[async_tp] Base support ag-transpose-mm(mat_B) case

ghstack-source-id: edd51b9c46e46e8eca0c45e0ea53c1b26b375c01
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163069

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -5,9 +5,9 @@ GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
 # Set CUDA architecture lists to match x86 build_cuda.sh
 if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
+    export TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;8.0;9.0"
 elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+    export TORCH_CUDA_ARCH_LIST="7.0;8.0;9.0;10.0;12.0"
 elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
 fi
@@ -31,7 +31,8 @@ pip install -r /pytorch/requirements.txt
 pip install auditwheel==6.2.0 wheel
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
     export USE_SYSTEM_NCCL=1
@@ -41,9 +42,13 @@ else
         echo "Bundling CUDA libraries with wheel for aarch64."
     else
         echo "Using nvidia libs from pypi for aarch64."
+        # Fix platform constraints in PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64
+        # Replace 'platform_machine == "x86_64"' with 'platform_machine == "aarch64"'
+        export PYTORCH_EXTRA_INSTALL_REQUIREMENTS="${PYTORCH_EXTRA_INSTALL_REQUIREMENTS//platform_machine == \'x86_64\'/platform_machine == \'aarch64\'}"
         echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
         export USE_NVIDIA_PYPI_LIBS=1
     fi
 
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -138,8 +138,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     folder = os.path.dirname(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}")
 
     # Check if we should use PyPI NVIDIA libraries or bundle system libraries
     use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
@@ -213,8 +211,7 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
         ]
 
         # CUDA version-specific libraries
-        if "13" in desired_cuda:
-            minor_version = desired_cuda[-1]
+        if "130" in desired_cuda:
             version_specific_libs = [
                 "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
                 "/usr/local/cuda/lib64/libcublas.so.13",
@@ -224,7 +221,7 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
                 "/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}",
+                "/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")
@@ -240,8 +237,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
                 "/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
@@ -280,7 +275,14 @@ 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(
@@ -317,7 +319,7 @@ if __name__ == "__main__":
     ).decode()
 
     print("Building PyTorch wheel")
-    build_vars = ""
+    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
     # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
     if enable_cuda:
         build_vars += "MAX_JOBS=5 "

.ci/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/executorch.txt

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

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

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

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

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

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-5ae38bdb0dc066c5823e34dc9797afb9de42c866
+fccfc522864cf8bc172abe0cd58ae5581e2d44b9

.ci/docker/common/install_executorch.sh

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

.ci/docker/common/install_rocm.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/requirements-ci.txt

@@ -93,9 +93,8 @@ librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
 #Pinned versions:
 #test that import:
 
-mypy==1.16.0 ; platform_system != "Windows"
+mypy==1.16.0
 # Pin MyPy version because new errors are likely to appear with each release
-# Skip on Windows as lots of type annotations are POSIX specific
 #Description: linter
 #Pinned versions: 1.16.0
 #test that import: test_typing.py, test_type_hints.py

.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

@@ -41,6 +41,7 @@ def sample_vllm_test_library():
                 "pytest -v -s basic_correctness/test_cumem.py",
                 "pytest -v -s basic_correctness/test_basic_correctness.py",
                 "pytest -v -s basic_correctness/test_cpu_offload.py",
+                "VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py",
             ],
         },
         "vllm_basic_models_test": {
@@ -67,11 +68,14 @@ def sample_vllm_test_library():
                         "-v",
                         "-s",
                         "entrypoints/llm",
+                        "--ignore=entrypoints/llm/test_lazy_outlines.py",
                         "--ignore=entrypoints/llm/test_generate.py",
+                        "--ignore=entrypoints/llm/test_generate_multiple_loras.py",
                         "--ignore=entrypoints/llm/test_collective_rpc.py",
                     ]
                 ),
-                "pytest -v -s entrypoints/llm/test_generate.py",
+                "pytest -v -s entrypoints/llm/test_lazy_outlines.py",
+                "pytest -v -s entrypoints/llm/test_generate.py ",
                 "VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode",
             ],
         },

.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):
@@ -131,11 +125,6 @@ class VllmTestRunner(BaseRunner):
             # 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)
-
     def _install_wheels(self, params: VllmTestParameters):
         logger.info("Running vllm test with inputs: %s", params)
         if not pkg_exists("torch"):

.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/functorch_doc_push_script.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+# This is where the local pytorch install in the docker image is located
+pt_checkout="/var/lib/jenkins/workspace"
+source "$pt_checkout/.ci/pytorch/common_utils.sh"
+echo "functorch_doc_push_script.sh: Invoked with $*"
+
+set -ex -o pipefail
+
+version=${DOCS_VERSION:-nightly}
+echo "version: $version"
+
+# Build functorch docs
+pushd $pt_checkout/functorch/docs
+make html
+popd
+
+git clone https://github.com/pytorch/functorch -b gh-pages --depth 1 functorch_ghpages
+pushd functorch_ghpages
+
+if [ "$version" == "main" ]; then
+  version=nightly
+fi
+
+git rm -rf "$version" || true
+mv "$pt_checkout/functorch/docs/build/html" "$version"
+
+git add "$version" || true
+git status
+git config user.email "soumith+bot@pytorch.org"
+git config user.name "pytorchbot"
+# If there aren't changes, don't make a commit; push is no-op
+git commit -m "Generate Python docs from pytorch/pytorch@${GITHUB_SHA}" || true
+git status
+
+if [[ "${WITH_PUSH:-}" == true ]]; then
+  git push -u origin gh-pages
+fi
+
+popd

