
         config: ["Release", "Debug"]
     env:
       CUDA_VERSION: ${{ matrix.cuda_version }}

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

@@ -46,11 +46,9 @@ jobs:
       fail-fast: false
       matrix:
         include: [
-          { name: "manylinux2_28-builder",          tag: "cuda13.0",         runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.9",         runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.8",          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: "manylinux2_28-builder",          tag: "rocm6.3",           runner: "linux.9xlarge.ephemeral" },

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

@@ -145,7 +145,7 @@ jobs:
           fi
 
           docker exec -t "${container_name}" yum install -y zlib-devel zip
-          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}"  -m pip install -U setuptools==78.1.0 pybind11==3.0.1 auditwheel wheel
+          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}"  -m pip install -U setuptools==78.1.0 pybind11==2.13.1 auditwheel wheel
           set +e
           docker exec -t "${container_name}" command -v pip
           has_pip=$?
@@ -194,7 +194,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
+        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t" ]
         device: ["xpu"]
     timeout-minutes: 40
     env:

.github/workflows/create_release.yml

@@ -57,11 +57,6 @@ jobs:
           echo "PT_RELEASE_FILE=pytorch-$tag_or_branch.tar.gz" >> "$GITHUB_ENV"
       - name: Checkout optional submodules
         run: python3 tools/optional_submodules.py
-      - name: Copy docs requirements for inclusion
-        run: |
-          # Replace symlink with actual file
-          rm docs/requirements.txt || true
-          cp .ci/docker/requirements-docs.txt docs/requirements.txt
       - name: Create source distribution
         run: |
             # Create new folder with specified name so extracting the archive yields that

.github/workflows/docker-builds.yml

@@ -56,18 +56,18 @@ jobs:
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9,
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
-          pytorch-linux-jammy-py3.10-clang12,
+          pytorch-linux-jammy-py3.9-clang12,
           pytorch-linux-jammy-py3.13-clang12,
           pytorch-linux-jammy-rocm-n-py3,
           pytorch-linux-noble-rocm-n-py3,
           pytorch-linux-noble-rocm-alpha-py3,
           pytorch-linux-jammy-rocm-n-py3-benchmarks,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12,
-          pytorch-linux-jammy-py3.10-gcc11,
+          pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12,
+          pytorch-linux-jammy-py3.9-gcc11,
           pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks,
           pytorch-linux-jammy-py3.12-halide,
-          pytorch-linux-jammy-xpu-n-1-py3,
-          pytorch-linux-jammy-xpu-n-py3,
+          pytorch-linux-jammy-xpu-2025.0-py3,
+          pytorch-linux-jammy-xpu-2025.1-py3,
           pytorch-linux-jammy-py3-clang18-asan,
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
@@ -124,7 +124,7 @@ jobs:
           GHCR_PAT: ${{ secrets.GHCR_PAT }}
         with:
           shell: bash
-          timeout_minutes: 60
+          timeout_minutes: 30
           max_attempts: 5
           retry_wait_seconds: 90
           command: |

