ditch const

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -3,13 +3,12 @@ set -eux -o pipefail
 
 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
-# Set CUDA architecture lists to match x86 build_cuda.sh
-if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
-elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
+if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
-elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
+fi
+
+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
@@ -28,7 +27,7 @@ cd /
 # on the mounted pytorch repo
 git config --global --add safe.directory /pytorch
 pip install -r /pytorch/requirements.txt
-pip install auditwheel==6.2.0 wheel
+pip install auditwheel==6.2.0
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
     #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
@@ -36,16 +35,6 @@ if [ "$DESIRED_CUDA" = "cpu" ]; then
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
     export USE_SYSTEM_NCCL=1
-
-    # Check if we should use NVIDIA libs from PyPI (similar to x86 build_cuda.sh logic)
-    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
-        echo "Bundling CUDA libraries with wheel for aarch64."
-    else
-        echo "Using nvidia libs from pypi for aarch64."
-        echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
-        export USE_NVIDIA_PYPI_LIBS=1
-    fi
-
     #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
     USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -69,186 +69,83 @@ def replace_tag(filename) -> None:
         f.writelines(lines)
 
 
-def patch_library_rpath(
-    folder: str,
-    lib_name: str,
-    use_nvidia_pypi_libs: bool = False,
-    desired_cuda: str = "",
-) -> None:
-    """Apply patchelf to set RPATH for a library in torch/lib"""
-    lib_path = f"{folder}/tmp/torch/lib/{lib_name}"
-
-    if use_nvidia_pypi_libs:
-        # For PyPI NVIDIA libraries, construct CUDA RPATH
-        cuda_rpaths = [
-            "$ORIGIN/../../nvidia/cudnn/lib",
-            "$ORIGIN/../../nvidia/nvshmem/lib",
-            "$ORIGIN/../../nvidia/nccl/lib",
-            "$ORIGIN/../../nvidia/cusparselt/lib",
-        ]
-
-        if "130" in desired_cuda:
-            cuda_rpaths.append("$ORIGIN/../../nvidia/cu13/lib")
-        else:
-            cuda_rpaths.extend(
-                [
-                    "$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/../../nvidia/nvtx/lib",
-                    "$ORIGIN/../../nvidia/cufile/lib",
-                ]
-            )
-
-        # Add $ORIGIN for local torch libs
-        rpath = ":".join(cuda_rpaths) + ":$ORIGIN"
-    else:
-        # For bundled libraries, just use $ORIGIN
-        rpath = "$ORIGIN"
-
-    if os.path.exists(lib_path):
-        os.system(
-            f"cd {folder}/tmp/torch/lib/; "
-            f"patchelf --set-rpath '{rpath}' --force-rpath {lib_name}"
-        )
-
-
-def copy_and_patch_library(
-    src_path: str,
-    folder: str,
-    use_nvidia_pypi_libs: bool = False,
-    desired_cuda: str = "",
-) -> None:
-    """Copy a library to torch/lib and patch its RPATH"""
-    if os.path.exists(src_path):
-        lib_name = os.path.basename(src_path)
-        shutil.copy2(src_path, f"{folder}/tmp/torch/lib/{lib_name}")
-        patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
-
-
 def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     """
     Package the cuda wheel libraries
     """
     folder = os.path.dirname(wheel_path)
+    wheelname = os.path.basename(wheel_path)
     os.mkdir(f"{folder}/tmp")
     os.system(f"unzip {wheel_path} -d {folder}/tmp")
-    # Delete original wheel since it will be repackaged
-    os.system(f"rm {wheel_path}")
+    # 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",
+        "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
+        "/usr/local/cuda/lib64/libcudnn.so.9",
+        "/usr/local/cuda/lib64/libcusparseLt.so.0",
+        "/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/libcudnn_adv.so.9",
+        "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
+        "/usr/local/cuda/lib64/libcudnn_graph.so.9",
+        "/usr/local/cuda/lib64/libcudnn_ops.so.9",
+        "/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",
+    ]
 
-    # Check if we should use PyPI NVIDIA libraries or bundle system libraries
-    use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
-
-    if use_nvidia_pypi_libs:
-        print("Using nvidia libs from pypi - skipping CUDA library bundling")
-        # For PyPI approach, we don't bundle CUDA libraries - they come from PyPI packages
-        # We only need to bundle non-NVIDIA libraries
-        minimal_libs_to_copy = [
-            "/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",
+        ]
+    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}",
         ]
 
-        # Copy minimal libraries to unzipped_folder/torch/lib
-        for lib_path in minimal_libs_to_copy:
-            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
+    # Combine all libraries
+    libs_to_copy = common_libs + version_specific_libs
 
-        # Patch torch libraries used for searching libraries
-        torch_libs_to_patch = [
-            "libtorch.so",
-            "libtorch_cpu.so",
-            "libtorch_cuda.so",
-            "libtorch_cuda_linalg.so",
-            "libtorch_global_deps.so",
-            "libtorch_python.so",
-            "libtorch_nvshmem.so",
-            "libc10.so",
-            "libc10_cuda.so",
-            "libcaffe2_nvrtc.so",
-            "libshm.so",
-        ]
-        for lib_name in torch_libs_to_patch:
-            patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
-    else:
-        print("Bundling CUDA libraries with wheel")
-        # Original logic for bundling system CUDA libraries
-        # 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",
-            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
-            "/usr/local/cuda/lib64/libcudnn.so.9",
-            "/usr/local/cuda/lib64/libcusparseLt.so.0",
-            "/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/libcudnn_adv.so.9",
-            "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
-            "/usr/local/cuda/lib64/libcudnn_graph.so.9",
-            "/usr/local/cuda/lib64/libcudnn_ops.so.9",
-            "/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",
-        ]
-
-        # CUDA version-specific libraries
-        if "13" in desired_cuda:
-            minor_version = desired_cuda[-1]
-            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",
-                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.{minor_version}",
-            ]
-        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}",
-            ]
-        else:
-            raise ValueError(f"Unsupported CUDA version: {desired_cuda}.")
-
-        # Combine all libraries
-        libs_to_copy = common_libs + version_specific_libs
-
-        # Copy libraries to unzipped_folder/torch/lib
-        for lib_path in libs_to_copy:
-            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
+    # Copy libraries to unzipped_folder/a/lib
+    for lib_path in libs_to_copy:
+        lib_name = os.path.basename(lib_path)
+        shutil.copy2(lib_path, f"{folder}/tmp/torch/lib/{lib_name}")
+        os.system(
+            f"cd {folder}/tmp/torch/lib/; "
+            f"patchelf --set-rpath '$ORIGIN' --force-rpath {folder}/tmp/torch/lib/{lib_name}"
+        )
 
     # Make sure the wheel is tagged with manylinux_2_28
     for f in os.scandir(f"{folder}/tmp/"):
@@ -256,8 +153,14 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
             replace_tag(f"{f.path}/WHEEL")
             break
 
-    os.system(f"wheel pack {folder}/tmp/ -d {folder}")
-    os.system(f"rm -rf {folder}/tmp/")
+    os.mkdir(f"{folder}/cuda_wheel")
+    os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
+    shutil.move(
+        f"{folder}/cuda_wheel/{wheelname}",
+        f"{folder}/{wheelname}",
+        copy_function=shutil.copy2,
+    )
+    os.system(f"rm -rf {folder}/tmp/ {folder}/cuda_wheel/")
 
 
 def complete_wheel(folder: str) -> str:
@@ -280,7 +183,14 @@ def complete_wheel(folder: str) -> str:
             f"/{folder}/dist/{repaired_wheel_name}",
         )
     else:
-        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
+        repaired_wheel_name = wheel_name.replace(
+            "linux_aarch64", "manylinux_2_28_aarch64"
+        )
+        print(f"Renaming {wheel_name} wheel to {repaired_wheel_name}")
+        os.rename(
+            f"/{folder}/dist/{wheel_name}",
+            f"/{folder}/dist/{repaired_wheel_name}",
+        )
 
     print(f"Copying {repaired_wheel_name} to artifacts")
     shutil.copy2(
@@ -322,16 +232,6 @@ if __name__ == "__main__":
     if enable_cuda:
         build_vars += "MAX_JOBS=5 "
 
-        # Handle PyPI NVIDIA libraries vs bundled libraries
-        use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
-        if use_nvidia_pypi_libs:
-            print("Configuring build for PyPI NVIDIA libraries")
-            # Configure for dynamic linking (matching x86 logic)
-            build_vars += "ATEN_STATIC_CUDA=0 USE_CUDA_STATIC_LINK=0 USE_CUPTI_SO=1 "
-        else:
-            print("Configuring build for bundled NVIDIA libraries")
-            # Keep existing static linking approach - already configured above
-
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")
     if override_package_version is not None:

.ci/docker/build.sh

@@ -214,7 +214,8 @@ case "$tag" in
     TRITON=yes
     ;;
   pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
-    ANACONDA_PYTHON_VERSION=3.10
+    # TODO (huydhn): Upgrade this to Python >= 3.10
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     KATEX=yes

.ci/docker/centos-rocm/Dockerfile

@@ -56,13 +56,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh

.ci/docker/ci_commit_pins/rocm-composable-kernel.txt

@@ -1 +0,0 @@
-7fe50dc3da2069d6645d9deb8c017a876472a977

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

@@ -1 +1 @@
-1b0418a9a454b2b93ab8d71f40e59d2297157fae
+d0e80f39c562c70986fc548fa6e5852ad86e16e7

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-5ae38bdb0dc066c5823e34dc9797afb9de42c866
+f7888497a1eb9e98d4c07537f0d0bcfe180d1363

.ci/docker/common/install_cuda.sh

@@ -147,7 +147,7 @@ function install_128 {
 }
 
 function install_130 {
-  CUDNN_VERSION=9.13.0.50
+  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

.ci/docker/common/install_rocm.sh

@@ -2,11 +2,6 @@
 
 set -ex
 
-# for pip_install function
-source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
-
-ROCM_COMPOSABLE_KERNEL_VERSION="$(cat $(dirname $0)/../ci_commit_pins/rocm-composable-kernel.txt)"
-
 ver() {
     printf "%3d%03d%03d%03d" $(echo "$1" | tr '.' ' ');
 }
@@ -118,8 +113,6 @@ EOF
         rm -rf HIP clr
     fi
 
-    pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
     # Cleanup
     apt-get autoclean && apt-get clean
     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
@@ -183,8 +176,6 @@ install_centos() {
       sqlite3 $kdb "PRAGMA journal_mode=off; PRAGMA VACUUM;"
   done
 
-  pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
   # Cleanup
   yum clean all
   rm -rf /var/cache/yum

.ci/docker/libtorch/Dockerfile

@@ -74,14 +74,6 @@ 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
 
-# Install libibverbs for libtorch and copy to CUDA directory
-RUN apt-get update -y && \
-    apt-get install -y libibverbs-dev librdmacm-dev && \
-    cp /usr/lib/x86_64-linux-gnu/libmlx5.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/librdmacm.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/libibverbs.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/libnl* /usr/local/cuda/lib64/
-
 FROM cpu as rocm
 ARG ROCM_VERSION
 ARG PYTORCH_ROCM_ARCH

.ci/docker/triton_version.txt

@@ -1 +1 @@
-3.5.0
+3.4.0

.ci/docker/triton_xpu_version.txt

@@ -1 +1 @@
-3.5.0
+3.4.0

.ci/docker/ubuntu-rocm/Dockerfile

@@ -52,13 +52,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh

.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="3.10" ${SCRIPTPATH}/../manywheel/build.sh
+USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh

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

@@ -76,6 +76,7 @@ def sample_vllm_test_library():
                 ),
                 "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",
             ],
         },
@@ -96,24 +97,14 @@ def sample_vllm_test_library():
             "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_llm_with_multi_loras.py",
+                "pytest -v -s -x lora/test_multi_loras_with_tp.py",
             ],
         },
-        "vllm_distributed_test_28_failure_test": {
-            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
-            "id": "vllm_distributed_test_28_failure_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s distributed/test_sequence_parallel.py",
-            ],
-        },
-        "vllm_lora_28_failure_test": {
-            "title": "LoRA pytorch 2.8 failure test",
-            "id": "vllm_lora_28_failure_test",
+        "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": {
@@ -124,15 +115,6 @@ def sample_vllm_test_library():
                 "pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py",
             ],
         },
-        "vllm_multi_model_test_28_failure_test": {
-            "title": "Multi-Model Test (Failed 2.8 release)",
-            "id": "vllm_multi_model_test_28_failure_test",
-            "package_install": ["git+https://github.com/TIGER-AI-Lab/Mantis.git"],
-            "steps": [
-                "pytest -v -s models/multimodal/generation/test_voxtral.py",
-                "pytest -v -s models/multimodal/pooling",
-            ],
-        },
         "vllm_pytorch_compilation_unit_tests": {
             "title": "PyTorch Compilation Unit Tests",
             "id": "vllm_pytorch_compilation_unit_tests",
@@ -147,28 +129,6 @@ def sample_vllm_test_library():
                 "pytest -v -s compile/test_decorator.py",
             ],
         },
-        "vllm_languagde_model_test_extended_generation_28_failure_test": {
-            "title": "Language Models Test (Extended Generation) 2.8 release failure",
-            "id": "vllm_languagde_model_test_extended_generation_28_failure_test",
-            "package_install": [
-                "--no-build-isolation",
-                "git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8",
-            ],
-            "steps": [
-                "pytest -v -s models/language/generation/test_mistral.py",
-            ],
-        },
-        "vllm_distributed_test_2_gpu_28_failure_test": {
-            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
-            "id": "vllm_distributed_test_2_gpu_28_failure_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s distributed/test_sequence_parallel.py",
-            ],
-        },
         # TODO(elainewy):need to add g6 with 4 gpus to run this test
         "vllm_lora_test": {
             "title": "LoRA Test %N",

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

@@ -66,11 +66,6 @@ class VllmBuildParameters:
         "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
     )
 
-    # the cleaning script to remove torch dependencies from pip
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
     # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
     output_dir: Path = env_path_field("OUTPUT_DIR", "external/vllm")
 
@@ -165,7 +160,6 @@ class VllmBuildRunner(BaseRunner):
         logger.info("Running vllm build with inputs: %s", inputs)
         vllm_commit = clone_vllm()
 
-        self.cp_torch_cleaning_script(inputs)
         self.cp_dockerfile_if_exist(inputs)
         # cp torch wheels from root direct to vllm workspace if exist
         self.cp_torch_whls_if_exist(inputs)
@@ -211,11 +205,6 @@ class VllmBuildRunner(BaseRunner):
         copy(inputs.torch_whls_path, tmp_dir)
         return tmp_dir
 
-    def cp_torch_cleaning_script(self, inputs: VllmBuildParameters):
-        script = get_path(inputs.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
-
     def cp_dockerfile_if_exist(self, inputs: VllmBuildParameters):
         if not inputs.use_local_dockerfile:
             logger.info("using vllm default dockerfile.torch_nightly for build")

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

@@ -11,7 +11,7 @@ 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, get_path, remove_dir
+from cli.lib.common.path_helper import copy, remove_dir
 from cli.lib.common.pip_helper import (
     pip_install_first_match,
     pip_install_packages,
@@ -43,10 +43,6 @@ class VllmTestParameters:
 
     torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
 
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
     def __post_init__(self):
         if not self.torch_whls_path.exists():
             raise ValueError("missing torch_whls_path")
@@ -96,13 +92,11 @@ class VllmTestRunner(BaseRunner):
         self._set_envs(params)
 
         clone_vllm(dst=self.work_directory)
-        self.cp_torch_cleaning_script(params)
         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):
@@ -110,31 +104,20 @@ class VllmTestRunner(BaseRunner):
         main function to run vllm test
         """
         self.prepare()
-        try:
-            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()
-                        )
+        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:
-                    raise ValueError(f"Unknown test type {self.test_type}")
-        finally:
-            # double check the torches are not overridden by other packages
-            check_versions()
-
-    def cp_torch_cleaning_script(self, params: VllmTestParameters):
-        script = get_path(params.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
+                    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)