.ci/pytorch/numba-cuda-13.patch

@@ -1,25 +0,0 @@
-From 6e08c9d08e9de59c7af28b720289debbbd384764 Mon Sep 17 00:00:00 2001
-From: Michael Wang <13521008+isVoid@users.noreply.github.com>
-Date: Tue, 1 Apr 2025 17:28:05 -0700
-Subject: [PATCH] Avoid bumping certain driver API to avoid future breakage
- (#185)
-
-Co-authored-by: isVoid <isVoid@users.noreply.github.com>
----
- numba_cuda/numba/cuda/cudadrv/driver.py | 3 +++
- 1 file changed, 3 insertions(+)
-
-diff --git a/numba_cuda/numba/cuda/cudadrv/driver.py b/numba_cuda/numba/cuda/cudadrv/driver.py
-index 1641bf77..233e9ed7 100644
---- a/numba_cuda/numba/cuda/cudadrv/driver.py
-+++ b/numba_cuda/numba/cuda/cudadrv/driver.py
-@@ -365,6 +365,9 @@ def _find_api(self, fname):
-         else:
-             variants = ('_v2', '')
- 
-+        if fname in ("cuCtxGetDevice", "cuCtxSynchronize"):
-+            return getattr(self.lib, fname)
-+
-         for variant in variants:
-             try:
-                 return getattr(self.lib, f'{fname}{variant}')

.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

@@ -32,16 +32,6 @@ if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /v
   git config --global --add safe.directory /var/lib/jenkins/workspace
 fi
 
-
-# Patch numba to avoid CUDA-13 crash, see https://github.com/pytorch/pytorch/issues/162878
-NUMBA_CUDA_DIR=$(python -c "import os;import numba.cuda; print(os.path.dirname(numba.cuda.__file__))" 2>/dev/null || true)
-if [ -n "$NUMBA_CUDA_DIR" ]; then
-  NUMBA_PATCH="$(dirname "$(realpath "${BASH_SOURCE[0]}")")/numba-cuda-13.patch"
-  pushd "$NUMBA_CUDA_DIR"
-  patch -p4 <"$NUMBA_PATCH"
-  popd
-fi
-
 echo "Environment variables:"
 env
 
@@ -1550,10 +1540,14 @@ test_executorch() {
   install_torchvision
   install_torchaudio
 
-  INSTALL_SCRIPT="$(pwd)/.ci/docker/common/install_executorch.sh"
-
   pushd /executorch
-  "${INSTALL_SCRIPT}" setup_executorch
+
+  export PYTHON_EXECUTABLE=python
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
+
+  # NB: We need to rebuild ExecuTorch runner here because it depends on PyTorch
+  # from the PR
+  bash .ci/scripts/setup-linux.sh --build-tool cmake
 
   echo "Run ExecuTorch unit tests"
   pytest -v -n auto
@@ -1567,6 +1561,10 @@ test_executorch() {
 
   popd
 
+  # Test torchgen generated code for Executorch.
+  echo "Testing ExecuTorch op registration"
+  "$BUILD_BIN_DIR"/test_edge_op_registration
+
   assert_git_not_dirty
 }
 
@@ -1574,7 +1572,6 @@ test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
         test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
-        distributed/elastic/timer/api_test distributed/elastic/timer/local_timer_example distributed/elastic/timer/local_timer_test \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1724,6 +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/win-test-helpers/build_pytorch.bat

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

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

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

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

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

.ci/pytorch/win-test.sh

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

.ci/pytorch/windows/internal/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"
 
 ###########################################################
 
@@ -125,6 +125,7 @@ popd
 export TH_BINARY_BUILD=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
 export MACOSX_DEPLOYMENT_TARGET=11.0
+export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
@@ -132,19 +133,25 @@ RENAME_WHEEL=true
 case $desired_python in
     3.14t)
         echo "Using 3.14 deps"
-        mac_version='macosx-11.0-arm64'
         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'
         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"
         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)
@@ -169,12 +176,17 @@ case $desired_python in
         ;;
 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=(
     "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 +200,7 @@ 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" --plat-name ${mac_version}
 
 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

@@ -21,7 +21,6 @@ self-hosted-runner:
     - 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/reuse-old-whl/reuse_old_whl.py

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

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

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

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+27fc2493d383354a008106f22f3be232badee9a1

.github/ci_commit_pins/fbgemm_rocm.txt

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

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-9d1c50a5ac8726f4af0d4a4e85ad4d26a674ad26
+e10fef08838612b4560e9c72e5cb1414a5edfa13

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-c77852e117bdf056c8e9a087e51d6f65cf6ba53d
+6c5478ff7c3d50dd1e3047d72ec5909bea474073

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -82,10 +82,16 @@ 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; \
+        dnf install -y git curl wget sudo vim; \
     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 $CUDA_VERSION | cut -d. -f1,2)/compat/
+
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version >/dev/null 2>&1; then \
@@ -214,16 +220,11 @@ ARG SCCACHE_S3_NO_CREDENTIALS=0
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$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=${SCCACHE_S3_NO_CREDENTIALS} \
@@ -284,7 +285,7 @@ RUN if command -v apt-get >/dev/null; then \
         && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
         && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
     else \
-        dnf install -y git curl wget sudo; \
+        dnf install -y git curl wget sudo vim; \
     fi \
     && python3 --version && python3 -m pip --version
 
@@ -297,6 +298,12 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
     echo "[INFO] 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 \

.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/labeler.yml

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

.github/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/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