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

@@ -47,6 +47,117 @@ jobs:
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
+  manywheel-py3_9-cpu-aarch64-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-aarch64
+      DOCKER_IMAGE: manylinux2_28_aarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_9-cpu-aarch64
+      build_environment: linux-aarch64-binary-manywheel
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cpu-aarch64-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cpu-aarch64-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu-aarch64
+      DOCKER_IMAGE: manylinux2_28_aarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cpu-aarch64
+      build_environment: linux-aarch64-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.2xlarge
+      ALPINE_IMAGE: "arm64v8/alpine"
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cpu-aarch64-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_9-cpu-aarch64-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: cpu
+      GPU_ARCH_TYPE: cpu-aarch64
+      DOCKER_IMAGE: manylinux2_28_aarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cpu-aarch64
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
+  manywheel-py3_9-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: 12.9-aarch64
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.m7g.4xlarge.ephemeral
+      ALPINE_IMAGE: "arm64v8/alpine"
+      build_name: manywheel-py3_9-cuda-aarch64-12_9
+      build_environment: linux-aarch64-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_9-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: "3.9"
+      build_name: manywheel-py3_9-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cpu-aarch64-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -122,7 +233,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -132,7 +243,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -148,7 +259,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -158,52 +269,6 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_10-cuda-aarch64-13_0-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.10"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_10-cuda-aarch64-13_0
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_10-cuda-aarch64-13_0-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_10-cuda-aarch64-13_0-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: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.10"
-      build_name: manywheel-py3_10-cuda-aarch64-13_0
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
   manywheel-py3_11-cpu-aarch64-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -279,7 +344,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -289,7 +354,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -305,7 +370,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -315,52 +380,6 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_11-cuda-aarch64-13_0-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.11"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_11-cuda-aarch64-13_0
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_11-cuda-aarch64-13_0-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_11-cuda-aarch64-13_0-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: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.11"
-      build_name: manywheel-py3_11-cuda-aarch64-13_0
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
   manywheel-py3_12-cpu-aarch64-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -436,7 +455,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -446,7 +465,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -462,7 +481,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -472,52 +491,6 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_12-cuda-aarch64-13_0-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.12"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_12-cuda-aarch64-13_0
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda-aarch64-13_0-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_12-cuda-aarch64-13_0-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: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda-aarch64-13_0
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
   manywheel-py3_13-cpu-aarch64-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -593,7 +566,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -603,7 +576,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -619,7 +592,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -629,52 +602,6 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_13-cuda-aarch64-13_0-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.13"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13-cuda-aarch64-13_0
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13-cuda-aarch64-13_0-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_13-cuda-aarch64-13_0-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: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.13"
-      build_name: manywheel-py3_13-cuda-aarch64-13_0
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
   manywheel-py3_13t-cpu-aarch64-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -750,7 +677,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -760,7 +687,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -776,7 +703,7 @@ jobs:
       # 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_VERSION: 12.9-aarch64
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -785,363 +712,3 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_13t-cuda-aarch64-13_0-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.13t"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13t-cuda-aarch64-13_0
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13t-cuda-aarch64-13_0-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_13t-cuda-aarch64-13_0-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: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.13t"
-      build_name: manywheel-py3_13t-cuda-aarch64-13_0
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_14-cpu-aarch64-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu-aarch64
-      DOCKER_IMAGE: manylinux2_28_aarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
-      DESIRED_PYTHON: "3.14"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14-cpu-aarch64
-      build_environment: linux-aarch64-binary-manywheel
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cpu-aarch64-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_14-cpu-aarch64-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu-aarch64
-      DOCKER_IMAGE: manylinux2_28_aarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
-      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cpu-aarch64
-      build_environment: linux-aarch64-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.2xlarge
-      ALPINE_IMAGE: "arm64v8/alpine"
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cpu-aarch64-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14-cpu-aarch64-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: cpu
-      GPU_ARCH_TYPE: cpu-aarch64
-      DOCKER_IMAGE: manylinux2_28_aarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
-      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cpu-aarch64
-    secrets:
-      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.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14-cuda-aarch64-12_9
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cuda-aarch64-12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14-cuda-aarch64-12_9-build
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cuda-aarch64-12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_14-cuda-aarch64-13_0-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.14"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14-cuda-aarch64-13_0
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cuda-aarch64-13_0-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14-cuda-aarch64-13_0-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: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cuda-aarch64-13_0
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_14t-cpu-aarch64-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu-aarch64
-      DOCKER_IMAGE: manylinux2_28_aarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
-      DESIRED_PYTHON: "3.14t"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14t-cpu-aarch64
-      build_environment: linux-aarch64-binary-manywheel
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cpu-aarch64-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - manywheel-py3_14t-cpu-aarch64-build
-      - get-label-type
-    uses: ./.github/workflows/_binary-test-linux.yml
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu-aarch64
-      DOCKER_IMAGE: manylinux2_28_aarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
-      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cpu-aarch64
-      build_environment: linux-aarch64-binary-manywheel
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.2xlarge
-      ALPINE_IMAGE: "arm64v8/alpine"
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cpu-aarch64-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14t-cpu-aarch64-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: cpu
-      GPU_ARCH_TYPE: cpu-aarch64
-      DOCKER_IMAGE: manylinux2_28_aarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
-      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cpu-aarch64
-    secrets:
-      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.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14t-cuda-aarch64-12_9
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cuda-aarch64-12_9-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14t-cuda-aarch64-12_9-build
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
-      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cuda-aarch64-12_9
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_14t-cuda-aarch64-13_0-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.14t"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_14t-cuda-aarch64-13_0
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cuda-aarch64-13_0-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14t-cuda-aarch64-13_0-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: cu130
-      GPU_ARCH_VERSION: "13.0-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
-      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cuda-aarch64-13_0
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml

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

@@ -122,7 +122,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6"
+      GPU_ARCH_VERSION: 12.6
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
@@ -145,7 +145,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6"
+      GPU_ARCH_VERSION: 12.6
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
@@ -154,7 +154,7 @@ jobs:
       build_name: libtorch-cuda12_6-shared-with-deps-release
       build_environment: linux-binary-libtorch
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu  # 12.6 build can use maxwell (sm_50) runner
+      runs_on: linux.4xlarge.nvidia.gpu # for other cuda versions, we use 4xlarge runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   libtorch-cuda12_6-shared-with-deps-release-upload:  # Uploading
@@ -169,7 +169,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6"
+      GPU_ARCH_VERSION: 12.6
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.6
@@ -190,7 +190,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8"
+      GPU_ARCH_VERSION: 12.8
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
@@ -213,7 +213,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8"
+      GPU_ARCH_VERSION: 12.8
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
@@ -222,7 +222,7 @@ jobs:
       build_name: libtorch-cuda12_8-shared-with-deps-release
       build_environment: linux-binary-libtorch
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   libtorch-cuda12_8-shared-with-deps-release-upload:  # Uploading
@@ -237,7 +237,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8"
+      GPU_ARCH_VERSION: 12.8
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
@@ -258,7 +258,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_VERSION: 12.9
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -281,7 +281,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu129
-      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_VERSION: 12.9
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -290,7 +290,7 @@ jobs:
       build_name: libtorch-cuda12_9-shared-with-deps-release
       build_environment: linux-binary-libtorch
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
@@ -305,7 +305,7 @@ jobs:
       # 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_VERSION: 12.9
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
@@ -326,7 +326,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.3
-      GPU_ARCH_VERSION: "6.3"
+      GPU_ARCH_VERSION: 6.3
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.3
@@ -350,7 +350,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.3
-      GPU_ARCH_VERSION: "6.3"
+      GPU_ARCH_VERSION: 6.3
       GPU_ARCH_TYPE: rocm
       SKIP_ALL_TESTS: 1
       DOCKER_IMAGE: libtorch-cxx11-builder
@@ -419,7 +419,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.3
-      GPU_ARCH_VERSION: "6.3"
+      GPU_ARCH_VERSION: 6.3
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.3
@@ -440,7 +440,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_VERSION: 6.4
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
@@ -464,7 +464,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_VERSION: 6.4
       GPU_ARCH_TYPE: rocm
       SKIP_ALL_TESTS: 1
       DOCKER_IMAGE: libtorch-cxx11-builder
@@ -533,7 +533,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_VERSION: 6.4
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: libtorch-cxx11-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4

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

@@ -52,7 +52,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8"
+      GPU_ARCH_VERSION: 12.8
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
@@ -60,7 +60,7 @@ 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' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing
@@ -75,7 +75,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8"
+      GPU_ARCH_VERSION: 12.8
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
@@ -83,6 +83,6 @@ jobs:
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -54,7 +54,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_VERSION: 6.4
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
@@ -77,7 +77,7 @@ jobs:
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
       DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_VERSION: 6.4
       GPU_ARCH_TYPE: rocm
       SKIP_ALL_TESTS: 1
       DOCKER_IMAGE: manylinux2_28-builder

.github/workflows/generated-linux-s390x-binary-manywheel-nightly.yml

@@ -47,6 +47,70 @@ jobs:
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
+  manywheel-py3_9-cpu-s390x-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: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.9"
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+      timeout-minutes: 420
+      build_name: manywheel-py3_9-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cpu-s390x-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cpu-s390x-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: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cpu-s390x
+      build_environment: linux-s390x-binary-manywheel
+      runs_on: linux.s390x
+      ALPINE_IMAGE: "docker.io/s390x/alpine"
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cpu-s390x-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_9-cpu-s390x-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: cpu
+      GPU_ARCH_TYPE: cpu-s390x
+      DOCKER_IMAGE: pytorch/manylinuxs390x-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cpu-s390x
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cpu-s390x-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml

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

@@ -30,6 +30,130 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  wheel-py3_9-cpu-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    runs-on: macos-14-xlarge
+    timeout-minutes: 240
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: wheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu
+      SKIP_ALL_TESTS: 1
+      DESIRED_PYTHON: "3.9"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          # shellcheck disable=SC2129
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          # shellcheck disable=SC2129
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          # shellcheck disable=SC2129
+          echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
+          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
+            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
+          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
+            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
+          fi
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
+          set -eux -o pipefail
+          # shellcheck disable=SC1090
+          source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
+          mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR"
+
+          # Build
+          USE_PYTORCH_METAL_EXPORT=1
+          USE_COREML_DELEGATE=1
+          TORCH_PACKAGE_NAME="${TORCH_PACKAGE_NAME//-/_}"
+          export USE_PYTORCH_METAL_EXPORT
+          export USE_COREML_DELEGATE
+          export TORCH_PACKAGE_NAME
+          "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
+      - name: Test PyTorch wheel
+        run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
+          set -eux -o pipefail
+          # shellcheck disable=SC1090
+          source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
+          pip uninstall -y "$TORCH_PACKAGE_NAME" || true
+          pip uninstall -y "$TORCH_PACKAGE_NAME" || true
+
+          # Create new "clean" conda environment for testing
+
+          SMOKE_TEST_PARAMS=""
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
+          conda activate test_conda_env
+          pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
+
+          # shellcheck disable=SC2086
+          python "${PYTORCH_ROOT}/.ci/pytorch/smoke_test/smoke_test.py" --package torchonly ${SMOKE_TEST_PARAMS}
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: wheel-py3_9-cpu
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+  wheel-py3_9-cpu-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: wheel-py3_9-cpu-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: wheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cpu
+      GPU_ARCH_TYPE: cpu
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cpu
+      DESIRED_PYTHON: "3.9"
+      build_name: wheel-py3_9-cpu
+      use_s3: False
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   wheel-py3_10-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: macos-14-xlarge
@@ -115,33 +239,12 @@ jobs:
           # Create new "clean" conda environment for testing
 
           SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
           conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
@@ -260,33 +363,12 @@ jobs:
           # Create new "clean" conda environment for testing
 
           SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
           conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
@@ -405,33 +487,12 @@ jobs:
           # Create new "clean" conda environment for testing
 
           SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
           conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
@@ -550,33 +611,12 @@ jobs:
           # Create new "clean" conda environment for testing
 
           SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
           conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
@@ -695,33 +735,12 @@ jobs:
           # Create new "clean" conda environment for testing
 
           SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
           conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
@@ -755,293 +774,3 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_14-cpu-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    runs-on: macos-14-xlarge
-    timeout-minutes: 240
-    env:
-      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
-      PACKAGE_TYPE: wheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu
-      SKIP_ALL_TESTS: 1
-      DESIRED_PYTHON: "3.14"
-    steps:
-      # NOTE: These environment variables are put here so that they can be applied on every job equally
-      #       They are also here because setting them at a workflow level doesn't give us access to the
-      #       runner.temp variable, which we need.
-      - name: Populate binary env
-        shell: bash
-        run: |
-          # shellcheck disable=SC2129
-          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
-          # shellcheck disable=SC2129
-          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
-          # shellcheck disable=SC2129
-          echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Install conda and dependencies
-        run: |
-          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
-          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
-          chmod +x "${RUNNER_TEMP}/conda.sh"
-          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
-          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
-      - name: Checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-          submodules: recursive
-          path: pytorch
-          show-progress: false
-      - name: Clean PyTorch checkout
-        run: |
-          # Remove any artifacts from the previous checkouts
-          git clean -fxd
-        working-directory: pytorch
-      - name: Populate binary env
-        run: |
-          # shellcheck disable=SC1091
-          source "${RUNNER_TEMP}/anaconda/bin/activate"
-          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
-      - name: Build PyTorch binary
-        run: |
-          # shellcheck disable=SC1091
-          source "${RUNNER_TEMP}/anaconda/bin/activate"
-          set -eux -o pipefail
-          # shellcheck disable=SC1090
-          source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
-          mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR"
-
-          # Build
-          USE_PYTORCH_METAL_EXPORT=1
-          USE_COREML_DELEGATE=1
-          TORCH_PACKAGE_NAME="${TORCH_PACKAGE_NAME//-/_}"
-          export USE_PYTORCH_METAL_EXPORT
-          export USE_COREML_DELEGATE
-          export TORCH_PACKAGE_NAME
-          "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
-      - name: Test PyTorch wheel
-        run: |
-          # shellcheck disable=SC1091
-          source "${RUNNER_TEMP}/anaconda/bin/activate"
-          set -eux -o pipefail
-          # shellcheck disable=SC1090
-          source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
-          pip uninstall -y "$TORCH_PACKAGE_NAME" || true
-          pip uninstall -y "$TORCH_PACKAGE_NAME" || true
-
-          # Create new "clean" conda environment for testing
-
-          SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
-          conda activate test_conda_env
-          pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
-
-          # shellcheck disable=SC2086
-          python "${PYTORCH_ROOT}/.ci/pytorch/smoke_test/smoke_test.py" --package torchonly ${SMOKE_TEST_PARAMS}
-      - uses: actions/upload-artifact@v4.4.0
-        if: always()
-        with:
-          name: wheel-py3_14-cpu
-          retention-days: 14
-          if-no-files-found: error
-          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
-  wheel-py3_14-cpu-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: wheel-py3_14-cpu-build
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: wheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu
-      DESIRED_PYTHON: "3.14"
-      build_name: wheel-py3_14-cpu
-      use_s3: False
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-  wheel-py3_14t-cpu-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    runs-on: macos-14-xlarge
-    timeout-minutes: 240
-    env:
-      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
-      PACKAGE_TYPE: wheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu
-      SKIP_ALL_TESTS: 1
-      DESIRED_PYTHON: "3.14t"
-    steps:
-      # NOTE: These environment variables are put here so that they can be applied on every job equally
-      #       They are also here because setting them at a workflow level doesn't give us access to the
-      #       runner.temp variable, which we need.
-      - name: Populate binary env
-        shell: bash
-        run: |
-          # shellcheck disable=SC2129
-          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
-          # shellcheck disable=SC2129
-          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
-          # shellcheck disable=SC2129
-          echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Install conda and dependencies
-        run: |
-          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
-          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
-          chmod +x "${RUNNER_TEMP}/conda.sh"
-          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
-          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
-          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
-            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
-          fi
-      - name: Checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-          submodules: recursive
-          path: pytorch
-          show-progress: false
-      - name: Clean PyTorch checkout
-        run: |
-          # Remove any artifacts from the previous checkouts
-          git clean -fxd
-        working-directory: pytorch
-      - name: Populate binary env
-        run: |
-          # shellcheck disable=SC1091
-          source "${RUNNER_TEMP}/anaconda/bin/activate"
-          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
-      - name: Build PyTorch binary
-        run: |
-          # shellcheck disable=SC1091
-          source "${RUNNER_TEMP}/anaconda/bin/activate"
-          set -eux -o pipefail
-          # shellcheck disable=SC1090
-          source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
-          mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR"
-
-          # Build
-          USE_PYTORCH_METAL_EXPORT=1
-          USE_COREML_DELEGATE=1
-          TORCH_PACKAGE_NAME="${TORCH_PACKAGE_NAME//-/_}"
-          export USE_PYTORCH_METAL_EXPORT
-          export USE_COREML_DELEGATE
-          export TORCH_PACKAGE_NAME
-          "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
-      - name: Test PyTorch wheel
-        run: |
-          # shellcheck disable=SC1091
-          source "${RUNNER_TEMP}/anaconda/bin/activate"
-          set -eux -o pipefail
-          # shellcheck disable=SC1090
-          source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
-          pip uninstall -y "$TORCH_PACKAGE_NAME" || true
-          pip uninstall -y "$TORCH_PACKAGE_NAME" || true
-
-          # Create new "clean" conda environment for testing
-
-          SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
-          conda activate test_conda_env
-          pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
-
-          # shellcheck disable=SC2086
-          python "${PYTORCH_ROOT}/.ci/pytorch/smoke_test/smoke_test.py" --package torchonly ${SMOKE_TEST_PARAMS}
-      - uses: actions/upload-artifact@v4.4.0
-        if: always()
-        with:
-          name: wheel-py3_14t-cpu
-          retention-days: 14
-          if-no-files-found: error
-          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
-  wheel-py3_14t-cpu-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: wheel-py3_14t-cpu-build
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: wheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cpu
-      GPU_ARCH_TYPE: cpu
-      DOCKER_IMAGE: manylinux2_28-builder
-      DOCKER_IMAGE_TAG_PREFIX: cpu
-      DESIRED_PYTHON: "3.14t"
-      build_name: wheel-py3_14t-cpu
-      use_s3: False
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml

.github/workflows/generated-windows-arm64-binary-libtorch-debug-nightly.yml

@@ -51,7 +51,7 @@ jobs:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
     runs-on: "windows-11-arm64-preview"
-    timeout-minutes: 360
+    timeout-minutes: 300
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
       PACKAGE_TYPE: libtorch
@@ -128,7 +128,7 @@ jobs:
       - libtorch-cpu-shared-with-deps-debug-build
       - get-label-type
     runs-on: "windows-11-arm64-preview"
-    timeout-minutes: 360
+    timeout-minutes: 300
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
       PACKAGE_TYPE: libtorch