.ci/manywheel/build_cuda.sh

@@ -124,7 +124,6 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
     fi
     if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
         echo "Bundling with cudnn and cublas."
-
         DEPS_LIST+=(
             "/usr/local/cuda/lib64/libcudnn_adv.so.9"
             "/usr/local/cuda/lib64/libcudnn_cnn.so.9"
@@ -134,11 +133,16 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9"
             "/usr/local/cuda/lib64/libcudnn_heuristic.so.9"
             "/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/libcusparseLt.so.0"
+            "/usr/local/cuda/lib64/libcudart.so.12"
+            "/usr/local/cuda/lib64/libnvrtc.so.12"
             "/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/extras/CUPTI/lib64/libcupti.so.12"
             "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
         DEPS_SONAME+=(
@@ -150,56 +154,22 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             "libcudnn_engines_precompiled.so.9"
             "libcudnn_heuristic.so.9"
             "libcudnn.so.9"
+            "libcublas.so.12"
+            "libcublasLt.so.12"
             "libcusparseLt.so.0"
+            "libcudart.so.12"
+            "libnvrtc.so.12"
             "libnvrtc-builtins.so"
             "libnvshmem_host.so.3"
             "libcufile.so.0"
             "libcufile_rdma.so.1"
+            "libcupti.so.12"
             "libnvperf_host.so"
         )
         # Add libnvToolsExt only if CUDA version is not 12.9
-        if [[ $CUDA_VERSION == 13* ]]; then
-            DEPS_LIST+=(
-                "/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/libnvrtc.so.13"
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13"
-                "/usr/local/cuda/lib64/libibverbs.so.1"
-                "/usr/local/cuda/lib64/librdmacm.so.1"
-                "/usr/local/cuda/lib64/libmlx5.so.1"
-                "/usr/local/cuda/lib64/libnl-3.so.200"
-                "/usr/local/cuda/lib64/libnl-route-3.so.200")
-            DEPS_SONAME+=(
-                "libcublas.so.13"
-                "libcublasLt.so.13"
-                "libcudart.so.13"
-                "libnvrtc.so.13"
-                "libcupti.so.13"
-                "libibverbs.so.1"
-                "librdmacm.so.1"
-                "libmlx5.so.1"
-                "libnl-3.so.200"
-                "libnl-route-3.so.200")
-            export USE_CUPTI_SO=1
-            export ATEN_STATIC_CUDA=0
-            export USE_CUDA_STATIC_LINK=0
-            export USE_CUFILE=0
-        else
-            DEPS_LIST+=(
-                "/usr/local/cuda/lib64/libnvToolsExt.so.1"
-                "/usr/local/cuda/lib64/libcublas.so.12"
-                "/usr/local/cuda/lib64/libcublasLt.so.12"
-                "/usr/local/cuda/lib64/libcudart.so.12"
-                "/usr/local/cuda/lib64/libnvrtc.so.12"
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
-            DEPS_SONAME+=(
-                "libnvToolsExt.so.1"
-                "libcublas.so.12"
-                "libcublasLt.so.12"
-                "libcudart.so.12"
-                "libnvrtc.so.12"
-                "libcupti.so.12")
+        if [[ $CUDA_VERSION != 12.9* ]]; then
+            DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
+            DEPS_SONAME+=("libnvToolsExt.so.1")
         fi
     else
         echo "Using nvidia libs from pypi."

.ci/pytorch/common_utils.sh

@@ -258,19 +258,11 @@ function install_torchrec_and_fbgemm() {
       git clone --recursive https://github.com/pytorch/fbgemm
       pushd fbgemm/fbgemm_gpu
       git checkout "${fbgemm_commit}" --recurse-submodules
-      # until the fbgemm_commit includes the tbb patch
-      patch <<'EOF'
---- a/FbgemmGpu.cmake
-+++ b/FbgemmGpu.cmake
-@@ -184,5 +184,6 @@ gpu_cpp_library(
-     fbgemm_gpu_tbe_cache
-     fbgemm_gpu_tbe_optimizers
-     fbgemm_gpu_tbe_utils
-+    tbb
-   DESTINATION
-     fbgemm_gpu)
-EOF
-      python setup.py bdist_wheel --build-variant=rocm
+      python setup.py bdist_wheel \
+        --build-variant=rocm \
+        -DHIP_ROOT_DIR="${ROCM_PATH}" \
+        -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
+        -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
       popd
 
       # Save the wheel before cleaning up

.ci/pytorch/macos-test.sh

@@ -199,7 +199,7 @@ torchbench_setup_macos() {
   git checkout "$(cat ../.github/ci_commit_pins/vision.txt)"
   git submodule update --init --recursive
   python setup.py clean
-  python -m pip install -e . -v --no-build-isolation
+  python setup.py develop
   popd
 
   pushd torchaudio
@@ -208,7 +208,7 @@ torchbench_setup_macos() {
   git submodule update --init --recursive
   python setup.py clean
   #TODO: Remove me, when figure out how to make TorchAudio find brew installed openmp
-  USE_OPENMP=0 python -m pip install -e . -v --no-build-isolation
+  USE_OPENMP=0 python setup.py develop
   popd
 
   checkout_install_torchbench

.ci/pytorch/smoke_test/smoke_test.py

@@ -386,8 +386,8 @@ def smoke_test_compile(device: str = "cpu") -> None:
 
 
 def smoke_test_nvshmem() -> None:
-    if not torch.cuda.is_available() or target_os == "windows":
-        print("Windows platform or CUDA is not available, skipping NVSHMEM test")
+    if not torch.cuda.is_available():
+        print("CUDA is not available, skipping NVSHMEM test")
         return
 
     # Check if NVSHMEM is compiled in current build
@@ -396,9 +396,7 @@ def smoke_test_nvshmem() -> None:
     except ImportError:
         # Not built with NVSHMEM support.
         # torch is not compiled with NVSHMEM prior to 2.9
-        from torch.torch_version import TorchVersion
-
-        if TorchVersion(torch.__version__) < (2, 9):
+        if torch.__version__ < "2.9":
             return
         else:
             # After 2.9: NVSHMEM is expected to be compiled in current build

.ci/pytorch/test.sh

@@ -1721,6 +1721,11 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   test_inductor_shard "${SHARD_NUMBER}"
+  if [[ "${SHARD_NUMBER}" == 1 ]]; then
+    if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then
+      test_inductor_distributed
+    fi
+  fi
 elif [[ "${TEST_CONFIG}" == *einops* ]]; then
   test_einops
 elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then

.ci/pytorch/windows/internal/copy.bat

@@ -1,20 +1,12 @@
-
-if %CUDA_VERSION% geq 130 (
-    set "dll_path=bin\x64"
-) else (
-    set "dll_path=bin"
-)
-
-copy "%CUDA_PATH%\%dll_path%\cusparse*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cublas*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cudart*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\curand*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cufft*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cusolver*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\nvrtc*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\nvJitLink_*.dll*"  pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cusparse*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cublas*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cudart*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\curand*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cufft*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cusolver*64_*.dll*" pytorch\torch\lib
 
 copy "%CUDA_PATH%\bin\cudnn*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\nvrtc*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\cupti64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\nvperf_host*.dll*" pytorch\torch\lib
 
@@ -28,3 +20,8 @@ copy "%libuv_ROOT%\bin\uv.dll" pytorch\torch\lib
 if exist "C:\Windows\System32\zlibwapi.dll" (
     copy "C:\Windows\System32\zlibwapi.dll"  pytorch\torch\lib
 )
+
+::copy nvJitLink dll is requires for cuda 12+
+if exist "%CUDA_PATH%\bin\nvJitLink_*.dll*" (
+    copy "%CUDA_PATH%\bin\nvJitLink_*.dll*"  pytorch\torch\lib
+)

.ci/pytorch/windows/internal/driver_update.bat

@@ -1,9 +1,9 @@
-set WIN_DRIVER_VN=580.88
-set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe" & REM @lint-ignore
-curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe
+set WIN_DRIVER_VN=528.89
+set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe" & REM @lint-ignore
+curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe
 if errorlevel 1 exit /b 1
 
-start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe -s -noreboot
+start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe -s -noreboot
 if errorlevel 1 exit /b 1
 
-del %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe || ver > NUL
+del %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe || ver > NUL

.ci/wheel/build_wheel.sh

@@ -85,7 +85,7 @@ mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
 # Create an isolated directory to store this builds pytorch checkout and conda
 # installation
 if [[ -z "$MAC_PACKAGE_WORK_DIR" ]]; then
-    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_${DESIRED_PYTHON}_$(date +%H%M%S)"
+    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_conda_${DESIRED_PYTHON}_$(date +%H%M%S)"
 fi
 mkdir -p "$MAC_PACKAGE_WORK_DIR" || true
 if [[ -n ${GITHUB_ACTIONS} ]]; then
@@ -96,11 +96,11 @@ fi
 whl_tmp_dir="${MAC_PACKAGE_WORK_DIR}/dist"
 mkdir -p "$whl_tmp_dir"
 
-mac_version='macosx-11_0-arm64'
+mac_version='macosx_11_0_arm64'
 libtorch_arch='arm64'
 
 # Create a consistent wheel package name to rename the wheel to
-wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version//[-,]/_}.whl"
+wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version}.whl"
 
 ###########################################################
 
@@ -124,57 +124,93 @@ popd
 
 export TH_BINARY_BUILD=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
-export MACOSX_DEPLOYMENT_TARGET=11.0
+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"
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
 RENAME_WHEEL=true
 case $desired_python in
     3.14t)
         echo "Using 3.14 deps"
-        mac_version='macosx-11.0-arm64'
+        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"
-        mac_version='macosx-11.0-arm64'
+        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"
+        CONDA_ENV_CREATE_FLAGS="python-freethreading"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+        desired_python="3.13"
         RENAME_WHEEL=false
         ;;
     3.13)
         echo "Using 3.13 deps"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
         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"
         ;;
     3.11)
         echo "Using 3.11 deps"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
         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"
+        ;;
+    3.9)
+        echo "Using 3.9 deps"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
         NUMPY_PINNED_VERSION="==2.0.2"
         ;;
     *)
-        echo "Unsupported version $desired_python"
-        exit 1
+        echo "Using default deps"
+        NUMPY_PINNED_VERSION="==1.11.3"
         ;;
 esac
 
+# Install into a fresh env
+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${NUMPY_PINNED_VERSION}"
 )
-python -mvenv ~/${desired_python}-build
-source ~/${desired_python}-build/bin/activate
-retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements.txt"
+retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements-build.txt"
+pip install requests ninja typing-extensions
+retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
 retry brew install libomp
 
 # For USE_DISTRIBUTED=1 on macOS, this enables gloo, which needs libuv, which
@@ -188,7 +224,7 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-_PYTHON_HOST_PLATFORM=${mac_version} ARCHFLAGS="-arch arm64" python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name "${mac_version//[-.]/_}"
+python setup.py bdist_wheel -d "$whl_tmp_dir"
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.flake8

@@ -73,7 +73,7 @@ exclude =
     ./docs/src,
     ./functorch/docs,
     ./functorch/examples,
-    ./functorch/docs/source/tutorials,
+    ./functorch/notebooks,
     ./scripts,
     ./test/generated_type_hints_smoketest.py,
     ./third_party,

.github/actionlint.yaml

@@ -12,16 +12,13 @@ self-hosted-runner:
     - linux.9xlarge.ephemeral
     - am2.linux.9xlarge.ephemeral
     - linux.12xlarge
-    - linux.12xlarge.memory
     - linux.24xlarge
-    - linux.24xlarge.memory
     - linux.24xlarge.ephemeral
     - linux.24xlarge.amd
     - linux.arm64.2xlarge
     - linux.arm64.2xlarge.ephemeral
     - linux.arm64.m7g.4xlarge
     - linux.arm64.m7g.4xlarge.ephemeral
-    - linux.arm64.r7g.12xlarge.memory
     - linux.4xlarge.nvidia.gpu
     - linux.8xlarge.nvidia.gpu
     - linux.16xlarge.nvidia.gpu

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

@@ -4,11 +4,6 @@ name: Build External packages
 description: build external packages for PyTorch
 
 inputs:
-  cuda-version:
-    description: CUDA version to use
-    type: string
-    required: true
-    default: '12.8.1'
   cuda-arch-list:
     description: TORCH_CUDA_ARCH_LIST (e.g., "8.0;8.9;9.0")
     type: string
@@ -49,12 +44,11 @@ runs:
       env:
         SCCACHE_BUCKET: ossci-compiler-cache-circleci-v2
         SCCACHE_REGION: us-east-1
-        CUDA_VERSION: ${{ inputs.cuda-version }}
         TORCH_CUDA_ARCH_LIST: ${{ inputs.cuda-arch-list }}
         BASE_IMAGE: ${{ inputs.docker-image }}
         BUILD_TARGETS: ${{ inputs.build-targets }}
-        PARENT_OUTPUT_DIR: ${{ inputs.output-dir }}
-        TORCH_WHEELS_PATH: ${{ inputs.torch-wheel-dir }}
+        PARENT_OUTPUT_DIR: ${{ inputs.output-dir}}
+
       shell: bash
       run: |
         set -euo pipefail
@@ -75,6 +69,7 @@ runs:
           export OUTPUT_DIR
           echo "Building external package: $target in directory $OUTPUT_DIR"
           python3 -m cli.run build external "$target"
+
         done
 
         END_TIME=$(date +%s)

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+0757bbb660855272f7dd8d31cc84e7c631522805

.github/ci_commit_pins/fbgemm_rocm.txt

@@ -1 +1 @@
-08ae0af1395c8d8471f4025deb6af9aef90b342f
+7f1de94a4c2d14f59ad4ca84538c36084ea6b2c8

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-973c9d01da863cac9c51e8a5c0d390fc84b84fbc
+862f2ef893d9751db0a92bd2d4ae0e3d9677872f

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-c77852e117bdf056c8e9a087e51d6f65cf6ba53d
+763e5b78d4fcd74a9e812256656c075f99d9a781

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -12,46 +12,54 @@ ARG BUILD_BASE_IMAGE=torch-nightly-base
 # by default, it uses devel-ubuntu22.04 official image.
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
-# The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
-ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
 
-
-#################### TORCH NIGHTLY BASE IMAGE ####################
+#################### TORCH NIGHTLY  BASE IMAGE ####################
 # A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
+From nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
+ARG CUDA_VERSION=12.8.1
+ARG PYTHON_VERSION=3.12
+ARG TARGETPLATFORM
+ENV DEBIAN_FRONTEND=noninteractive
 
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-ARG GET_PIP_URL
+RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
+    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
 
-# Install Python and other dependencies
-RUN apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-    && add-apt-repository -y ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
-    && python3 --version && python3 -m pip --version
+# Install Python and other dependencies if it does not existed
+RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
+      echo "Installing Python ${PYTHON_VERSION}..." && \
+      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
+      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
+      apt-get update -y && \
+      apt-get install -y ccache software-properties-common git curl sudo && \
+      for i in 1 2 3; do \
+        add-apt-repository -y ppa:deadsnakes/ppa && break || \
+        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+      done && \
+      apt-get update -y && \
+      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
+      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
+      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
+      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
+      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
+   else \
+      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
+   fi \
+   && python3 --version && python3 -m pip --version
 
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
 # Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
 RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if command -v apt-get >/dev/null; then \
-        if [ "$current_gcc_version" -lt 10 ]; then \
-            echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-            apt-get update \
-            && apt-get install -y gcc-10 g++-10 \
-            && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
-            && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-        else \
-            echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-        fi \
-    fi \
-    && gcc --version && g++ --version
+    if [ "$current_gcc_version" -lt 10 ]; then \
+      echo "GCC version is $current_gcc_version, installing gcc-10..."; \
+      apt-get update && \
+      apt-get install -y gcc-10 g++-10 && \
+      update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 && \
+      update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
+    else \
+      echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
+    fi && \
+    gcc --version && g++ --version
 
 # install uv for faster pip installs
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -71,20 +79,11 @@ ENV UV_LINK_MODE=copy
 FROM ${BUILD_BASE_IMAGE} AS base
 USER root
 
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
-# Install some system dependencies and double check python version
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
-    else \
-        dnf install -y git curl wget sudo; \
-    fi \
-    && python3 --version && python3 -m pip --version
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -119,15 +118,17 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     if [ -n "$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=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
-        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
+        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
+        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
+        uv pip install --system "${torch_whl}[opt-einsum]"; \
+        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: $PINNED_TORCH_VERSION"; \
-        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu128; \
     else \
         echo "[INFO] Installing torch nightly with latest one to build vllm"; \
-        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128; \
     fi
 
 # Install numba 0.61.2 for cuda environment
@@ -136,11 +137,12 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Install common dependencies from vllm common.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
+uv pip install --system -r requirements/common.txt
+
 
 # Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}
+ARG exformer_cuda_arch_list='7.5;8.0+PTX;9.0a'
+ENV TORCH_CUDA_ARCH_LIST=${exformer_cuda_arch_list}
 
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
@@ -151,8 +153,8 @@ 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/v0.0.32.post2
-ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
+# sha for https://github.com/facebookresearch/xformers/tree/v0.0.31
+ARG XFORMERS_COMMIT=eb0946a363464da96ea40afd1a7f72a907c25497
 ENV CCACHE_DIR=/root/.cache/ccache
 
 RUN --mount=type=cache,target=/root/.cache/ccache \
@@ -186,6 +188,11 @@ RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
 FROM base AS build
 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 . .
 