@@ -43,55 +43,55 @@ CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-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'"
     ),
     "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==9.13.0.50; 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 | "

.github/scripts/generate_ci_workflows.py

@@ -135,7 +135,7 @@ ROCM_SMOKE_WORKFLOWS = [
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
             arches=["6.4"],
-            python_versions=["3.10"],
+            python_versions=["3.9"],
         ),
         ciflow_config=CIFlowConfig(
             labels={

.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/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 +61,23 @@ 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" >> "${GITHUB_PATH}"
       !{{ common.checkout(deep_clone=False, directory="pytorch") }}
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -94,6 +94,8 @@ jobs:
 {%- if config["package_type"] == "wheel" %}
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -104,9 +106,33 @@ 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/_binary-test-linux.yml

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

.github/workflows/_docs.yml

@@ -75,6 +75,10 @@ jobs:
             runner: ${{ inputs.runner_prefix }}linux.2xlarge
             # It takes less than 30m to finish python docs unless there are issues
             timeout-minutes: 30
+          - docs_type: functorch
+            runner: ${{ inputs.runner_prefix }}linux.2xlarge
+            # It takes less than 15m to finish functorch docs unless there are issues
+            timeout-minutes: 15
     # Set a fixed name for this job instead of using the current matrix-generated name, i.e. build-docs (cpp, linux.12xlarge, 180)
     # The current name requires updating the database last docs push query from test-infra every time the matrix is updated
     name: build-docs-${{ matrix.docs_type }}-${{ inputs.push }}
@@ -207,6 +211,16 @@ jobs:
           path: cppdocs/
           s3-prefix: pytorch/pytorch/${{ github.event.pull_request.number }}/cppdocs
 
+      - name: Upload functorch Docs Preview
+        uses: seemethere/upload-artifact-s3@baba72d0712b404f646cebe0730933554ebce96a # v5.1.0
+        if: ${{ github.event_name == 'pull_request' && matrix.docs_type == 'functorch' && steps.build-docs.outcome == 'success' }}
+        with:
+          retention-days: 14
+          s3-bucket: doc-previews
+          if-no-files-found: error
+          path: functorch_ghpages/nightly/
+          s3-prefix: pytorch/pytorch/${{ github.event.pull_request.number }}/functorchdocs
+
       - name: Teardown Linux
         uses: pytorch/test-infra/.github/actions/teardown-linux@main
         if: always()

.github/workflows/_linux-test.yml

@@ -169,7 +169,7 @@ jobs:
         id: install-nvidia-driver
         uses: pytorch/test-infra/.github/actions/setup-nvidia@main
         with:
-          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '580.82.07' }}
+          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '570.133.07' }}
         if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && !contains(matrix.runner, 'b200') }}
 
       - name: Setup GPU_FLAG for docker run

.github/workflows/_rocm-test.yml

@@ -62,11 +62,6 @@ on:
         required: false
         type: number
         default: 1
-    secrets:
-      HUGGING_FACE_HUB_TOKEN:
-        required: false
-        description: |
-          HF Auth token to avoid rate limits when downloading models or datasets from hub
 env:
   GIT_DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
 
@@ -81,9 +76,10 @@ jobs:
     strategy:
       matrix: ${{ fromJSON(inputs.test-matrix) }}
       fail-fast: false
-    runs-on: ${{ matrix.runner }}
     timeout-minutes: ${{ matrix.mem_leak_check == 'mem_leak_check' && 600 || inputs.timeout-minutes }}
+    runs-on: ${{ matrix.runner }}
     steps:
+      # [see note: pytorch repo ref]
       - name: Checkout PyTorch
         uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
         with:
@@ -135,9 +131,6 @@ jobs:
 
       - name: Start monitoring script
         id: monitor-script
-        if: ${{ !inputs.disable-monitor }}
-        shell: bash
-        continue-on-error: true
         env:
           JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
           JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
@@ -145,6 +138,9 @@ jobs:
           WORKFLOW_RUN_ID: ${{github.run_id}}
           MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
+        if: ${{ !inputs.disable-monitor }}
+        shell: bash
+        continue-on-error: true
         run: |
           python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7
           python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
@@ -182,12 +178,6 @@ jobs:
         run: |
           echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
 
-      - name: Preserve github env variables for use in docker
-        shell: bash
-        run: |
-          env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-          env | grep '^CI' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-
       - name: Test
         id: test
         env:
@@ -203,22 +193,20 @@ jobs:
           JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
           BRANCH: ${{ steps.parse-ref.outputs.branch }}
           SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha || github.sha }}
-          TEST_CONFIG: ${{ matrix.config }}
-          SHARD_NUMBER: ${{ matrix.shard }}
-          NUM_TEST_SHARDS: ${{ matrix.num_shards }}
-          REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
           CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
           VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
           TEST_SHOWLOCALS: ${{ steps.keep-going.outputs.ci-test-showlocals }}
           NO_TEST_TIMEOUT: ${{ steps.keep-going.outputs.ci-no-test-timeout }}
           NO_TD: ${{ steps.keep-going.outputs.ci-no-td }}
+          TEST_CONFIG: ${{ matrix.config }}
+          SHARD_NUMBER: ${{ matrix.shard }}
+          NUM_TEST_SHARDS: ${{ matrix.num_shards }}
+          REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
           DOCKER_IMAGE: ${{ inputs.docker-image }}
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
           PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
           TESTS_TO_INCLUDE: ${{ inputs.tests-to-include }}
           DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
         timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
         run: |
           set -x
@@ -248,7 +236,6 @@ jobs:
             -e GITHUB_RUN_ATTEMPT \
             -e JOB_ID \
             -e JOB_NAME \
-            -e BASE_SHA \
             -e BRANCH \
             -e SHA1 \
             -e AWS_DEFAULT_REGION \
@@ -266,12 +253,10 @@ jobs:
             -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
             -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
             -e TESTS_TO_INCLUDE \
-            -e HUGGING_FACE_HUB_TOKEN \
             -e DASHBOARD_TAG \
             --env-file="${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --ulimit core=0 \
-            --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --security-opt seccomp=unconfined \
             --cap-add=SYS_PTRACE \
             --shm-size="8g" \

.github/workflows/_win-build.yml

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

.github/workflows/_win-test.yml

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

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

@@ -12,9 +12,6 @@ on:
     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' }}
@@ -27,33 +24,21 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [ '3.12' ]
-        # TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
-        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
+        # TODO (huydhn): Add cu130 https://github.com/pytorch/pytorch/pull/162000#issuecomment-3261541554
         device: [ 'cu128', 'cu129' ]
+        runner: [ 'linux.12xlarge.memory' ]
         include:
-          - platform: manylinux_2_28_x86_64
-            device: cu128
+          - device: cu128
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.8'
-            runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_x86_64
-            device: cu129
+          - 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 }}"
+    name: "Build ${{ matrix.device }} vLLM wheel"
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 480
     env:
       PY_VERS: ${{ matrix.python-version }}
       MANYLINUX_IMAGE: ${{ matrix.manylinux-image }}
-      PLATFORM: ${{ matrix.platform }}
+      PLATFORM: 'manylinux_2_28_x86_64'
       BUILD_DEVICE: ${{ matrix.device }}
     steps:
       - name: Setup SSH (Click me for login details)
@@ -74,6 +59,20 @@ jobs:
         run: |
           set -eux
 
+          # 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 \
+            -v "${GITHUB_WORKSPACE}:/pytorch" \
+            -v "${RUNNER_TEMP}/artifacts:/artifacts" \
+            -w /artifacts/ \
+            "${MANYLINUX_IMAGE}"
+          )
+
           # Determine python executable for given version (copied from build-triton-wheel)
           case $PY_VERS in
           3.10)
@@ -103,21 +102,6 @@ jobs:
             ;;
           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/${BUILD_DEVICE}"
@@ -129,6 +113,7 @@ jobs:
             --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
 
           # Save this for later
+          echo "PYTHON_EXECUTABLE=${PYTHON_EXECUTABLE}" >> "$GITHUB_ENV"
           echo "container_name=${container_name}" >> "$GITHUB_ENV"
 
       - name: Build vLLM wheel
@@ -146,12 +131,41 @@ jobs:
           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}" bash -c "
+            set -eux
+
+            nightly=\$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2 | cut -d'.' -f4)
+
+            pushd externals/vllm/wheels
+            for package in xformers flashinfer-python vllm; do
+              pushd \$package
+              auditwheel repair --plat \$PLATFORM *.whl \
+                --exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
+              repair_wheel=\$(find wheelhouse -name *\${PLATFORM}*)
+              repair_wheel=\$(basename \${repair_wheel})
+              popd
+
+              cp \${package}/wheelhouse/\${repair_wheel} .
+              version=\$(unzip -p \$repair_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
+
+              if [[ \$package == vllm ]]; then
+                new_wheel=\${repair_wheel/\$version/1.0.0.\$nightly}
+              else
+                major_version=\$(echo \$version | tr '.+' '.' | cut -d'.' -f1-3)
+                new_wheel=\${repair_wheel/\$version/\$major_version.\$nightly}
+              fi
+
+              mv -- \$repair_wheel \$new_wheel
+              rm -rf \$package
+            done
+            popd
+          "
+
           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 }}-${{ matrix.python-version }}
+          name: vllm-wheel-${{ matrix.device }}-${{ matrix.python-version }}-${{ env.PLATFORM }}
           if-no-files-found: error
           path: ${{ runner.temp }}/artifacts/externals/vllm/wheels/*.whl
 
@@ -161,29 +175,27 @@ jobs:
 
   # Copied from build-triton-wheel workflow (mostly)
   upload-wheel:
-    name: "Upload ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
+    name: "Upload ${{ matrix.device }} vLLM wheel"
     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') || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && 'nightly-wheel-upload' || '' }}
+    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') || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
+        if: ${{ github.event_name == 'push' && github.event.ref == 'refs/heads/main' }}
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
         with:
           role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_nightly_build_wheels
@@ -207,15 +219,15 @@ jobs:
         run: |
           set -eux
           mkdir -p "${RUNNER_TEMP}/artifacts/"
-          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-"${PLATFORM}"-*/* "${RUNNER_TEMP}/artifacts/"
+          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-*/* "${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' }}
+      - name: Set DRY_RUN (only for tagged pushes)
+        if: ${{ github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v')) }}
         shell: bash
         run: |
           echo "DRY_RUN=disabled" >> "$GITHUB_ENV"
 
-      - name: Set UPLOAD_CHANNEL
+      - name: Set UPLOAD_CHANNEL (only for tagged pushes)
         if: ${{ github.event_name == 'push' && startsWith(github.event.ref, 'refs/tags/v') }}
         shell: bash
         run: |