 RUN python3 use_existing_torch.py
@@ -214,16 +221,11 @@ ARG SCCACHE_S3_NO_CREDENTIALS=0
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
-        echo "Installing sccache..."; \
-        if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-            SCCACHE_ARCHIVE="sccache-v0.8.1-aarch64-unknown-linux-musl"; \
-        else \
-            SCCACHE_ARCHIVE="sccache-v0.8.1-x86_64-unknown-linux-musl"; \
-        fi; \
-        curl -L -o sccache.tar.gz "https://github.com/mozilla/sccache/releases/download/v0.8.1/${SCCACHE_ARCHIVE}.tar.gz" \
+        echo "Installing sccache..." \
+        && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
         && tar -xzf sccache.tar.gz \
-        && sudo mv "${SCCACHE_ARCHIVE}"/sccache /usr/bin/sccache \
-        && rm -rf sccache.tar.gz "${SCCACHE_ARCHIVE}" \
+        && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
+        && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
         && export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
         && export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
         && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
@@ -249,9 +251,9 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
     fi
 
-RUN echo "[INFO] Listing current directory:" && \
+RUN echo "[DEBUG] Listing  current directory:" && \
     ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
+    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
 #################### WHEEL BUILD IMAGE ####################
@@ -261,42 +263,51 @@ RUN echo "[INFO] Listing current directory:" && \
 # Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 USER root
-
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-ARG GET_PIP_URL
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
 # prepare for environment starts
 WORKDIR /workspace
 
-# Install Python and other dependencies
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-        && add-apt-repository -y ppa:deadsnakes/ppa \
-        && apt-get update -y \
-        && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-        && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-        && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-        && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-        && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
-    else \
-        dnf install -y git curl wget sudo; \
-    fi \
-    && python3 --version && python3 -m pip --version
+RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
+    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
+
+# Install Python and other dependencies if it does not existed
+RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
+      echo "Installing Python ${PYTHON_VERSION}..." && \
+      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
+      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
+      apt-get update -y && \
+      apt-get install -y ccache software-properties-common git curl sudo && \
+      for i in 1 2 3; do \
+        add-apt-repository -y ppa:deadsnakes/ppa && break || \
+        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+      done && \
+      apt-get update -y && \
+      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
+      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
+      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
+      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
+      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
+   else \
+      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
+   fi \
+   && python3 --version && python3 -m pip --version
+
 
 # Get the torch versions, and whls used in previous stagtes for consistency
 COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
-RUN echo "[INFO] Listing current directory before torch install step:" && \
+RUN echo "[DEBUG] Listing current directory before torch install step:" && \
     ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
+    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
+
+
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version > /dev/null 2>&1; then \
@@ -316,13 +327,15 @@ RUN --mount=type=bind,source=${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 \
         torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
+        vision_whl=$(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}'"; \
-        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
+        uv pip install --system "${torch_whl}[opt-einsum]"; \
+        uv pip install --system "${vision_whl}"; \
+        uv pip install --system "${audio_whl}"; \
     else \
         echo "[INFO] Installing torch versions from torch_build_versions.txt"; \
-        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu128; \
     fi
 
 # Install the vllm wheel from previous stage
@@ -333,8 +346,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
 
+
 # Build flashinfer from source.
-ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
+ARG torch_cuda_arch_list='8.0;8.9;9.0a'
 # install package for build flashinfer
 # see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
 
@@ -402,6 +416,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/nightly_torch_test.txt
 
+# Workaround for #17068
+# pinned commit for v2.2.4
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@95d8aba8a8c75aedcaa6143713b11e745e7cd0d9#egg=mamba-ssm"
+
 # Logging to confirm the torch versions
 RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
 

.github/ci_configs/vllm/use_existing_torch.py

@@ -1,17 +0,0 @@
-import glob
-
-
-requires_files = glob.glob("requirements/*.txt")
-requires_files += ["pyproject.toml"]
-for file in requires_files:
-    print(f">>> cleaning {file}")
-    with open(file) as f:
-        lines = f.readlines()
-    if "torch" in "".join(lines).lower():
-        print("removed:")
-        with open(file, "w") as f:
-            for line in lines:
-                if "torch" not in line.lower():
-                    f.write(line)
-    print(f"<<< done cleaning {file}")
-    print()

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

@@ -15,7 +15,7 @@ optree==0.13.0
 packaging==23.1
 parameterized==0.8.1
 pillow==10.3.0
-protobuf==5.29.5
+protobuf==5.29.4
 psutil==5.9.8
 pygments==2.15.0
 pytest-cpp==2.3.0
@@ -26,7 +26,7 @@ pytest-xdist==3.3.1
 pytest==7.3.2
 pyyaml==6.0.2
 scipy==1.12.0
-setuptools==78.1.1
+setuptools==72.1.0
 sympy==1.13.3
 tlparse==0.4.0
 tensorboard==2.13.0

.github/scripts/build_triton_wheel.py

@@ -84,7 +84,6 @@ def build_triton(
                 ["git", "checkout", f"release/{ver}.{rev}.x"], cwd=triton_basedir
             )
         else:
-            check_call(["git", "fetch", "origin", commit_hash], cwd=triton_basedir)
             check_call(["git", "checkout", commit_hash], cwd=triton_basedir)
 
         # change built wheel name and version

.github/scripts/docathon-label-sync.py

@@ -39,9 +39,7 @@ def main() -> None:
     pull_request_label_names = [label.name for label in pull_request_labels]
     issue_label_names = [label.name for label in issue_labels]
     labels_to_add = [
-        label
-        for label in issue_label_names
-        if label not in pull_request_label_names and label != "actionable"
+        label for label in issue_label_names if label not in pull_request_label_names
     ]
     if not labels_to_add:
         print("The pull request already has the same labels.")

.github/scripts/generate_binary_build_matrix.py

@@ -16,16 +16,18 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
+    "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",
 }
 
@@ -38,60 +40,77 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.9-aarch64", "13.0-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
     "12.6": (
-        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "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-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'"
     ),
     "12.8": (
-        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
+        "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'"
+    ),
+    "12.9": (
+        "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "13.0": (
-        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | "
-        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' | "
-        "nvidia-curand==10.4.0.35; platform_system == 'Linux' | "
-        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
-        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
-        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
-        "nvidia-cufile==1.15.0.42; platform_system == 'Linux'"
+        "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 | "
@@ -221,6 +240,8 @@ 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 libtorch_variants is None:
@@ -322,7 +343,7 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.8", "12.6"]
+                arch_version in ["13.0", "12.9", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -386,5 +407,6 @@ def generate_wheels_matrix(
 
 
 validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
 validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")

.github/scripts/generate_ci_workflows.py

@@ -22,7 +22,7 @@ LABEL_CIFLOW_BINARIES = "ciflow/binaries"
 LABEL_CIFLOW_PERIODIC = "ciflow/periodic"
 LABEL_CIFLOW_BINARIES_LIBTORCH = "ciflow/binaries_libtorch"
 LABEL_CIFLOW_BINARIES_WHEEL = "ciflow/binaries_wheel"
-LABEL_CIFLOW_ROCM = "ciflow/rocm"
+LABEL_CIFLOW_ROCM = "ciflow/rocm-mi300"
 
 
 @dataclass
@@ -139,8 +139,6 @@ ROCM_SMOKE_WORKFLOWS = [
         ),
         ciflow_config=CIFlowConfig(
             labels={
-                LABEL_CIFLOW_BINARIES,
-                LABEL_CIFLOW_BINARIES_WHEEL,
                 LABEL_CIFLOW_ROCM,
             },
             isolated_workflow=True,

.github/scripts/prepare_vllm_wheels.sh

@@ -1,94 +0,0 @@
-#!/usr/bin/env bash
-
-set -eux
-
-torch_version=$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-nightly=$(echo ${torch_version} | cut -d'.' -f4)
-
-# Copied from .ci/manywheel/build_common.sh
-make_wheel_record() {
-  fpath=$1
-  if echo $fpath | grep RECORD >/dev/null 2>&1; then
-    echo "$fpath,,"
-  else
-    fhash=$(openssl dgst -sha256 -binary $fpath | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
-    fsize=$(ls -nl $fpath | awk '{print $5}')
-    echo "$fpath,sha256=$fhash,$fsize"
-  fi
-}
-
-change_wheel_version() {
-  local package=$1
-  local wheel=$2
-  local f_version=$3
-  local t_version=$4
-
-  # Extract the wheel
-  ${PYTHON_EXECUTABLE} -mwheel unpack $wheel
-
-  mv "${package}-${f_version}" "${package}-${t_version}"
-  # Change the version from f_version to t_version in the dist-info dir
-  pushd "${package}-${t_version}"
-  mv "${package}-${f_version}.dist-info" "${package}-${t_version}.dist-info"
-
-  pushd "${package}-${t_version}.dist-info"
-  sed -i "s/${package}-${f_version}.dist-info/${package}-${t_version}.dist-info/g" RECORD
-
-  # Update the version in METADATA and its SHA256 hash
-  sed -i "s/Version: ${f_version}/Version: ${t_version}/g" METADATA
-  # then add PyTorch nightly dependency of vLLM
-  if [[ "${package}" == vllm ]] || [[ "${package}" == xformers ]]; then
-    sed -i "/License-File/a\Requires-Dist: torch==${torch_version}" METADATA
-  fi
-  sed -i '/METADATA,sha256/d' RECORD
-  popd
-
-  make_wheel_record "${package}-${t_version}.dist-info/METADATA" >> "${package}-${t_version}.dist-info/RECORD"
-  popd
-
-  # Repack the wheel
-  ${PYTHON_EXECUTABLE} -mwheel pack "${package}-${t_version}"
-
-  # Clean up
-  rm -rf "${package}-${t_version}"
-}
-
-repackage_wheel() {
-  local package=$1
-  pushd $package
-
-  local orig_wheel=$(find . -name *${package//-/_}*)
-  local orig_version=$(unzip -p $orig_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-
-  local version=""
-  if [[ "${package}" == vllm ]]; then
-    # Copied from vllm/.buildkite/scripts/upload-wheels.sh
-    version=1.0.0
-  else
-    version=$(echo $orig_version | tr '.+' '.' | cut -d'.' -f1-3)
-  fi
-  local nightly_version=$version.$nightly
-
-  # Use nightly version
-  change_wheel_version ${package//-/_} $orig_wheel $orig_version $nightly_version
-  # Clean up
-  rm "${orig_wheel}"
-
-  auditwheel repair --plat $PLATFORM *.whl \
-    --exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
-  local repair_wheel=$(find wheelhouse -name *${PLATFORM}*)
-  local repair_wheel=$(basename ${repair_wheel})
-  popd
-
-  cp ${package}/wheelhouse/${repair_wheel} .
-  rm -rf $package
-}
-
-# Require to re-package the wheel
-${PYTHON_EXECUTABLE} -mpip install wheel==0.45.1
-
-pushd externals/vllm/wheels
-for package in xformers flashinfer-python vllm; do
-  repackage_wheel $package
-done
-popd

.github/templates/linux_binary_build_workflow.yml.j2

@@ -171,7 +171,7 @@ jobs:
       - name: Teardown XPU
         uses: ./.github/actions/teardown-xpu
     {%- else %}
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: !{{ common.timeout_minutes }}
     !{{ upload.binary_env(config) }}
     steps:

.github/templates/macos_binary_build_workflow.yml.j2

@@ -22,16 +22,6 @@ name: !{{ build_environment }}
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
 {%- endmacro %}
 
-{%- macro setup_python(py_ver) -%}
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "!{{ (py_ver.strip('t') + '.4') if '3.14' not in py_ver else '3.14.0-rc.2' }}"
-          freethreaded: !{{ "true" if py_ver.endswith('t') else "false" }}
-{%- endmacro %}
-
 on:
 # TODO: Migrate to new ciflow trigger, reference https://github.com/pytorch/pytorch/pull/70321
   push:
@@ -71,13 +61,28 @@ jobs:
     {%- endif %}
     steps:
       !{{ set_runner_specific_vars() }}
-      !{{ setup_python(config.get("python_version", "3.10")) }}
+      - 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" >> "${GITHUB_ENV}"
+          fi
       !{{ common.checkout(deep_clone=False, directory="pytorch") }}
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -94,6 +99,8 @@ jobs:
 {%- if config["package_type"] == "wheel" %}
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -104,9 +111,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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

.github/templates/upload.yml.j2

@@ -33,7 +33,7 @@
   {%- if is_windows %}
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
   {%- endif %}
 
 {%- else %}

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

@@ -47,11 +47,12 @@ jobs:
       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: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm6.3",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cpu",               runner: "linux.9xlarge.ephemeral" },