.github/workflows/docker-builds.yml

@@ -71,7 +71,8 @@ jobs:
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter,
-          pytorch-linux-jammy-py3-clang12-executorch,
+          # Executorch pin needs update
+          # pytorch-linux-jammy-py3-clang12-executorch,
           pytorch-linux-jammy-py3.12-triton-cpu,
           pytorch-linux-noble-riscv64-py3.12-gcc14
         ]

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

@@ -132,7 +132,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -178,7 +178,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -224,7 +224,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_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==9.13.0.50; 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 }}
@@ -335,7 +335,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -381,7 +381,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -427,7 +427,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_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==9.13.0.50; 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 }}
@@ -538,7 +538,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -584,7 +584,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -630,7 +630,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_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==9.13.0.50; 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 }}
@@ -741,7 +741,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -787,7 +787,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -833,7 +833,7 @@ 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==9.13.0.50; 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 }}
@@ -944,7 +944,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -990,7 +990,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1036,7 +1036,7 @@ 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==9.13.0.50; 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 }}
@@ -1147,7 +1147,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1193,7 +1193,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1239,7 +1239,7 @@ 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==9.13.0.50; 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 }}
@@ -1350,7 +1350,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1396,7 +1396,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1442,7 +1442,7 @@ 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==9.13.0.50; 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-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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_8-test:  # Testing
@@ -259,7 +259,7 @@ jobs:
       runner_prefix: "${{ 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==9.13.0.50; 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda13_0-test:  # Testing
@@ -719,7 +719,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_6-test:  # Testing
@@ -785,7 +785,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_8-test:  # Testing
@@ -851,7 +851,7 @@ jobs:
       runner_prefix: "${{ 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==9.13.0.50; 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda13_0-test:  # Testing
@@ -1311,7 +1311,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_6-test:  # Testing
@@ -1377,7 +1377,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing
@@ -1443,7 +1443,7 @@ jobs:
       runner_prefix: "${{ 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==9.13.0.50; 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda13_0-test:  # Testing
@@ -1903,7 +1903,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_6-test:  # Testing
@@ -1969,7 +1969,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_8-test:  # Testing
@@ -2035,7 +2035,7 @@ jobs:
       runner_prefix: "${{ 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==9.13.0.50; 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda13_0-test:  # Testing
@@ -2495,7 +2495,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_6-test:  # Testing
@@ -2561,7 +2561,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_8-test:  # Testing
@@ -2627,7 +2627,7 @@ jobs:
       runner_prefix: "${{ 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==9.13.0.50; 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda13_0-test:  # Testing
@@ -3087,7 +3087,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_6-test:  # Testing
@@ -3153,7 +3153,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_8-test:  # Testing
@@ -3219,7 +3219,7 @@ jobs:
       runner_prefix: "${{ 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==9.13.0.50; 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda13_0-test:  # Testing
@@ -3679,7 +3679,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_6-test:  # Testing
@@ -3745,7 +3745,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_8-test:  # Testing
@@ -3811,7 +3811,7 @@ jobs:
       runner_prefix: "${{ 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==9.13.0.50; 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

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

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

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

@@ -60,13 +60,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -81,9 +81,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 +56,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -77,9 +77,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -95,6 +99,8 @@ jobs:
           "${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 +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
@@ -166,13 +196,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -187,9 +217,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -205,6 +239,8 @@ jobs:
           "${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 +251,33 @@ 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 +336,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -297,9 +357,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -315,6 +379,8 @@ jobs:
           "${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 +391,33 @@ 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 +476,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -407,9 +497,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -425,6 +519,8 @@ jobs:
           "${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 +531,33 @@ 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 +616,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -517,9 +637,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -535,6 +659,8 @@ jobs:
           "${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 +671,33 @@ 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 +756,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -627,9 +777,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -645,6 +799,8 @@ jobs:
           "${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 +811,33 @@ 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 +896,13 @@ 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" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -737,9 +917,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${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 "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -755,6 +939,8 @@ jobs:
           "${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 +951,33 @@ 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/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-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/pull.yml

@@ -127,8 +127,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # More memory is needed to build with asan
-      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
@@ -318,6 +316,32 @@ jobs:
         ]}
     secrets: inherit
 
+  linux-jammy-py3-clang12-executorch-build:
+    if: false  # Docker build needs pin update
+    name: linux-jammy-py3-clang12-executorch
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-py3-clang12-executorch
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang12-executorch
+      test-matrix: |
+        { include: [
+          { config: "executorch", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-py3-clang12-executorch-test:
+    name: linux-jammy-py3-clang12-executorch
+    uses: ./.github/workflows/_linux-test.yml
+    needs: linux-jammy-py3-clang12-executorch-build
+    if: false # Has been broken for a while
+    with:
+      build-environment: linux-jammy-py3-clang12-executorch
+      docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
+    secrets: inherit
+
   linux-jammy-cuda12_8-py3_10-gcc9-inductor-build:
     name: cuda12.8-py3.10-gcc9-sm75
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/slow.yml

@@ -140,8 +140,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # More memory is needed to build with asan
-      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

.github/workflows/trunk.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: [
@@ -255,31 +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: ${{ needs.verify-cachebench-cpu-build.outputs.test-matrix }}
     secrets: inherit
-
-  linux-jammy-py3-clang12-executorch-build:
-    name: linux-jammy-py3-clang12-executorch
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3-clang12-executorch
-      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang12-executorch
-      test-matrix: |
-        { include: [
-          { config: "executorch", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-py3-clang12-executorch-test:
-    name: linux-jammy-py3-clang12-executorch
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-py3-clang12-executorch-build
-    with:
-      build-environment: linux-jammy-py3-clang12-executorch
-      docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/vllm.yml

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

.gitignore

@@ -259,9 +259,6 @@ gen
 .pytest_cache
 aten/build/*
 
-# Linker scripts for prioritized text optimization
-cmake/linker_script.ld
-
 # Bram
 plsdontbreak
 
@@ -392,5 +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/**',
@@ -123,7 +123,6 @@ is_formatter = true
 code = 'MYPY'
 include_patterns = [
     'setup.py',
-    'functorch/dim/**/*.py',
     'torch/**/*.py',
     'torch/**/*.pyi',
     'caffe2/**/*.py',
@@ -965,6 +964,7 @@ exclude_patterns = [
     'test/jit/**',  # should be run through test/test_jit.py
     'test/ao/sparsity/**',  # should be run through test/test_ao_sparsity.py
     'test/fx/**',  # should be run through test/test_fx.py
+    'test/bottleneck_test/**',  # excluded by test/run_test.py
     'test/package/**',  # excluded by test/run_test.py
     'test/distributed/argparse_util_test.py',
     'test/distributed/bin/test_script.py',
@@ -1410,6 +1410,8 @@ exclude_patterns = [
     'torch/utils/benchmark/utils/timer.py',
     'torch/utils/benchmark/utils/valgrind_wrapper/__init__.py',
     'torch/utils/benchmark/utils/valgrind_wrapper/timer_interface.py',
+    'torch/utils/bottleneck/__init__.py',
+    'torch/utils/bottleneck/__main__.py',
     'torch/utils/bundled_inputs.py',
     'torch/utils/checkpoint.py',
     'torch/utils/collect_env.py',
@@ -1566,6 +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",

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)
@@ -380,13 +379,6 @@ cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
                        OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                         "CPU_AARCH64" OFF)
-# prioritized text linker, ON by default for AArch64+Linux, option visible to all AArch64, x86 and ppc64le.
-set(USE_PRIORITIZED_TEXT_DEFAULT OFF)
-if(LINUX AND CPU_AARCH64)
-  set(USE_PRIORITIZED_TEXT_DEFAULT ON)
-endif()
-cmake_dependent_option(USE_PRIORITIZED_TEXT_FOR_LD "Use prioritized text linker for ld."
-  "${USE_PRIORITIZED_TEXT_DEFAULT}" "CPU_INTEL OR CPU_AARCH64 OR CPU_POWER" OFF)
 
 option(USE_MIMALLOC "Use mimalloc" OFF)
 # Enable third party mimalloc library to improve memory allocation performance
@@ -664,11 +656,6 @@ endif(MSVC)
 
 string(APPEND CMAKE_CUDA_FLAGS " -Xfatbin -compress-all")
 
-# Set linker max-page-size to 64KiB on AArch64 Linux
-if(LINUX AND CPU_AARCH64)
-  add_link_options_if_supported("-z,max-page-size=0x10000")
-endif()
-
 # Set INTERN_BUILD_MOBILE for all mobile builds. Components that are not
 # applicable to mobile are disabled by this variable. Setting
 # `BUILD_PYTORCH_MOBILE_WITH_HOST_TOOLCHAIN` environment variable can force it
@@ -886,7 +873,7 @@ 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(
@@ -902,9 +889,9 @@ IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
   set(USE_FBGEMM_GENAI off)
 endif()
 
-# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100.
-if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0" AND CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8 AND NOT WIN32)
-  message(STATUS "Setting USE_FBGEMM_GENAI to ON, doing CUDA build for SM100a")
+# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100
+if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0a")
+  message(WARNING "Setting USE_FBGEMM_GENAI to ON for CUDA build on SM100")
   set(USE_FBGEMM_GENAI ON)
 endif()
 
@@ -921,7 +908,7 @@ 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()
@@ -1433,57 +1420,3 @@ if(BUILD_BUNDLE_PTXAS AND USE_CUDA)
   install(PROGRAMS "${PROJECT_BINARY_DIR}/ptxas"
           DESTINATION "${CMAKE_INSTALL_BINDIR}")
 endif()
-
-if(USE_PRIORITIZED_TEXT_FOR_LD)
-  add_compile_options(
-    $<$<COMPILE_LANGUAGE:C,CXX>:-ffunction-sections>
-    $<$<COMPILE_LANGUAGE:C,CXX>:-fdata-sections>
-  )
-  set(LINKER_SCRIPT_FILE_OUT "${CMAKE_SOURCE_DIR}/cmake/linker_script.ld")
-  set(LINKER_SCRIPT_FILE_IN "${CMAKE_SOURCE_DIR}/cmake/prioritized_text.txt")
-
-  add_custom_command(
-    OUTPUT "${LINKER_SCRIPT_FILE_OUT}"
-    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py --filein "${LINKER_SCRIPT_FILE_IN}" --fout "${LINKER_SCRIPT_FILE_OUT}"
-    DEPENDS ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py "${LINKER_SCRIPT_FILE_IN}"
-    COMMENT "Generating prioritized text linker files"
-    VERBATIM
-  )
-
-  add_custom_target(generate_linker_script DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
-
-  if(BUILD_PYTHON)
-    set(LINKER_OPT_TARGETS torch_python)
-  endif()
-
-  if(NOT BUILD_LIBTORCHLESS)
-    list(APPEND LINKER_OPT_TARGETS torch_cpu c10)
-    if(USE_CUDA)
-      list(APPEND LINKER_OPT_TARGETS torch_cuda c10_cuda)
-    endif()
-    if(USE_XPU)
-      list(APPEND LINKER_OPT_TARGETS torch_xpu c10_xpu)
-    endif()
-    if(USE_ROCM)
-      list(APPEND LINKER_OPT_TARGETS torch_hip c10_hip)
-    endif()
-  endif()
-
-  foreach(tgt IN LISTS LINKER_OPT_TARGETS)
-    if(TARGET ${tgt})
-      add_dependencies("${tgt}" generate_linker_script)
-      target_link_options_if_supported(${tgt} "-T,${LINKER_SCRIPT_FILE_OUT}")
-      set_property(TARGET ${tgt} APPEND PROPERTY LINK_DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
-    else()
-       message(WARNING "Requested target '${tgt}' for linker script optimization was not found.")
-    endif()
-  endforeach()
-
-else()
-  if(LINUX AND CPU_AARCH64)
-    message(WARNING [[
-    It is strongly recommend to enable linker script optimization for all AArch64 Linux builds.
-    To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
-    ]])
-  endif()
-endif()