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

@@ -1,236 +0,0 @@
-name: Build vLLM wheels
-
-on:
-  push:
-    branches:
-      - main
-    paths:
-      - .github/workflows/build-vllm-wheel.yml
-      - .github/ci_commit_pins/vllm.txt
-  workflow_dispatch:
-  pull_request:
-    paths:
-      - .github/workflows/build-vllm-wheel.yml
-      - .github/ci_commit_pins/vllm.txt
-  schedule:
-    # every morning at 01:30PM UTC, 9:30AM EST, 6:30AM PST
-    - cron: 30 13 * * *
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
-  cancel-in-progress: true
-
-jobs:
-  build-wheel:
-    if: github.repository_owner == 'pytorch'
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: [ '3.12' ]
-        # TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
-        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129' ]
-        include:
-          - platform: manylinux_2_28_x86_64
-            device: cu128
-            manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.8'
-            runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_x86_64
-            device: cu129
-            manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
-            runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_aarch64
-            device: cu128
-            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
-            runner: linux.arm64.r7g.12xlarge.memory
-          - platform: manylinux_2_28_aarch64
-            device: cu129
-            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
-            runner: linux.arm64.r7g.12xlarge.memory
-    name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
-    runs-on: ${{ matrix.runner }}
-    timeout-minutes: 480
-    env:
-      PY_VERS: ${{ matrix.python-version }}
-      MANYLINUX_IMAGE: ${{ matrix.manylinux-image }}
-      PLATFORM: ${{ matrix.platform }}
-      BUILD_DEVICE: ${{ matrix.device }}
-    steps:
-      - name: Setup SSH (Click me for login details)
-        uses: pytorch/test-infra/.github/actions/setup-ssh@main
-        with:
-          github-secret: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Checkout PyTorch
-        uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
-        with:
-          submodules: false
-
-      - name: Setup Linux
-        uses: ./.github/actions/setup-linux
-
-      - name: Get latest PyTorch nightly
-        shell: bash
-        run: |
-          set -eux
-
-          # Determine python executable for given version (copied from build-triton-wheel)
-          case $PY_VERS in
-          3.10)
-            PYTHON_EXECUTABLE=/opt/python/cp310-cp310/bin/python
-            ;;
-          3.11)
-            PYTHON_EXECUTABLE=/opt/python/cp311-cp311/bin/python
-            ;;
-          3.12)
-            PYTHON_EXECUTABLE=/opt/python/cp312-cp312/bin/python
-            ;;
-          3.13)
-            PYTHON_EXECUTABLE=/opt/python/cp313-cp313/bin/python
-            ;;
-          3.13t)
-            PYTHON_EXECUTABLE=/opt/python/cp313-cp313t/bin/python
-            ;;
-          3.14)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314/bin/python
-            ;;
-          3.14t)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314t/bin/python
-            ;;
-          *)
-            echo "Unsupported python version ${PY_VERS}"
-            exit 1
-            ;;
-          esac
-
-          # Keep PyTorch nightly wheel here so that we can install it later during
-          # vLLM build process
-          mkdir -p "${RUNNER_TEMP}/artifacts/"
-
-          container_name=$(docker run \
-            --tty \
-            --detach \
-            -e PLATFORM \
-            -e PYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" \
-            -v "${GITHUB_WORKSPACE}:/pytorch" \
-            -v "${RUNNER_TEMP}/artifacts:/artifacts" \
-            -w /artifacts/ \
-            "${MANYLINUX_IMAGE}"
-          )
-
-          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip install \
-            --pre torch torchvision torchaudio \
-            --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
-
-          # I wonder if there is a command to both download and install the wheels
-          # in one go
-          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip download \
-            --pre torch torchvision torchaudio \
-            --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
-
-          # Save this for later
-          echo "container_name=${container_name}" >> "$GITHUB_ENV"
-
-      - name: Build vLLM wheel
-        uses: ./.github/actions/build-external-packages
-        with:
-          build-targets: vllm
-          docker-image: ${{ env.MANYLINUX_IMAGE }}
-          cuda-arch-list: '8.0;8.9;9.0;10.0;12.0'
-          torch-wheel-dir: ${{ runner.temp }}/artifacts
-          output-dir: ${{ runner.temp }}/artifacts/externals
-
-      - name: Prepare vLLM wheel
-        shell: bash
-        run: |
-          set -eux
-
-          # Get these wheels ready, the vllm renaming logic is copied from its .buildkite/scripts/upload-wheels.sh
-          docker exec -t "${container_name}" bash -c /pytorch/.github/scripts/prepare_vllm_wheels.sh
-          docker exec -t "${container_name}" chown -R 1000:1000 /artifacts
-
-      - uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
-        with:
-          name: vllm-wheel-${{ matrix.device }}-${{ matrix.platform }}-${{ matrix.python-version }}
-          if-no-files-found: error
-          path: ${{ runner.temp }}/artifacts/externals/vllm/wheels/*.whl
-
-      - name: Teardown Linux
-        uses: pytorch/test-infra/.github/actions/teardown-linux@main
-        if: always()
-
-  # Copied from build-triton-wheel workflow (mostly)
-  upload-wheel:
-    name: "Upload ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
-    needs:
-      - build-wheel
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129' ]
-    env:
-      PLATFORM: ${{ matrix.platform }}
-      BUILD_DEVICE: ${{ matrix.device }}
-    permissions:
-      id-token: write
-      contents: read
-    container:
-      image: continuumio/miniconda3:4.12.0
-    environment: ${{ (github.event_name == 'push' && github.event.ref == 'refs/heads/main') && 'nightly-wheel-upload' || '' }}
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Configure AWS credentials(PyTorch account) for main
-        if: ${{ github.event_name == 'push' && github.event.ref == 'refs/heads/main' }}
-        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-        with:
-          role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_nightly_build_wheels
-          aws-region: us-east-1
-
-      - name: Configure AWS credentials(PyTorch account) for RC builds
-        if: ${{ github.event_name == 'push' &&  (startsWith(github.event.ref, 'refs/tags/') && !startsWith(github.event.ref, 'refs/tags/ciflow/')) }}
-        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-        with:
-          role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_test_build_wheels
-          aws-region: us-east-1
-
-      - name: Download Build Artifacts
-        uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
-        with:
-          # Download all available artifacts
-          path: ${{ runner.temp }}/artifacts-all
-
-      - name: Select Wheel Artifacts
-        shell: bash
-        run: |
-          set -eux
-          mkdir -p "${RUNNER_TEMP}/artifacts/"
-          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-"${PLATFORM}"-*/* "${RUNNER_TEMP}/artifacts/"
-
-      - name: Set DRY_RUN
-        if: ${{ (github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v'))) || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
-        shell: bash
-        run: |
-          echo "DRY_RUN=disabled" >> "$GITHUB_ENV"
-
-      - name: Set UPLOAD_CHANNEL
-        if: ${{ github.event_name == 'push' && startsWith(github.event.ref, 'refs/tags/v') }}
-        shell: bash
-        run: |
-          set -ex
-
-          if [[ "${GITHUB_REF_NAME}" = *-rc[0-9]* ]]; then
-            echo "UPLOAD_CHANNEL=test" >> "$GITHUB_ENV"
-          fi
-
-      - name: Upload binaries
-        env:
-          PACKAGE_TYPE: wheel
-          UPLOAD_SUBFOLDER: ${{ env.BUILD_DEVICE }}
-          PKG_DIR: ${{ runner.temp }}/artifacts
-        shell: bash
-        run: |
-          set -ex
-          bash .circleci/scripts/binary_upload.sh

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

@@ -112,7 +112,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_10-cuda-aarch64-12_6-build:
+  manywheel-py3_10-cuda-aarch64-12_9-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -121,85 +121,39 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_10-cuda-aarch64-12_6
+      build_name: manywheel-py3_10-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; 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_10-cuda-aarch64-12_6-upload:  # Uploading
+  manywheel-py3_10-cuda-aarch64-12_9-upload:  # Uploading
     if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       id-token: write
       contents: read
-    needs: manywheel-py3_10-cuda-aarch64-12_6-build
+    needs: manywheel-py3_10-cuda-aarch64-12_9-build
     with:
       PYTORCH_ROOT: /pytorch
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.10"
-      build_name: manywheel-py3_10-cuda-aarch64-12_6
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_10-cuda-aarch64-12_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.10"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_10-cuda-aarch64-12_8
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_10-cuda-aarch64-12_8-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_10-cuda-aarch64-12_8-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: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.10"
-      build_name: manywheel-py3_10-cuda-aarch64-12_8
+      build_name: manywheel-py3_10-cuda-aarch64-12_9
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -224,7 +178,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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 }}
@@ -315,7 +269,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_11-cuda-aarch64-12_6-build:
+  manywheel-py3_11-cuda-aarch64-12_9-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -324,85 +278,39 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_11-cuda-aarch64-12_6
+      build_name: manywheel-py3_11-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; 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_11-cuda-aarch64-12_6-upload:  # Uploading
+  manywheel-py3_11-cuda-aarch64-12_9-upload:  # Uploading
     if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       id-token: write
       contents: read
-    needs: manywheel-py3_11-cuda-aarch64-12_6-build
+    needs: manywheel-py3_11-cuda-aarch64-12_9-build
     with:
       PYTORCH_ROOT: /pytorch
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.11"
-      build_name: manywheel-py3_11-cuda-aarch64-12_6
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_11-cuda-aarch64-12_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.11"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_11-cuda-aarch64-12_8
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_11-cuda-aarch64-12_8-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_11-cuda-aarch64-12_8-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: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.11"
-      build_name: manywheel-py3_11-cuda-aarch64-12_8
+      build_name: manywheel-py3_11-cuda-aarch64-12_9
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -427,7 +335,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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 }}
@@ -518,7 +426,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_12-cuda-aarch64-12_6-build:
+  manywheel-py3_12-cuda-aarch64-12_9-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -527,85 +435,39 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_12-cuda-aarch64-12_6
+      build_name: manywheel-py3_12-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; 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_12-cuda-aarch64-12_6-upload:  # Uploading
+  manywheel-py3_12-cuda-aarch64-12_9-upload:  # Uploading
     if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       id-token: write
       contents: read
-    needs: manywheel-py3_12-cuda-aarch64-12_6-build
+    needs: manywheel-py3_12-cuda-aarch64-12_9-build
     with:
       PYTORCH_ROOT: /pytorch
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda-aarch64-12_6
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_12-cuda-aarch64-12_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.12"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_12-cuda-aarch64-12_8
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda-aarch64-12_8-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_12-cuda-aarch64-12_8-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: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda-aarch64-12_8
+      build_name: manywheel-py3_12-cuda-aarch64-12_9
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -630,7 +492,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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 }}
@@ -721,7 +583,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_13-cuda-aarch64-12_6-build:
+  manywheel-py3_13-cuda-aarch64-12_9-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -730,85 +592,39 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13-cuda-aarch64-12_6
+      build_name: manywheel-py3_13-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; 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_13-cuda-aarch64-12_6-upload:  # Uploading
+  manywheel-py3_13-cuda-aarch64-12_9-upload:  # Uploading
     if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       id-token: write
       contents: read
-    needs: manywheel-py3_13-cuda-aarch64-12_6-build
+    needs: manywheel-py3_13-cuda-aarch64-12_9-build
     with:
       PYTORCH_ROOT: /pytorch
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.13"
-      build_name: manywheel-py3_13-cuda-aarch64-12_6
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_13-cuda-aarch64-12_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.13"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13-cuda-aarch64-12_8
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13-cuda-aarch64-12_8-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_13-cuda-aarch64-12_8-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: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.13"
-      build_name: manywheel-py3_13-cuda-aarch64-12_8
+      build_name: manywheel-py3_13-cuda-aarch64-12_9
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -833,7 +649,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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 }}
@@ -924,7 +740,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_13t-cuda-aarch64-12_6-build:
+  manywheel-py3_13t-cuda-aarch64-12_9-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -933,85 +749,39 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
       ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13t-cuda-aarch64-12_6
+      build_name: manywheel-py3_13t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; 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_13t-cuda-aarch64-12_6-upload:  # Uploading
+  manywheel-py3_13t-cuda-aarch64-12_9-upload:  # Uploading
     if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       id-token: write
       contents: read
-    needs: manywheel-py3_13t-cuda-aarch64-12_6-build
+    needs: manywheel-py3_13t-cuda-aarch64-12_9-build
     with:
       PYTORCH_ROOT: /pytorch
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.13t"
-      build_name: manywheel-py3_13t-cuda-aarch64-12_6
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_13t-cuda-aarch64-12_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.13t"
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.arm64.m7g.4xlarge.ephemeral
-      ALPINE_IMAGE: "arm64v8/alpine"
-      build_name: manywheel-py3_13t-cuda-aarch64-12_8
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_13t-cuda-aarch64-12_8-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_13t-cuda-aarch64-12_8-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: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.13t"
-      build_name: manywheel-py3_13t-cuda-aarch64-12_8
+      build_name: manywheel-py3_13t-cuda-aarch64-12_9
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -1036,7 +806,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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 }}
@@ -1127,7 +897,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_14-cuda-aarch64-12_6-build:
+  manywheel-py3_14-cuda-aarch64-12_9-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -1136,85 +906,39 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.14"
       runner_prefix: "${{ 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_6
+      build_name: manywheel-py3_14-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; 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_6-upload:  # Uploading
+  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_6-build
+    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: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cuda-aarch64-12_6
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_14-cuda-aarch64-12_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.14"
-      runner_prefix: "${{ 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_8
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14-cuda-aarch64-12_8-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14-cuda-aarch64-12_8-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: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.14"
-      build_name: manywheel-py3_14-cuda-aarch64-12_8
+      build_name: manywheel-py3_14-cuda-aarch64-12_9
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -1239,7 +963,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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 }}
@@ -1330,7 +1054,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  manywheel-py3_14t-cuda-aarch64-12_6-build:
+  manywheel-py3_14t-cuda-aarch64-12_9-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -1339,85 +1063,39 @@ jobs:
       PACKAGE_TYPE: manywheel
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.14t"
       runner_prefix: "${{ 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_6
+      build_name: manywheel-py3_14t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; 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_6-upload:  # Uploading
+  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_6-build
+    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: cu126
-      GPU_ARCH_VERSION: "12.6-aarch64"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cuda-aarch64-12_6
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
-  manywheel-py3_14t-cuda-aarch64-12_8-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: manywheel
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.14t"
-      runner_prefix: "${{ 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_8
-      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
-      timeout-minutes: 420
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_14t-cuda-aarch64-12_8-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: manywheel-py3_14t-cuda-aarch64-12_8-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: cu128
-      GPU_ARCH_VERSION: "12.8-aarch64"
-      GPU_ARCH_TYPE: cuda-aarch64
-      DOCKER_IMAGE: manylinuxaarch64-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.14t"
-      build_name: manywheel-py3_14t-cuda-aarch64-12_8
+      build_name: manywheel-py3_14t-cuda-aarch64-12_9
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -1442,7 +1120,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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 }}

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

@@ -248,7 +248,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  libtorch-cuda13_0-shared-with-deps-release-build:
+  libtorch-cuda12_9-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -257,22 +257,22 @@ jobs:
       PACKAGE_TYPE: libtorch
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: libtorch-cuda13_0-shared-with-deps-release
+      build_name: libtorch-cuda12_9-shared-with-deps-release
       build_environment: linux-binary-libtorch
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-cuda13_0-shared-with-deps-release-test:  # Testing
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
     if: ${{ github.repository_owner == 'pytorch' }}
     needs:
-      - libtorch-cuda13_0-shared-with-deps-release-build
+      - libtorch-cuda12_9-shared-with-deps-release-build
       - get-label-type
     uses: ./.github/workflows/_binary-test-linux.yml
     with:
@@ -280,38 +280,38 @@ jobs:
       PACKAGE_TYPE: libtorch
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-cuda13_0-shared-with-deps-release
+      build_name: libtorch-cuda12_9-shared-with-deps-release
       build_environment: linux-binary-libtorch
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-cuda13_0-shared-with-deps-release-upload:  # Uploading
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
     if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       id-token: write
       contents: read
-    needs: libtorch-cuda13_0-shared-with-deps-release-test
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
     with:
       PYTORCH_ROOT: /pytorch
       PACKAGE_TYPE: libtorch
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-cuda13_0-shared-with-deps-release
+      build_name: libtorch-cuda12_9-shared-with-deps-release
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
@@ -342,7 +342,7 @@ jobs:
     needs:
       - libtorch-rocm6_3-shared-with-deps-release-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -456,7 +456,7 @@ jobs:
     needs:
       - libtorch-rocm6_4-shared-with-deps-release-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch

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

@@ -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' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing

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

@@ -127,7 +127,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | 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-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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_6-test:  # Testing
@@ -193,7 +193,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_8-test:  # Testing
@@ -241,6 +241,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_10-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_10-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_10-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -259,7 +325,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda13_0-test:  # Testing
@@ -332,7 +398,7 @@ jobs:
     needs:
       - manywheel-py3_10-rocm6_3-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -443,7 +509,7 @@ jobs:
     needs:
       - manywheel-py3_10-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -719,7 +785,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | 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-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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_6-test:  # Testing
@@ -785,7 +851,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_8-test:  # Testing
@@ -833,6 +899,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_11-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_11-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_11-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_11-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -851,7 +983,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda13_0-test:  # Testing
@@ -924,7 +1056,7 @@ jobs:
     needs:
       - manywheel-py3_11-rocm6_3-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -1035,7 +1167,7 @@ jobs:
     needs:
       - manywheel-py3_11-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -1311,7 +1443,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | 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-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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_6-test:  # Testing
@@ -1377,7 +1509,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing
@@ -1425,6 +1557,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_12-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_12-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_12-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_12-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1443,7 +1641,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda13_0-test:  # Testing
@@ -1516,7 +1714,7 @@ jobs:
     needs:
       - manywheel-py3_12-rocm6_3-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -1627,7 +1825,7 @@ jobs:
     needs:
       - manywheel-py3_12-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -1903,7 +2101,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | 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-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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_6-test:  # Testing
@@ -1969,7 +2167,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_8-test:  # Testing
@@ -2017,6 +2215,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2035,7 +2299,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda13_0-test:  # Testing
@@ -2108,7 +2372,7 @@ jobs:
     needs:
       - manywheel-py3_13-rocm6_3-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -2219,7 +2483,7 @@ jobs:
     needs:
       - manywheel-py3_13-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -2495,7 +2759,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | 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-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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_6-test:  # Testing
@@ -2561,7 +2825,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_8-test:  # Testing
@@ -2609,6 +2873,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13t-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2627,7 +2957,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda13_0-test:  # Testing
@@ -2700,7 +3030,7 @@ jobs:
     needs:
       - manywheel-py3_13t-rocm6_3-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -2811,7 +3141,7 @@ jobs:
     needs:
       - manywheel-py3_13t-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -3087,7 +3417,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | 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-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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_6-test:  # Testing
@@ -3153,7 +3483,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_8-test:  # Testing
@@ -3201,6 +3531,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3219,7 +3615,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda13_0-test:  # Testing
@@ -3292,7 +3688,7 @@ jobs:
     needs:
       - manywheel-py3_14-rocm6_3-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -3403,7 +3799,7 @@ jobs:
     needs:
       - manywheel-py3_14-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -3679,7 +4075,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; 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==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | 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-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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_6-test:  # Testing
@@ -3745,7 +4141,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_8-test:  # Testing
@@ -3793,6 +4189,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14t-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' 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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3811,7 +4273,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda13_0-test:  # Testing
@@ -3884,7 +4346,7 @@ jobs:
     needs:
       - manywheel-py3_14t-rocm6_3-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch
@@ -3995,7 +4457,7 @@ jobs:
     needs:
       - manywheel-py3_14t-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch

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

@@ -10,9 +10,7 @@ on:
     branches:
       - main
     tags:
-      - 'ciflow/binaries/*'
-      - 'ciflow/binaries_wheel/*'
-      - 'ciflow/rocm/*'
+      - 'ciflow/rocm-mi300/*'
   workflow_dispatch:
 
 permissions:
@@ -69,7 +67,7 @@ jobs:
     needs:
       - manywheel-py3_9-rocm6_4-build
       - get-label-type
-    runs-on: linux.rocm.gpu.mi250
+    runs-on: linux.rocm.gpu.gfx942.1
     timeout-minutes: 240
     env:
       PYTORCH_ROOT: /pytorch

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

@@ -46,7 +46,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -60,13 +60,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.10.4"
-          freethreaded: false
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -81,9 +86,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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}"

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

@@ -56,13 +56,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.10.4"
-          freethreaded: false
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -77,9 +82,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -95,6 +104,8 @@ jobs:
           "${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}"
@@ -105,9 +116,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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
@@ -166,13 +201,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.11.4"
-          freethreaded: false
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -187,9 +227,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -205,6 +249,8 @@ jobs:
           "${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}"
@@ -215,9 +261,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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
@@ -276,13 +346,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.12.4"
-          freethreaded: false
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -297,9 +372,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -315,6 +394,8 @@ jobs:
           "${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}"
@@ -325,9 +406,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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
@@ -386,13 +491,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.13.4"
-          freethreaded: false
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -407,9 +517,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -425,6 +539,8 @@ jobs:
           "${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}"
@@ -435,9 +551,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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
@@ -496,13 +636,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.13.4"
-          freethreaded: true
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -517,9 +662,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -535,6 +684,8 @@ jobs:
           "${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}"
@@ -545,9 +696,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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
@@ -606,13 +781,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0-rc.2"
-          freethreaded: false
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -627,9 +807,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -645,6 +829,8 @@ jobs:
           "${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}"
@@ -655,9 +841,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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
@@ -716,13 +926,18 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0-rc.2"
-          freethreaded: true
+      - 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" >> "${GITHUB_ENV}"
+          fi
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -737,9 +952,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -755,6 +974,8 @@ jobs:
           "${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}"
@@ -765,9 +986,33 @@ jobs:
 
           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
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          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

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

@@ -64,7 +64,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Populate binary env
         shell: cmd
@@ -141,7 +141,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Populate binary env
         shell: cmd
@@ -201,7 +201,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cpu-shared-with-deps-debug
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -64,7 +64,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Populate binary env
         shell: cmd
@@ -141,7 +141,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Populate binary env
         shell: cmd
@@ -201,7 +201,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cpu-shared-with-deps-release
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -51,7 +51,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -166,7 +166,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash

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

@@ -58,7 +58,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -173,7 +173,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -283,7 +283,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cpu-shared-with-deps-debug
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -306,7 +306,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -422,7 +422,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -533,7 +533,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cuda12_6-shared-with-deps-debug
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -556,7 +556,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -672,7 +672,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -783,11 +783,261 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cuda12_8-shared-with-deps-debug
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-debug-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.9"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-debug-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-debug-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.9"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-debug-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-debug-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.9"
+      build_name: libtorch-cuda12_9-shared-with-deps-debug
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
@@ -806,7 +1056,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -922,7 +1172,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -1033,7 +1283,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cuda13_0-shared-with-deps-debug
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -51,7 +51,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -166,7 +166,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash

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

@@ -58,7 +58,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -173,7 +173,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -283,7 +283,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cpu-shared-with-deps-release
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -306,7 +306,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -422,7 +422,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -533,7 +533,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cuda12_6-shared-with-deps-release
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -556,7 +556,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -672,7 +672,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -783,11 +783,261 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cuda12_8-shared-with-deps-release
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.9"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-release-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.9"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.9"
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
@@ -806,7 +1056,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      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
@@ -922,7 +1172,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Display EC2 information
         shell: bash
@@ -1033,7 +1283,7 @@ jobs:
       LIBTORCH_VARIANT: shared-with-deps
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       build_name: libtorch-cuda13_0-shared-with-deps-release
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/inductor-nightly.yml

@@ -37,7 +37,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-default-label-prefix
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       test-matrix: |
@@ -56,7 +56,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: nightly-dynamo-benchmarks-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.nightly-dynamo-benchmarks-build.outputs.docker-image }}
       test-matrix: ${{ needs.nightly-dynamo-benchmarks-build.outputs.test-matrix }}
       timeout-minutes: 720

.github/workflows/inductor-perf-test-nightly-x86-zen.yml

@@ -43,11 +43,6 @@ on:
         required: false
         type: boolean
         default: false
-      freezing:
-        description: Run freezing?
-        required: false
-        type: boolean
-        default: true
       benchmark_configs:
         description: The list of configs used the benchmark
         required: false
@@ -80,7 +75,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -106,8 +101,8 @@ jobs:
     needs: inductor-build
     if: github.event.schedule == '0 7 * * *'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
-      dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true-freezing-true
+      build-environment: linux-jammy-py3.9-gcc11-build
+      dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
       timeout-minutes: 720
@@ -121,9 +116,10 @@ jobs:
     name: inductor-test
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-build
+    if: github.event_name == 'workflow_dispatch'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
-      dashboard-tag: training-${{ inputs.training || 'false' }}-inference-${{ inputs.inference || 'true' }}-default-${{ inputs.default || 'true' }}-dynamic-${{ inputs.dynamic || 'true' }}-cppwrapper-${{ inputs.cppwrapper || 'true' }}-aotinductor-${{ inputs.aotinductor || 'true' }}-freezing-${{ inputs.freezing || 'true' }}
+      build-environment: linux-jammy-py3.9-gcc11-build
+      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
       timeout-minutes: 720

.github/workflows/inductor-perf-test-nightly-x86.yml

@@ -80,7 +80,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -107,7 +107,7 @@ jobs:
     needs: inductor-build
     if: github.event.schedule == '0 7 * * *'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true-freezing-true
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
@@ -124,7 +124,7 @@ jobs:
     needs: inductor-build
     if: github.event_name == 'workflow_dispatch'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-freezing-${{ inputs.freezing }}
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}

.github/workflows/inductor-periodic.yml

@@ -39,7 +39,7 @@ jobs:
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
-      cuda-arch-list: '8.0;8.6'
+      cuda-arch-list: '8.6'
       test-matrix: |
         { include: [
           { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
@@ -62,7 +62,7 @@ jobs:
           { config: "dynamic_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "dynamic_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "dynamic_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "aot_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.aws.a100" },
+          { config: "aot_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "aot_inductor_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "aot_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "aot_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
@@ -154,7 +154,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-default-label-prefix
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       test-matrix: |
@@ -200,7 +200,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: periodic-dynamo-benchmarks-cpu-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/inductor-rocm.yml

@@ -3,10 +3,18 @@ name: inductor-rocm
 on:
   push:
     branches:
-      - main
+      #- main
       - release/*
     tags:
       - ciflow/inductor-rocm/*
+  schedule:
+    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
+    # Also run less frequently on weekends.
+    - cron: 45 0,8,16 * * 1-5
+    - cron: 45 4 * * 0,6
+    - cron: 45 4,12,20 * * 1-5
+    - cron: 45 12 * * 0,6
+    - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
   workflow_dispatch:
 
 concurrency:

.github/workflows/inductor-unittest.yml

@@ -110,7 +110,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
@@ -127,7 +127,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-cpu-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/inductor.yml

@@ -79,7 +79,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
@@ -101,7 +101,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-cpu-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/nightly.yml

@@ -54,7 +54,7 @@ jobs:
       - get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11
+      build-environment: linux-jammy-py3.9-gcc11
       docker-image: ${{ needs.docs-build.outputs.docker-image }}
       push: ${{ github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' || startsWith(github.event.ref, 'refs/tags/v') }}
       run-doxygen: true

.github/workflows/operator_benchmark.yml

@@ -14,10 +14,6 @@ on:
   schedule:
     # Run at 07:00 UTC every Sunday
     - cron: 0 7 * * 0
-  pull_request:
-    paths:
-      - benchmarks/operator_benchmark/**
-      - .github/workflows/operator_benchmark.yml
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
@@ -33,7 +29,7 @@ jobs:
     name: opbenchmark-build
     uses: ./.github/workflows/_linux-build.yml
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -46,7 +42,7 @@ jobs:
     name: opbenchmark-on-demand-build
     uses: ./.github/workflows/_linux-build.yml
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -59,7 +55,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: opbenchmark-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
       test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/rocm.yml

@@ -3,13 +3,19 @@ name: rocm
 on:
   push:
     branches:
-      - main
+  #     - main
       - release/*
     tags:
       - ciflow/rocm/*
   workflow_dispatch:
   schedule:
-    - cron: 29 8 * * *  # about 1:29am PDT
+    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
+    # Also run less frequently on weekends.
+    - cron: 45 0,8,16 * * 1-5
+    - cron: 45 4 * * 0,6
+    - cron: 45 4,12,20 * * 1-5
+    - cron: 45 12 * * 0,6
+    - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}

.github/workflows/trunk.yml

@@ -240,7 +240,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11
+      build-environment: linux-jammy-py3.9-gcc11
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -255,7 +255,7 @@ jobs:
       - verify-cachebench-cpu-build
       - target-determination
     with:
-      build-environment: linux-jammy-py3.10-gcc11
+      build-environment: linux-jammy-py3.9-gcc11
       docker-image: ${{ needs.verify-cachebench-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.verify-cachebench-cpu-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/vllm.yml

@@ -2,9 +2,6 @@ name: vllm-test
 
 on:
   push:
-    branches:
-      - main
-      - release/*
     tags:
       - ciflow/vllm/*
   workflow_dispatch:
@@ -48,18 +45,14 @@ jobs:
           { config: "vllm_basic_models_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_entrypoints_test", shard: 1, num_shards: 1,runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_regression_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
+          { config: "vllm_lora_280_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_multi_model_processor_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_pytorch_compilation_unit_tests", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
-          { config: "vllm_lora_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
-          { config: "vllm_multi_model_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
-          { config: "vllm_languagde_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
-          { config: "vllm_distributed_test_2_gpu_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 0, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 1, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 2, num_shards: 4,  runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 3, num_shards: 4,  runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
-          { config: "vllm_lora_tp_test_distributed", shard: 1, num_shards: 1, runner: "linux.g6.12xlarge.nvidia.gpu"},
-          { config: "vllm_distributed_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.12xlarge.nvidia.gpu"}
+          { config: "vllm_lora_tp_test_distributed", shard: 1, num_shards: 1, runner: "linux.aws.h100.4"},
         ]}
     secrets: inherit
 

.gitignore

@@ -389,5 +389,3 @@ android/pytorch_android_torchvision/.cxx
 
 # Claude Code local configuration
 CLAUDE.local.md
-/test_*.py
-/debug_*.py

.lintrunner.toml

@@ -13,7 +13,7 @@ exclude_patterns = [
     '**/fb/**',
     'functorch/docs/**',
     'functorch/examples/**',
-    'functorch/docs/source/tutorials/**',
+    'functorch/notebooks/**',
     'torch/_inductor/fx_passes/serialized_patterns/**',
     'torch/_inductor/autoheuristic/artifacts/**',
     'scripts/**',
@@ -1568,6 +1568,7 @@ include_patterns = [
 exclude_patterns = [
     'caffe2/**',
     'functorch/docs/**',
+    'functorch/notebooks/**',
     'torch/_inductor/fx_passes/serialized_patterns/**',
     'torch/_inductor/autoheuristic/artifacts/**',
     'test/dynamo/cpython/**',

BUILD.bazel

@@ -810,7 +810,7 @@ cc_library(
     name = "torch_python",
     srcs = libtorch_python_core_sources
         + if_cuda(libtorch_python_cuda_sources)
-        + libtorch_python_distributed_sources
+        + if_cuda(libtorch_python_distributed_sources)
         + GENERATED_AUTOGRAD_PYTHON,
     hdrs = glob([
         "torch/csrc/generic/*.cpp",

CLAUDE.md

@@ -1,15 +0,0 @@
-# Testing
-
-Use our test class and test runner:
-
-```
-from torch.testing._internal.common_utils import run_tests, TestCase
-
-class TestFeature(TestCase):
-    ...
-
-if __name__ == "__main__":
-    run_tests()
-```
-
-To test Tensor equality, use assertEqual.

CMakeLists.txt

@@ -234,7 +234,6 @@ cmake_dependent_option(INSTALL_TEST "Install test binaries if BUILD_TEST is on"
 option(USE_CPP_CODE_COVERAGE "Compile C/C++ with code coverage flags" OFF)
 option(USE_COLORIZE_OUTPUT "Colorize output during compilation" ON)
 option(USE_ASAN "Use Address+Undefined Sanitizers" OFF)
-option(USE_LSAN "Use Leak Sanitizer" OFF)
 option(USE_TSAN "Use Thread Sanitizer" OFF)
 option(USE_CUDA "Use CUDA" ON)
 option(USE_XPU "Use XPU" ON)
@@ -874,28 +873,17 @@ cmake_dependent_option(
   "Whether to build the flash_attention kernel for scaled dot product attention.\
   Will be disabled if not supported by the platform"
   ON
-  "(USE_CUDA AND NOT MSVC) OR USE_ROCM"
+  "USE_CUDA OR USE_ROCM;NOT MSVC"
   OFF)
 
 cmake_dependent_option(
   USE_FBGEMM_GENAI
   "Whether to build FBGEMM GenAI quantized GEMM kernels.\
   Will be disabled if not supported by the platform"
-  ON
-  "USE_ROCM"
+  OFF
+  "USE_CUDA OR USE_ROCM"
   OFF)
 
-IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-  message(WARNING "Unsupported ROCM arch for FBGEMM GenAI, will set USE_FBGEMM_GENAI to OFF")
-  set(USE_FBGEMM_GENAI off)
-endif()
-
-# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100.
-if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0" AND CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
-  message(STATUS "Setting USE_FBGEMM_GENAI to ON, doing CUDA build for SM100a")
-  set(USE_FBGEMM_GENAI ON)
-endif()
-
 # CAVEAT: Again, Flash Attention2 will error while building for sm52 while Mem
 # Eff Attention won't
 cmake_dependent_option(
@@ -909,7 +897,7 @@ cmake_dependent_option(
 # USE_FLASH_ATTENTION -> USE_ROCM -> Dependencies.cmake -> aotriton.cmake
 #
 if(USE_ROCM)
-  if(USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION)
+  if(UNIX AND (USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION))
     include(cmake/External/aotriton.cmake)
   endif()
 endif()

CONTRIBUTING.md

@@ -88,13 +88,13 @@ source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
 
 * If you want to have no-op incremental rebuilds (which are fast), see [Make no-op build fast](#make-no-op-build-fast) below.
 
-* When installing with `python -m pip install -e . -v --no-build-isolation` (in contrast to `python -m pip install . -v --no-build-isolation`) Python runtime will use
+* When installing with `python -m pip install -e .` (in contrast to `python -m pip install .`) Python runtime will use
   the current local source-tree when importing `torch` package. (This is done by creating [`.egg-link`](https://wiki.python.org/moin/PythonPackagingTerminology#egg-link) file in `site-packages` folder)
   This way you do not need to repeatedly install after modifying Python files (`.py`).
   However, you would need to reinstall if you modify Python interface (`.pyi`, `.pyi.in`) or non-Python files (`.cpp`, `.cc`, `.cu`, `.h`, ...).
 
 
-  One way to avoid running `python -m pip install -e . -v --no-build-isolation` every time one makes a change to C++/CUDA/ObjectiveC files on Linux/Mac,
+  One way to avoid running `python -m pip install -e .` every time one makes a change to C++/CUDA/ObjectiveC files on Linux/Mac,
   is to create a symbolic link from `build` folder to `torch/lib`, for example, by issuing following:
   ```bash
   pushd torch/lib; sh -c "ln -sf ../../build/lib/libtorch_cpu.* ."; popd
@@ -116,7 +116,7 @@ source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
 
   Next run `python setup.py clean`. After that, you can install in editable mode again.
 
-* If you run into errors when running `python -m pip install -e . -v --no-build-isolation`, here are some debugging steps:
+* If you run into errors when running `python -m pip install -e .`, here are some debugging steps:
   1. Run `printf '#include <stdio.h>\nint main() { printf("Hello World");}'|clang -x c -; ./a.out` to make sure
   your CMake works and can compile this simple Hello World program without errors.
   2. Nuke your `build` directory. The `setup.py` script compiles binaries into the `build` folder and caches many
@@ -129,10 +129,10 @@ source venv/bin/activate  # or `& .\venv\Scripts\Activate.ps1` on Windows
       git clean -xdf
       python setup.py clean
       git submodule update --init --recursive
-      python -m pip install --group dev
+      python -m pip install -r requirements.txt
       python -m pip install --no-build-isolation -v -e .
       ```
-  4. The main step within `python -m pip install -e . -v --no-build-isolation` is running `make` from the `build` directory. If you want to
+  4. The main step within `python -m pip install -e .` is running `cmake --build build` from the `build` directory. If you want to
     experiment with some environment variables, you can pass them into the command:
       ```bash
       ENV_KEY1=ENV_VAL1[, ENV_KEY2=ENV_VAL2]* CMAKE_FRESH=1 python -m pip install --no-build-isolation -v -e .
@@ -259,7 +259,6 @@ dependencies as well as the nightly binaries into the repo directory.
       support for PyTorch.
 * [tools](tools) - Code generation scripts for the PyTorch library.
   See [README](tools/README.md) of this directory for more details.
-* [torchgen](torchgen) - contains the logic and tooling for generating PyTorch's low-level C++ and Python bindings from operator definitions, typically specified in native_functions.yaml
 * [test](test) - Python unit tests for PyTorch Python frontend.
   * [test_torch.py](test/test_torch.py) - Basic tests for PyTorch
     functionality.
@@ -295,7 +294,7 @@ The following packages should be installed with `pip`:
 - `pytest` - recommended to run tests more selectively
 Running
 ```
-pip install --group dev
+pip install -r requirements.txt
 ```
 will install these dependencies for you.
 
@@ -646,9 +645,9 @@ can be selected interactively with your mouse to zoom in on a particular part of
 the program execution timeline. The `--native` command-line option tells
 `py-spy` to record stack frame entries for PyTorch C++ code. To get line numbers
 for C++ code it may be necessary to compile PyTorch in debug mode by prepending
-your `python -m pip install -e . -v --no-build-isolation` call to compile
-PyTorch with `DEBUG=1`. Depending on your operating system it may also be
-necessary to run `py-spy` with root privileges.
+your `python -m pip install -e .` call to compile PyTorch with `DEBUG=1`.
+Depending on your operating system it may also be necessary to run `py-spy` with
+root privileges.
 
 `py-spy` can also work in an `htop`-like "live profiling" mode and can be
 tweaked to adjust the stack sampling rate, see the `py-spy` readme for more
@@ -656,10 +655,10 @@ details.
 
 ## Managing multiple build trees
 
-One downside to using `python -m pip install -e . -v --no-build-isolation` is
-that your development version of PyTorch will be installed globally on your
-account (e.g., if you run `import torch` anywhere else, the development version
-will be used).
+One downside to using `python -m pip install -e .` is that your development
+version of PyTorch will be installed globally on your account (e.g., if
+you run `import torch` anywhere else, the development version will be
+used).
 
 If you want to manage multiple builds of PyTorch, you can make use of
 [venv environments](https://docs.python.org/3/library/venv.html) to maintain
@@ -720,7 +719,7 @@ options.
 
 ### Code completion and IDE support
 
-When using `python -m pip install -e . -v --no-build-isolation`, PyTorch will generate
+When using `python -m pip install -e .`, PyTorch will generate
 a `compile_commands.json` file that can be used by many editors
 to provide command completion and error highlighting for PyTorch's
 C++ code. You need to `pip install ninja` to generate accurate

README.md

@@ -243,7 +243,7 @@ git submodule update --init --recursive
 
 ```bash
 # Run this command from the PyTorch directory after cloning the source code using the “Get the PyTorch Source“ section above
-pip install --group dev
+pip install -r requirements.txt
 ```
 
 **On Linux**
@@ -394,7 +394,7 @@ On macOS
 