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

@@ -265,14 +265,6 @@ IF(USE_FBGEMM_GENAI)
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
     list(FILTER fbgemm_genai_native_cuda_cu INCLUDE REGEX ${FBGEMM_CUTLASS_KERNELS_REGEX})
 
-    # 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()
-
     file(GLOB_RECURSE fbgemm_genai_native_cuda_cpp
       "${FBGEMM_GENAI_SRCS}/common/*.cpp"
     )

aten/src/ATen/Context.cpp

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

aten/src/ATen/DLConvertor.cpp

@@ -65,24 +65,14 @@ DLDataType getDLDataType(const Tensor& t) {
       break;
     // TODO(#146647): use macro here instead of spelling out each shell dtype
     case ScalarType::Float8_e5m2:
-      dtype.code = DLDataTypeCode::kDLFloat8_e5m2;
-      break;
     case ScalarType::Float8_e5m2fnuz:
-      dtype.code = DLDataTypeCode::kDLFloat8_e5m2fnuz;
-      break;
     case ScalarType::Float8_e4m3fn:
-      dtype.code = DLDataTypeCode::kDLFloat8_e4m3fn;
-      break;
     case ScalarType::Float8_e4m3fnuz:
-      dtype.code = DLDataTypeCode::kDLFloat8_e4m3fnuz;
-      break;
     case ScalarType::Float8_e8m0fnu:
-      dtype.code = DLDataTypeCode::kDLFloat8_e8m0fnu;
+      TORCH_CHECK_BUFFER(false, "float8 types are not supported by dlpack");
       break;
     case ScalarType::Float4_e2m1fn_x2:
-      dtype.code = DLDataTypeCode::kDLFloat4_e2m1fn;
-      dtype.lanes = 2;
-      dtype.bits = 4;
+      TORCH_CHECK_BUFFER(false, "float4 types are not supported by dlpack");
       break;
     case ScalarType::QInt8:
     case ScalarType::QUInt8:
@@ -187,11 +177,7 @@ static Device getATenDevice(DLDeviceType type, c10::DeviceIndex index, void* dat
 
 ScalarType toScalarType(const DLDataType& dtype) {
   ScalarType stype = ScalarType::Undefined;
-  if (dtype.code != DLDataTypeCode::kDLFloat4_e2m1fn) {
-    TORCH_CHECK_BUFFER(
-        dtype.lanes == 1,
-        "ATen does not support lanes != 1 for dtype code", std::to_string(dtype.code));
-  }
+  TORCH_CHECK_BUFFER(dtype.lanes == 1, "ATen does not support lanes != 1");
   switch (dtype.code) {
     case DLDataTypeCode::kDLUInt:
       switch (dtype.bits) {
@@ -283,73 +269,6 @@ ScalarType toScalarType(const DLDataType& dtype) {
               false, "Unsupported kDLBool bits ", std::to_string(dtype.bits));
       }
       break;
-    case DLDataTypeCode::kDLFloat8_e5m2:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e5m2;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e5m2 bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e5m2fnuz:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e5m2fnuz;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e5m2fnuz bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e4m3fn:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e4m3fn;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e4m3fn bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e4m3fnuz:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e4m3fnuz;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e4m3fnuz bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e8m0fnu:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e8m0fnu;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e8m0fnu bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat4_e2m1fn:
-      switch (dtype.bits) {
-        case 4:
-          switch (dtype.lanes) {
-            case 2:
-              stype = ScalarType::Float4_e2m1fn_x2;
-              break;
-            default:
-              TORCH_CHECK_BUFFER(
-                false, "Unsupported kDLFloat4_e2m1fn lanes ", std::to_string(dtype.lanes));
-          }
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat4_e2m1fn bits ", std::to_string(dtype.bits));
-      }
-      break;
     default:
       TORCH_CHECK_BUFFER(false, "Unsupported code ", std::to_string(dtype.code));
   }
@@ -435,8 +354,8 @@ T* toDLPackImpl(const Tensor& src) {
   atDLMTensor->tensor.dl_tensor.device = torchDeviceToDLDevice(src.device());
   atDLMTensor->tensor.dl_tensor.ndim = static_cast<int32_t>(src.dim());
   atDLMTensor->tensor.dl_tensor.dtype = getDLDataType(src);
-  atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(view.sizes().data());
-  atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(view.strides().data());
+  atDLMTensor->tensor.dl_tensor.shape = view.sizes().data();
+  atDLMTensor->tensor.dl_tensor.strides = view.strides().data();
   atDLMTensor->tensor.dl_tensor.byte_offset = 0;
   fillVersion(&atDLMTensor->tensor);
 