 ```bash
 export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
-MACOSX_DEPLOYMENT_TARGET=11.0 CMAKE_ONLY=1 python setup.py build
+MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ CMAKE_ONLY=1 python setup.py build
 ccmake build  # or cmake-gui build
 ```
 

RELEASE.md

@@ -50,7 +50,6 @@ Following is the Release Compatibility Matrix for PyTorch releases:
 
 | PyTorch version | Python | C++ | Stable CUDA | Experimental CUDA | Stable ROCm |
 | --- | --- | --- | --- | --- | --- |
-| 2.9 | >=3.10, <=(3.14, 3.14t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 13.0 (CUDNN 9.13.0.50) | ROCm 6.4 |
 | 2.8 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 12.9 (CUDNN 9.10.2.21) | ROCm 6.4 |
 | 2.7 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8 (CUDNN 9.1.0.70), CUDA 12.6 (CUDNN 9.5.1.17) | CUDA 12.8 (CUDNN 9.7.1.26) | ROCm 6.3 |
 | 2.6 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8, CUDA 12.4 (CUDNN 9.1.0.70) | CUDA 12.6 (CUDNN 9.5.1.17) | ROCm 6.2.4 |

SECURITY.md

@@ -16,8 +16,6 @@ However, if you believe you have found a security vulnerability in PyTorch, we e
 
 Please report security issues using https://github.com/pytorch/pytorch/security/advisories/new
 
-All reports submitted thru the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.
-
 Please refer to the following page for our responsible disclosure policy, reward guidelines, and those things that should not be reported:
 
 https://www.facebook.com/whitehat

aten/src/ATen/CMakeLists.txt

@@ -252,88 +252,47 @@ if(USE_MEM_EFF_ATTENTION)
   list(APPEND ATen_ATTENTION_KERNEL_SRCS ${mem_eff_attention_cuda_kernels_cu})
 endif()
 
+IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
+  message(WARNING "Unsupported ROCM arch for FBGEMM GenAI, will set USE_FBGEMM_GENAI to OFF")
+  set(USE_FBGEMM_GENAI off)
+endif()
+
 # FBGEMM GenAI
 IF(USE_FBGEMM_GENAI)
   set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
-  set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
-  if(USE_CUDA)
-    # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
-    # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
-    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
-    file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
-      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
-      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
-    list(FILTER fbgemm_genai_native_cuda_cu INCLUDE REGEX ${FBGEMM_CUTLASS_KERNELS_REGEX})
+  set(FBGEMM_GENAI_DIR ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
 
-    # PyTorch is not built for 10.0a in CI, due to lack of portability,
-    # so we need to explicitly build these files for 10.0a.
-    foreach(cu_file ${fbgemm_genai_native_cuda_cu})
-      _BUILD_FOR_ADDITIONAL_ARCHS(
-        "${cu_file}"
-        "100a")
-    endforeach()
+  if(USE_ROCM)
+    # Only include the kernels we want to build to avoid increasing binary size.
+    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
+      "${FBGEMM_GENAI_DIR}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
+      "${FBGEMM_GENAI_DIR}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
+    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-    file(GLOB_RECURSE fbgemm_genai_native_cuda_cpp
-      "${FBGEMM_GENAI_SRCS}/common/*.cpp"
-    )
-
-    # Combine all source files into a single list
-    list(APPEND fbgemm_genai_all_sources
-      ${fbgemm_genai_native_cuda_cu}
-      ${fbgemm_genai_native_cuda_cpp}
-    )
-
-    # Now, create the library and provide the sources at the same time
-    add_library(fbgemm_genai OBJECT ${fbgemm_genai_all_sources})
+    # Add additional HIPCC compiler flags for performance
+    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+      -mllvm
+      -amdgpu-coerce-illegal-types=1
+      -mllvm
+      -enable-post-misched=0
+      -mllvm
+      -greedy-reverse-local-assignment=1
+      -fhip-new-launch-api)
 
+    hip_add_library(
+      fbgemm_genai STATIC
+      ${fbgemm_genai_native_rocm_hip}
+      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
     set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-
-    set(fbgemm_genai_mx8mx8bf16_grouped
-      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
-    )
+    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
 
     target_include_directories(fbgemm_genai PUBLIC
-      ${FBGEMM_THIRD_PARTY}/cutlass/include
-      ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${fbgemm_genai_mx8mx8bf16_grouped}
-      ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
-      ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
+      # FBGEMM version of Composable Kernel is used due to some customizations
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
+      ${FBGEMM_GENAI_DIR}/include/
+      ${FBGEMM_GENAI_DIR}/common/include/
     )
-  else()
-    if(USE_ROCM)
-      # Only include the kernels we want to build to avoid increasing binary size.
-      file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-      set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
-
-      # Add additional HIPCC compiler flags for performance
-      set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-        -mllvm
-        -amdgpu-coerce-illegal-types=1
-        -mllvm
-        -enable-post-misched=0
-        -mllvm
-        -greedy-reverse-local-assignment=1
-        -fhip-new-launch-api)
-
-      hip_add_library(
-        fbgemm_genai STATIC
-        ${fbgemm_genai_native_rocm_hip}
-        HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
-      set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-      target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
-
-      target_include_directories(fbgemm_genai PUBLIC
-        # FBGEMM version of Composable Kernel is used due to some customizations
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-        ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
-        ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
-      )
-    endif()
   endif()
 endif()
 
@@ -676,26 +635,12 @@ if(USE_CUDA AND NOT USE_ROCM)
   add_definitions(-DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
-
-  # Add FBGEMM_GENAI include directories for torch_ops.h
-  if(USE_FBGEMM_GENAI)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
-  endif()
-
   if($ENV{ATEN_STATIC_CUDA})
-    if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
-      list(APPEND ATen_CUDA_DEPENDENCY_LIBS
-          ${CUDA_LIBRARIES}
-          CUDA::cusparse_static
-          CUDA::cufft_static_nocallback)
-    else()
-      list(APPEND ATen_CUDA_DEPENDENCY_LIBS
-          ${CUDA_LIBRARIES}
-          CUDA::cusparse_static
-          CUDA::cufft_static)
-    endif()
-
+    list(APPEND ATen_CUDA_DEPENDENCY_LIBS
+      ${CUDA_LIBRARIES}
+      CUDA::cusparse_static
+      CUDA::cufft_static_nocallback
+    )
    if(NOT BUILD_LAZY_CUDA_LINALG)
      list(APPEND ATen_CUDA_DEPENDENCY_LIBS
        CUDA::cusolver_static

aten/src/ATen/DLConvertor.cpp

@@ -308,44 +308,17 @@ void fillVersion<DLManagedTensorVersioned>(
 // constructed out of ATen tensor
 template <class T>
 T* toDLPackImpl(const Tensor& src) {
-  auto view = src;
-
-  // Detect whether there is need to normalize the strides
-  // Background: gh-83069
-  //
-  // However, normalizing strides can come at a high-cost
-  // to slow down toDLPack conversion 3x, so we
-  // only normalize if needed.
-  //
-  // The following code detects whether the src follows
-  // a continuous pattern. If the src follows such pattern (common-case)
-  // then we do not need to normalize the strides.
-  bool need_normalize_strides = false;
-  int64_t expected_stride = 1;
-  for (int i = src.dim() - 1; i >= 0; i--) {
-    // detect if we do not meet continuous pattern
-    // and the size is 1, so there is opportunity to normalize
-    if (src.stride(i) != expected_stride && src.size(i) == 1) {
-      need_normalize_strides = true;
-      break;
+  // create a new tensor with possibly normalized strides
+  // gh-83069
+  auto shape = src.sizes();
+  auto strides = src.strides().vec();
+  for (int i = 0; i < src.dim(); i++) {
+    if (shape[i] < 2) {
+      strides[i] = 1;
     }
-    expected_stride *= src.size(i);
-  }
-
-  // less common case, try normalizing the strides
-  if (need_normalize_strides) {
-    // create a new tensor with possibly normalized strides
-    // gh-83069
-    auto shape = src.sizes();
-    auto strides = src.strides().vec();
-    for (int i = 0; i < src.dim(); i++) {
-      if (shape[i] < 2) {
-        strides[i] = 1;
-      }
-    }
-    view = src.as_strided(shape, strides, src.storage_offset());
   }
 
+  auto view = src.as_strided(shape, strides, src.storage_offset());
   ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);
   atDLMTensor->handle = view;
   atDLMTensor->tensor.manager_ctx = atDLMTensor;

aten/src/ATen/DTensorState.cpp

@@ -1,17 +0,0 @@
-#include <ATen/DTensorState.h>
-
-namespace at {
-
-namespace {
-thread_local bool kDTensorAllowImplicitReplication = false;
-}
-
-bool get_dtensor_allow_implicit_replication() {
-  return kDTensorAllowImplicitReplication;
-}
-
-void set_dtensor_allow_implicit_replication(bool enabled) {
-  kDTensorAllowImplicitReplication = enabled;
-}
-
-} // namespace at

aten/src/ATen/DTensorState.h

@@ -1,34 +0,0 @@
-#pragma once
-
-#include <c10/macros/Macros.h>
-
-namespace at {
-
-TORCH_API bool get_dtensor_allow_implicit_replication();
-TORCH_API void set_dtensor_allow_implicit_replication(bool enabled);
-
-struct DTensorAllowImplicitReplication {
-  DTensorAllowImplicitReplication()
-      : prev_dtensor_allow_implicit_replication_(
-            get_dtensor_allow_implicit_replication()) {
-    set_dtensor_allow_implicit_replication(true);
-  }
-
-  DTensorAllowImplicitReplication(const DTensorAllowImplicitReplication&) =
-      delete;
-  DTensorAllowImplicitReplication& operator=(
-      const DTensorAllowImplicitReplication&) = delete;
-  DTensorAllowImplicitReplication(DTensorAllowImplicitReplication&&) = delete;
-  DTensorAllowImplicitReplication& operator=(
-      DTensorAllowImplicitReplication&&) = delete;
-
-  ~DTensorAllowImplicitReplication() {
-    set_dtensor_allow_implicit_replication(
-        prev_dtensor_allow_implicit_replication_);
-  }
-
- private:
-  bool prev_dtensor_allow_implicit_replication_;
-};
-
-} // namespace at

aten/src/ATen/SparseTensorImpl.h

@@ -133,12 +133,12 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
         "resize_ called on tensor with symbolic shape")
     TORCH_CHECK(
         sparse_dim + dense_dim == static_cast<int64_t>(size.size()),
-        "'len(size) == sparse_dim + dense_dim' is not satisfied: len(size) = ",
-        size.size(),
-        ", sparse_dim = ",
+        "number of dimensions must be sparse_dim (",
         sparse_dim,
-        ", dense_dim = ",
-        dense_dim);
+        ") + dense_dim (",
+        dense_dim,
+        "), but got ",
+        size.size());
     if (nnz() > 0) {
       [[maybe_unused]] auto constexpr alt_options_msg =
           "You could try the following options:\n\
@@ -254,12 +254,12 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
         "resize_and_clear_ called on tensor with symbolic shape")
     TORCH_CHECK(
         sparse_dim + dense_dim == static_cast<int64_t>(size.size()),
-        "'len(size) == sparse_dim + dense_dim' is not satisfied: len(size) = ",
-        size.size(),
-        ", sparse_dim = ",
+        "number of dimensions must be sparse_dim (",
         sparse_dim,
-        ", dense_dim = ",
-        dense_dim);
+        ") + dense_dim (",
+        dense_dim,
+        "), but got ",
+        size.size());
 
     set_sizes_and_strides(size, std::vector<int64_t>(size.size()));
     sparse_dim_ = sparse_dim;

aten/src/ATen/ThreadLocalState.cpp

@@ -8,7 +8,6 @@
 #include <ATen/record_function.h>
 #include <ATen/SavedTensorHooks.h>
 #include <ATen/FunctionalTensorWrapper.h>
-#include <ATen/DTensorState.h>
 
 namespace at {
 
@@ -20,7 +19,6 @@ ThreadLocalState::ThreadLocalState()
       torch_dispatch_mode_state_(c10::impl::TorchDispatchModeTLS::get_state()), python_dispatcher_state_(c10::impl::PythonDispatcherTLS::get_state()),
       python_torch_function_state_(at::impl::PythonTorchFunctionTLS::get_state()),
       saved_tensors_default_hooks_state_(at::SavedTensorDefaultHooks::get_tls_state()), functionalization_reapply_views_state_(at::functionalization::impl::getFunctionalizationReapplyViewsTLS()),
-      dtensor_allow_implicit_replication_(at::get_dtensor_allow_implicit_replication()),
       saved_objects_(at::impl::ThreadLocalPythonObjects::get_state()) {
 #if !defined(CAFFE2_IS_XPLAT_BUILD) && !defined(C10_MOBILE) && !defined(BUILD_LITE_INTERPRETER)
   for(size_t i=0; i<autocast_dtypes_.size(); i++) {
@@ -54,8 +52,6 @@ void ThreadLocalState::setThreadLocalState(
 
   c10::impl::PythonDispatcherTLS::set_state(state.python_dispatcher_state_);
 
-  at::set_dtensor_allow_implicit_replication(state.dtensor_allow_implicit_replication_);
-
   c10::ThreadLocalDebugInfo::_forceCurrentDebugInfo(state.debug_info_);
 
   c10::impl::_force_tls_local_dispatch_key_set(state.dispatch_key_);

aten/src/ATen/ThreadLocalState.h

@@ -75,8 +75,6 @@ class TORCH_API ThreadLocalState {
 
   bool functionalization_reapply_views_state_;
 
-  bool dtensor_allow_implicit_replication_;
-
   // TLS for arbitrary python objects that is registered via hooks
   at::impl::ThreadLocalPythonObjects saved_objects_;
 

aten/src/ATen/core/dynamic_type.h

@@ -64,7 +64,6 @@ constexpr DynamicTypeBits kDynamicClassTypeBit = DYNAMIC_TYPE_BIT(10);
   _(ScalarType, kDynamicIntTypeBit, 1)                                \
   _(Layout, kDynamicIntTypeBit, 1)                                        \
   _(SymInt, kDynamicIntTypeBit, 1)                                        \
-  _(SymBool, kDynamicIntTypeBit, 1)                                        \
   _(MemoryFormat, kDynamicIntTypeBit, 1)
 
 #define FORWARD_DECL_TYPE(NAME, _, __) struct NAME ## Type;

aten/src/ATen/cuda/CUDABlas.cpp

@@ -644,8 +644,6 @@ inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYP
   void * beta_ptr = &fbeta;
 #ifdef USE_ROCM
   int flag = 0;
-  rocblas_datatype c_type = std::is_same<C_Dtype, float>::value ? rocblas_datatype_f32_r : rocblas_datatype_f16_r;
-  rocblas_datatype d_type = c_type;
 #if USE_GEMM_FLAGS_FP16_ALT_IMPL
   flag = at::ROCmBackwardPassGuard::is_backward_pass() ? rocblas_gemm_flags_fp16_alt_impl : 0;
 #endif
@@ -654,8 +652,8 @@ inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYP
                                    hipOperationToRocOperation(opb), (int)m, (int)n, (int)k,
                                    (void*)alpha_ptr, a, rocblas_datatype_f16_r, (int)lda, stridea,
                                    b, rocblas_datatype_f16_r, (int)ldb, strideb,
-                                   (void*)beta_ptr, c, c_type, (int)ldc, stridec,
-                                   c, d_type, (int)ldc, stridec,
+                                   (void*)beta_ptr, c, rocblas_datatype_f16_r, (int)ldc, stridec,
+                                   c, rocblas_datatype_f16_r, (int)ldc, stridec,
                                    (int) num_batches, rocblas_datatype_f32_r, rocblas_gemm_algo_standard,
                                    0, flag)));
 #else
@@ -1098,8 +1096,6 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
   GEMM_CHECK_ARGVALUES(at::Half);
 #ifdef USE_ROCM
   int flag = 0;
-  rocblas_datatype c_type = std::is_same<C_Dtype, float>::value ? rocblas_datatype_f32_r : rocblas_datatype_f16_r;
-  rocblas_datatype d_type = c_type;
 #if USE_GEMM_FLAGS_FP16_ALT_IMPL
   flag = at::ROCmBackwardPassGuard::is_backward_pass() ? rocblas_gemm_flags_fp16_alt_impl : 0;
 #endif
@@ -1119,10 +1115,10 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
       ldb,
       beta_ptr,
       c,
-      c_type,
+      rocblas_datatype_f16_r,
       ldc,
       c,
-      d_type,
+      rocblas_datatype_f16_r,
       ldc,
       rocblas_datatype_f32_r,
       rocblas_gemm_algo_standard,
@@ -1941,11 +1937,11 @@ void scaled_gemm(
   computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_TRANSB, _cublasOpFromChar(transb));
   cublasLtMatmulDescAttributes_t matmulDescA = CUBLASLT_MATMUL_DESC_A_SCALE_POINTER;
   cublasLtMatmulDescAttributes_t matmulDescB = CUBLASLT_MATMUL_DESC_B_SCALE_POINTER;
-#if defined(USE_ROCM) && !defined(HIPBLASLT_OUTER_VEC) && defined(HIPBLASLT_VEC_EXT)
   // hipblaslt supported row-wise before cublas, and did so their own way (via
   // the SCALE_POINTERSs), but then migrated to match how cublas does it (via
   // the SCALE_MODEs). Here we check for this early custom mode.
   bool use_rowwise = (mat1_scaling_type == ScalingType::RowWise && mat2_scaling_type == ScalingType::RowWise);
+#if defined(USE_ROCM) && !defined(HIPBLASLT_OUTER_VEC) && defined(HIPBLASLT_VEC_EXT)
   if (use_rowwise) {
     matmulDescA = HIPBLASLT_MATMUL_DESC_A_SCALE_POINTER_VEC_EXT;
     matmulDescB = HIPBLASLT_MATMUL_DESC_B_SCALE_POINTER_VEC_EXT;
@@ -1960,12 +1956,8 @@ void scaled_gemm(
             }
   #endif
   }
-#elif (CUDA_VERSION < 12090) && !defined(USE_ROCM)
-  // hipblaslt supported row-wise before cublas, and did so their own way (via
-  // the SCALE_POINTERSs), but then migrated to match how cublas does it (via
-  // the SCALE_MODEs). Here we check for this early custom mode.
-  bool use_rowwise = (mat1_scaling_type == ScalingType::RowWise && mat2_scaling_type == ScalingType::RowWise);
-  // rowwise isn't supported using older cublaslt or older hipblaslt
+#else
+  // rowwise isn't supported using cublaslt or older hipblaslt
   TORCH_INTERNAL_ASSERT(use_rowwise == false, "rowwise scaled_gemm not supported with blaslt");
 #endif  // if defined(USE_ROCM) && !defined(HIPBLASLT_OUTER_VEC) && defined(HIPBLASLT_VEC_EXT)
   computeDesc.setAttribute(matmulDescA, mat1_scale_ptr);

aten/src/ATen/cuda/detail/OffsetCalculator.cuh

@@ -45,24 +45,6 @@ struct OffsetCalculator {
 
   C10_HOST_DEVICE offset_type get(index_t linear_idx) const {
     offset_type offsets;
-
-#if defined(USE_ROCM)
-    if ((dims > 0) && (dims <= 2)) {
-      auto divmod = sizes_[0].divmod(linear_idx);
-#pragma unroll
-      for (int arg = 0; arg < NARGS; arg++)
-        offsets[arg] = divmod.mod * strides_[0][arg];
-      if (dims >= 2) {
-        divmod = sizes_[1].divmod(divmod.div);
-#pragma unroll
-        for (int arg = 0; arg < NARGS; arg++)
-          offsets[arg] += divmod.mod * strides_[1][arg];
-      }
-      // [...]
-      return offsets;
-    }
-#endif
-
     #pragma unroll
     for (int arg = 0; arg < NARGS; arg++) {
       offsets[arg] = 0;

aten/src/ATen/cuda/nvrtc_stub/ATenNVRTC.h

@@ -117,8 +117,6 @@ namespace at::cuda {
   _(nvrtcGetPTXSize)                              \
   _(nvrtcGetPTX)                                  \
   _(cuModuleLoadData)                             \
-  _(cuModuleLoad)                                 \
-  _(cuGetErrorString)                             \
   _(cuModuleGetFunction)                          \
   _(HIPOCCUPANCYMAXACTIVEBLOCKSPERMULTIPROCESSOR) \
   _(nvrtcGetErrorString)                          \

aten/src/ATen/functorch/BatchRulesModules.cpp

@@ -7,7 +7,6 @@
 #include <ATen/functorch/BatchRulesHelper.h>
 #include <ATen/functorch/PlumbingHelper.h>
 #include <ATen/core/dispatch/Dispatcher.h>
-#include <ATen/DTensorState.h>
 
 #include <utility>
 
@@ -45,13 +44,8 @@ static std::tuple<Tensor, std::optional<int64_t>> embedding_batch_rule(
   const auto weight_ = reshape_dim_into(*weight_bdim, 0, weight);
   auto indices_ = moveBatchDimToFront(indices, indices_bdim);
 
-  {
-    // getStepTensor returns a regular Tensor. If indices_ is a DTensor
-    // we want to allow this mixed DTensor-Tensor operation.
-    at::DTensorAllowImplicitReplication guard;
-    const auto range = getStepTensor(indices, batch_size, num_embeddings);
-    indices_ = indices_ + range;
-  }
+  const auto range = getStepTensor(indices, batch_size, num_embeddings);
+  indices_ = indices_ + range;
   auto result = at::embedding_symint(weight_, indices_, std::move(padding_idx), scale_grad_by_freq, sparse);
   return std::make_tuple(std::move(result), 0);
 }

aten/src/ATen/native/Blas.cpp

@@ -9,7 +9,6 @@
 #include <ATen/native/mkldnn/Matmul.h>
 #include <ATen/native/mkldnn/Linear.h>
 #include <ATen/native/Resize.h>
-#include <ATen/native/GroupedMMUtils.h>
 #if !defined(__s390x__) && !defined(__powerpc__)
 #include <cpuinfo.h>
 #endif
@@ -333,23 +332,4 @@ _scaled_mm_cpu(const Tensor& mat_a, const Tensor& mat_b,
   return _scaled_mm_out_cpu(mat_a, mat_b, scale_a, scale_b, bias, scale_result, out_dtype, use_fast_accum, out);
 }
 
-// TODO(vasiliy, future PR): figure out why we need to declare this function, when
-// other functions that live in ATen/native/*.cpp without declarations
-// or headers work just fine.
-Tensor _grouped_mm(const Tensor& mat_a, const Tensor& mat_b,
-const std::optional<at::Tensor>& offs,
-const std::optional<at::Tensor>& bias,
-std::optional<c10::ScalarType> out_dtype);
-
-Tensor _grouped_mm(const Tensor& mat_a, const Tensor& mat_b,
-const std::optional<at::Tensor>& offs,
-const std::optional<at::Tensor>& bias,
-std::optional<c10::ScalarType> out_dtype) {
-  _grouped_mm_validate_inputs(mat_a, mat_b, offs, bias, out_dtype);
-  const auto out_dtype_ = _resolve_grouped_mm_out_dtype(mat_a, mat_b, out_dtype);
-  Tensor out = create_grouped_gemm_output_tensor(mat_a, mat_b, offs, out_dtype_);
-  _grouped_mm_fallback(mat_a, mat_b, offs, bias, out_dtype, out);
-  return out;
-}
-
 }  // namespace at::native

aten/src/ATen/native/Convolution.cpp

@@ -14,7 +14,6 @@
 #include <c10/util/accumulate.h>
 #include <c10/util/irange.h>
 #include <c10/macros/Macros.h>
-#include <algorithm>
 #include <limits>
 #include <utility>
 
@@ -301,50 +300,67 @@ struct ConvParams {
   bool allow_tf32{};
 
   bool is_strided() const {
-    return std::any_of(
-      stride.cbegin(), stride.cend(), [](const T& s) { return s != 1; });
+    bool is_strided = false;
+    for (const auto& s : stride) {
+      is_strided |= (s != 1);
+    }
+    return is_strided;
   }
 
   bool is_dilated() const {
-    return std::any_of(
-      dilation.cbegin(), dilation.cend(), [](const T& d) { return d != 1; });
+    bool is_dilated = false;
+    for (const auto& d : dilation) {
+      is_dilated |= (d != 1);
+    }
+    return is_dilated;
   }
 
   bool is_padded() const {
-    return std::any_of(
-      padding.cbegin(), padding.cend(), [](const T& p) { return p != 0; });
+    bool is_padded = false;
+    for (auto p : padding) {
+      is_padded |= (p != 0);
+    }
+    return is_padded;
   }
 
   bool is_output_padding_neg() const {
-    return std::any_of(
-      output_padding.cbegin(),
-      output_padding.cend(),
-      [](const T& p) { return p < 0; });
+    bool is_non_neg = false;
+    for (const auto& p : output_padding) {
+      is_non_neg |= (p < 0);
+    }
+    return is_non_neg;
   }
 
   bool is_output_padding_big() const {
-    // Revisit this with std::views::zip at C++20.
+    bool is_big = false;
     for (auto i: c10::irange(output_padding.size())) {
-      if (output_padding[i] >= stride[i]) {
-        return true;
-      }
+      is_big |= (output_padding[i] >= stride[i]);
     }
-    return false;
+    return is_big;
   }
 
   bool is_padding_neg() const {
-    return std::any_of(
-      padding.cbegin(), padding.cend(), [](const T& p) { return p < 0; });
+    bool is_non_neg = false;
+    for (const auto& p : padding) {
+      is_non_neg |= (p < 0);
+    }
+    return is_non_neg;
   }
 
   bool is_dilation_neg() const {
-    return std::any_of(
-      dilation.cbegin(), dilation.cend(), [](const T& d) { return d < 0; });
+    bool is_non_neg = false;
+    for (const auto& p : dilation) {
+      is_non_neg |= (p < 0);
+    }
+    return is_non_neg;
   }
 
   bool is_stride_nonpos() const {
-    return std::any_of(
-      stride.cbegin(), stride.cend(), [](const T& s) { return s <= 0; });
+    bool is_nonpos = false;
+    for (const auto& s : stride) {
+      is_nonpos |= (s <= 0);
+    }
+    return is_nonpos;
   }
 
   void view1d_as_2d() {

aten/src/ATen/native/GroupedMMUtils.h

@@ -1,167 +0,0 @@
-#pragma once
-
-#include <ATen/core/Tensor.h>
-#include <ATen/TensorUtils.h>
-
-#ifndef AT_PER_OPERATOR_HEADERS
-#include <ATen/CPUFunctions.h>
-#include <ATen/Functions.h>
-#include <ATen/NativeFunctions.h>
-#else
-#include <ATen/ops/bmm.h>
-#include <ATen/ops/empty.h>
-#include <ATen/ops/empty_strided.h>
-#include <ATen/ops/mm.h>
-#endif
-
-namespace at::native {
-
-inline bool check_valid_strides_and_return_transposed(const Tensor& mat) {
-  IntArrayRef tensor_strides = mat.strides();
-  IntArrayRef tensor_sizes = mat.sizes();
-  int end_dim = mat.dim() - 1;
-  int alignment = 16 / mat.element_size();
-  TORCH_CHECK(uint64_t(mat.data_ptr()) % 16 ==0, "expected data_ptr to be aligned to 16 bytes\n");
-  if ((tensor_strides[end_dim - 1] == 1) && (tensor_strides[end_dim] >= std::max<int64_t>(1, tensor_sizes[end_dim - 1]))) {
-    TORCH_CHECK(tensor_strides[end_dim] % alignment == 0, "strides should be multiple of 16 bytes");
-    return true;
-  } else if ((tensor_strides[end_dim] == 1) && (tensor_strides[end_dim - 1] >= std::max<int64_t>(1, tensor_sizes[end_dim]))) {
-    TORCH_CHECK(tensor_strides[end_dim - 1] % alignment == 0, "strides should be multiple of 16 bytes");
-    return false;
-  } else {
-    TORCH_CHECK(false, "Invalid strides/sizes, got ", mat.strides(), " for strides and ", mat.sizes(), " for sizes");
-  }
-}
-
-inline at::Tensor create_grouped_gemm_output_tensor(const Tensor& mat_a,
-const Tensor& mat_b,
-const std::optional<at::Tensor>& offs,
-c10::ScalarType out_dtype
-) {
-  c10::SmallVector<int64_t, 3> out_size;
-  const bool a_is_2d = mat_a.dim() == 2;
-  const bool b_is_2d = mat_b.dim() == 2;
-  if (a_is_2d) {
-    if (b_is_2d) {
-      out_size = {offs->size(0), mat_a.size(0), mat_b.size(1)};
-    } else {
-      TORCH_CHECK(offs->size(0) == mat_b.size(0), "matrix batch sizes have to match");
-      out_size = {mat_a.size(0), mat_b.size(-1)};
-    }
-  } else {
-    if (b_is_2d) {
-      // this case is not actually encountered for MoE gemms
-      TORCH_CHECK(offs->size(0) == mat_a.size(0), "matrix batch sizes have to match");
-      out_size = {mat_a.size(1), mat_b.size(1)};
-    } else { // regular bmm
-      TORCH_CHECK(mat_a.size(0) == mat_b.size(0), "batched dimension has to match");
-      out_size = {mat_a.size(0), mat_a.size(1), mat_b.size(-1)};
-    }
-  }
-
-  #ifndef USE_ROCM
-  // For TMA transfers, strides of output tensor have to be either
-  // 1, or aligned to 16 bytes.
-  const auto last_dim = out_size.size() - 1;
-  const auto alignment = 16 / c10::elementSize(out_dtype);
-  const int64_t size_padded = (out_size[last_dim] + alignment - 1) / alignment * alignment;
-  std::vector<int64_t> out_stride;
-  if (a_is_2d != b_is_2d) {
-    out_stride = {size_padded, 1};
-  } else {
-    out_stride = {out_size[1] * size_padded, size_padded, 1};
-  }
-  return at::empty_strided(out_size, out_stride, mat_a.options().dtype(out_dtype));
-  #else
-  return at::empty(out_size, mat_a.options().dtype(out_dtype));
-  #endif
-}
-
-inline void _grouped_mm_validate_inputs(const Tensor& mat_a, const Tensor& mat_b,
-const std::optional<at::Tensor>& offs,
-const std::optional<at::Tensor>& bias,
-std::optional<c10::ScalarType> out_dtype) {
-  TORCH_CHECK((mat_a.dtype() == at::kBFloat16) || (mat_a.dtype() == at::kFloat) || (mat_a.dtype() == at::kHalf), "Expected mat_a to be Float32, BFloat16 or Float16 matrix, got ", mat_a.scalar_type());
-  TORCH_CHECK((mat_b.dtype() == at::kBFloat16) || (mat_b.dtype() == at::kFloat) || (mat_b.dtype() == at::kHalf), "Expected mat_b to be Float32, BFloat16 or Float16 matrix, got ", mat_b.scalar_type());
-  TORCH_CHECK(mat_a.dim() == 2 || mat_a.dim() == 3, "mat_a has to be 2 or 3d");
-  TORCH_CHECK(mat_b.dim() == 2 || mat_b.dim() == 3, "mat_b has to be 2 or 3d");
-  const bool a_is_2d = mat_a.dim() == 2;
-  const bool b_is_2d = mat_b.dim() == 2;
-  if (!a_is_2d || !b_is_2d) {
-    TORCH_CHECK(mat_a.size(-1) == mat_b.size(-2), "contraction dimension of mat_a and mat_b must match");
-  }
-
-  // check that the strides are valid, the fn will throw an error if not
-  check_valid_strides_and_return_transposed(mat_a);
-  check_valid_strides_and_return_transposed(mat_b);
-  TORCH_CHECK(offs.has_value() ==  (a_is_2d || b_is_2d), "Have to provide offsets if there is a 2d matrix, or no offset if both matrices are 3d");
-
-  if (offs.has_value()) {
-    TORCH_CHECK(offs->dim() == 1, "offs has to be 1D");
-    TORCH_CHECK(offs->dtype() == at::kInt, "Offsets have to be int32");
-  }
-  TORCH_CHECK(!bias.has_value(), "Bias not supported yet");
-}
-
-inline c10::ScalarType _resolve_grouped_mm_out_dtype(const Tensor& mat_a, const Tensor& mat_b,
-std::optional<c10::ScalarType> out_dtype) {
-  const auto out_dtype_ = out_dtype.value_or(mat_a.scalar_type());
-  // TODO(future PR): enable float32 output dtype for bfloat16 and float16 inputs
-  TORCH_CHECK(out_dtype_ == mat_a.dtype(), "Grouped gemm output dtype must match `mat_a` dtype");
-  return out_dtype_;
-}
-
-
-inline void _grouped_mm_fallback(const Tensor& mat_a, const Tensor& mat_b,
-const std::optional<at::Tensor>& offs,
-const std::optional<at::Tensor>& bias,
-std::optional<c10::ScalarType> out_dtype,
-Tensor out) {
-  LOG(INFO) << "fallback path for `torch._grouped_mm`, performance may not be optimal";
-  const bool a_is_2d = mat_a.dim() == 2;
-  const bool b_is_2d = mat_b.dim() == 2;
-  if (a_is_2d && !b_is_2d) {
-    // 2d x 3d with offsets
-    int group_start_idx = 0;
-    auto offs_cpu = offs.value().cpu();
-    for (int group_idx = 0; group_idx < offs_cpu.size(0); group_idx++) {
-      int group_end_idx = offs_cpu[group_idx].item<int>();
-      auto mat_a_slice = mat_a.slice(0, group_start_idx, group_end_idx);
-      auto out_slice = out.slice(0, group_start_idx, group_end_idx);
-      at::mm_out(out_slice, mat_a_slice, mat_b[group_idx]);
-      group_start_idx = group_end_idx;
-    }
-
-  } else if (!a_is_2d && b_is_2d) {
-    // 3d x 2d with offsets
-    int group_start_idx = 0;
-    auto offs_cpu = offs.value().cpu();
-    for (int group_idx = 0; group_idx < offs_cpu.size(0); group_idx++) {
-      int group_end_idx = offs_cpu[group_idx].item<int>();
-      auto mat_b_slice = mat_b.slice(1, group_start_idx, group_end_idx);
-      auto out_slice = out.slice(1, group_start_idx, group_end_idx);
-      at::mm_out(out_slice, mat_a[group_idx], mat_b_slice);
-      group_start_idx = group_end_idx;
-    }
-
-  } else if (a_is_2d && b_is_2d) {
-    // 2d x 2d with offsets
-    int group_start_idx = 0;
-    auto offs_cpu = offs.value().cpu();
-    for (int group_idx = 0; group_idx < offs_cpu.size(0); group_idx++) {
-      int group_end_idx = offs_cpu[group_idx].item<int>();
-      auto mat_a_slice = mat_a.slice(1, group_start_idx, group_end_idx);
-      auto mat_b_slice = mat_b.slice(0, group_start_idx, group_end_idx);
-      auto out_slice = out[group_idx];
-      at::mm_out(out_slice, mat_a_slice, mat_b_slice);
-      group_start_idx = group_end_idx;
-    }
-
-  } else {
-    // 3d x 3d without offsets - regular bmm
-    at::bmm_out(out, mat_a, mat_b);
-  }
-}
-
-
-} // namespace at::native

aten/src/ATen/native/LinearAlgebra.cpp

@@ -1360,8 +1360,7 @@ Tensor outer(const Tensor& self, const Tensor& vec2) {
 #endif
 
 
-#if !defined(__aarch64__) || AT_MKLDNN_ACL_ENABLED()
-// Used by default on x86 platforms and on AArch64+ACL
+#if defined(__aarch64__) && AT_MKLDNN_ACL_ENABLED()
 static inline int64_t get_mkldnn_matmul_min_dim() {
   static auto value = [&] {
     const int64_t default_min_dim = [&] {
@@ -1396,6 +1395,8 @@ static inline bool apply_mkldnn_matmul_heur(int64_t m, int64_t k, int64_t n) {
   return at::globalContext().userEnabledMkldnn() && m > min_dim && k > min_dim && n > min_dim && m * k * n > min_size;
 }
 #endif
+
+
 static void addmm_impl_cpu_(
     Tensor &result, const Tensor &self, Tensor m1, Tensor m2, const Scalar& beta, const Scalar& alpha) {
   TORCH_INTERNAL_ASSERT(self.dim() == 2 && m1.dim() == 2 && m2.dim() == 2);
@@ -1771,8 +1772,8 @@ static inline void bmm_out_or_baddbmm_(const Tensor& self_or_result_, const Tens
     return (strides[2] == 1 && (sizes[1] == 1 || strides[1] >= sizes[2])) ||
         (strides[1] == 1 && (sizes[2] == 1 || strides[2] >= sizes[1]));
   };
-#if !defined(__aarch64__) || AT_MKLDNN_ACL_ENABLED()
-  // Always apply mkldnn heuristic on x86 platform, but on ARM only if compiled with ACL
+
+#if defined(__aarch64__) && AT_MKLDNN_ACL_ENABLED()
   bool apply_heur = apply_mkldnn_matmul_heur(batch1.sizes()[1], batch1.sizes()[2], batch2.sizes()[2]);
   if (apply_heur && use_mkldnn_matmul(batch1, batch2, self_or_result)) {
     try {
@@ -1784,6 +1785,7 @@ static inline void bmm_out_or_baddbmm_(const Tensor& self_or_result_, const Tens
     }
   }
 #endif
+
   if (contraction_size * res_rows * res_cols < 400) {
     if (is_bmm_out) {
       AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kBFloat16, kHalf, batch1.scalar_type(), "bmm", [&] {

aten/src/ATen/native/LossNLL.cpp

@@ -47,14 +47,10 @@ TORCH_META_FUNC(nll_loss_forward)
   TORCH_CHECK(
       target.dim() <= 1,
       "0D or 1D target tensor expected, multi-target not supported");
-  if (self.dim() == 1 && target.dim() == 1) {
-      TORCH_CHECK_VALUE(
-          target.size(0) == 1,
-          "For 1D input, 1D target must have size 1, but got target size: ",
-          target.size(0));
-  }
+
+  auto no_batch_dim = self.dim() == 1  && target.dim() == 0;
   TORCH_CHECK(
-      self.dim() == 1 || (self.size(0) == target.size(0)),
+      no_batch_dim || (self.size(0) == target.size(0)),
       "size mismatch (got input: ",
       self.sizes(),
       ", target: ",

aten/src/ATen/native/Normalization.cpp

@@ -624,9 +624,7 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, int64_t> _batch_norm_impl_index(
   if (backend == BatchNormBackend::Miopen) {
     return std::tuple_cat(
              at::miopen_batch_norm(
-               input.contiguous(input.suggest_memory_format()),
-               weight.contiguous(),
-               bias.contiguous(),
+               input.contiguous(), weight.contiguous(), bias.contiguous(),
                running_mean.defined() ? running_mean.contiguous() : running_mean,
                running_var.defined() ? running_var.contiguous() : running_var,
                training, momentum, eps),

aten/src/ATen/native/Onehot.cpp

@@ -1,6 +1,5 @@
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
 #include <ATen/core/Tensor.h>
-#include <ATen/DTensorState.h>
 
 #ifndef AT_PER_OPERATOR_HEADERS
 #include <ATen/Functions.h>
@@ -25,13 +24,8 @@ Tensor one_hot(const Tensor &self, int64_t num_classes) {
         if (num_classes == -1) {
           num_classes = self.max().item().toLong() + 1;
         }
-        {
-          // If `self` is a DTensor, then allow implicit replication
-          // of the `index` Tensor.
-          at::DTensorAllowImplicitReplication guard;
-          at::Tensor index = at::arange(num_classes, self.options());
-          return at::eq(self.unsqueeze(-1), index).to(kLong);
-        }
+        at::Tensor index = at::arange(num_classes, self.options());
+        return at::eq(self.unsqueeze(-1), index).to(kLong);
     }
 
     auto shape = self.sizes().vec();

aten/src/ATen/native/TensorAdvancedIndexing.cpp

@@ -2174,7 +2174,7 @@ static void _scatter_via_index_put(
   if (self.dim() == 1 || broadcast_index) {
     Tensor squeezed = index;
     if (broadcast_index && index.dim() > 1) {
-      for (int64_t d = index.dim() - 1; d >= 0; --d) {
+      for (const auto d : c10::irange(index.dim())) {
         if (d == dim) {
           continue;
         }