aten/src/ATen/FunctionalStorageImpl.cpp

@@ -102,7 +102,7 @@ FunctionalStorageImpl::FunctionalStorageImpl(const Tensor& base)
   // SparseTensorImpl has no storage, so we cannot query its nbytes.
   // (original_storage_size is only used for storage resizing in fsdp anyway, which does not apply to sparse)
   // Same for XLA
-  if (base.unsafeGetTensorImpl()->has_storage() && data_ptr().device().type() != c10::DeviceType::XLA) {
+  if (base.unsafeGetTensorImpl()->has_storage() && base.device().type() != c10::DeviceType::XLA) {
     original_storage_size_ = base.unsafeGetTensorImpl()->unsafe_storage().unsafeGetStorageImpl()->sym_nbytes();
   } else {
     original_storage_size_ = -1;

aten/src/ATen/FunctionalTensorWrapper.cpp

@@ -133,7 +133,7 @@ FunctionalTensorWrapper::FunctionalTensorWrapper(const Tensor& view_value, const
   : c10::TensorImpl(
       c10::DispatchKeySet(DispatchKey::Functionalize),
       view_value.dtype(),
-      base->storage().data_ptr().device()
+      view_value.device()
     ),
     value_(view_value),
     is_multi_output_view_(base->is_multi_output_view_ || meta.is_multi_output),
@@ -485,10 +485,7 @@ void FunctionalTensorWrapper::shallow_copy_from(const c10::intrusive_ptr<TensorI
 
 
 c10::Device FunctionalTensorWrapper::device_custom() const {
-  // The storage pointer already uses the underlying tensor custom device (if
-  // applicable) to extract the device. So, we dont have to recurse again by
-  // doing value_.unsafeGetTensorImpl()->device().
-  return storage().data_ptr().device();
+  return value_.unsafeGetTensorImpl()->device();
 }
 at::IntArrayRef FunctionalTensorWrapper::sizes_custom() const {
   return value_.unsafeGetTensorImpl()->sizes();

aten/src/ATen/InferSize.h

@@ -94,10 +94,10 @@ inline at::DimVector infer_size_dv(IntArrayRef shape, int64_t numel) {
 
 inline at::SymDimVector infer_size_dv(
     c10::SymIntArrayRef shape,
-    const c10::SymInt& numel) {
+    c10::SymInt numel) {
   auto res = at::SymDimVector(shape);
   infer_size_impl<c10::SymIntArrayRef, c10::SymInt, at::SymDimVector>(
-      shape, numel, res);
+      shape, std::move(numel), res);
   return res;
 }
 

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/core/Dict.h

@@ -6,6 +6,7 @@
 #include <c10/util/TypeList.h>
 #include <c10/util/intrusive_ptr.h>
 #include <c10/util/order_preserving_flat_hash_map.h>
+#include <optional>
 #include <ATen/core/TensorBody.h>
 #include <ATen/core/jit_type_base.h>
 

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

@@ -55,7 +55,8 @@ class TORCH_API CppSignature final {
   }
 
  private:
-  explicit CppSignature(std::type_index signature) : signature_(signature) {}
+  explicit CppSignature(std::type_index signature)
+      : signature_(std::move(signature)) {}
   std::type_index signature_;
 };
 

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

@@ -70,7 +70,7 @@ private:
 void _print_dispatch_trace(const std::string& label, const std::string& op_name, const DispatchKeySet& dispatchKeySet) {
   auto nesting_value = dispatch_trace_nesting_value();
   for (int64_t i = 0; i < nesting_value; ++i) std::cerr << " ";
-  std::cerr << label << " op=[" << op_name << "], key=[" << toString(dispatchKeySet.highestPriorityTypeId()) << "]" << '\n';
+  std::cerr << label << " op=[" << op_name << "], key=[" << toString(dispatchKeySet.highestPriorityTypeId()) << "]" << std::endl;
 }
 } // namespace detail
 
@@ -213,11 +213,9 @@ OperatorHandle Dispatcher::findOrRegisterName_(const OperatorName& op_name) {
 // Windows build doesn't produce the destructor symbol in PyTorch libs
 // causing a linker failure in downstream projects.
 // x-ref https://github.com/pytorch/pytorch/issues/70032
-#if defined(_WIN32)
 OperatorHandle::~OperatorHandle() = default;
-#endif
 
-RegistrationHandleRAII Dispatcher::registerLibrary(const std::string& ns, std::string debug) {
+RegistrationHandleRAII Dispatcher::registerLibrary(std::string ns, std::string debug) {
   std::lock_guard<std::mutex> lock(guard_->mutex);
   auto found = libraries_.find(ns);
   TORCH_CHECK(
@@ -308,7 +306,7 @@ PythonModuleMapType& pythonModulesSingleton() {
 
 }
 
-std::optional<std::pair<const char*, const char*>> Dispatcher::getPyStub(const OperatorName& op_name) {
+std::optional<std::pair<const char*, const char*>> Dispatcher::getPyStub(OperatorName op_name) {
   std::lock_guard<std::mutex> lock(guard_->mutex);
   auto found = pythonModulesSingleton().find(op_name);
   if (found == pythonModulesSingleton().end()) {
@@ -344,7 +342,7 @@ RegistrationHandleRAII Dispatcher::registerPythonModule(
   });
 }
 
-void Dispatcher::throwIfHasPythonModule(const OperatorName& op_name) {
+void Dispatcher::throwIfHasPythonModule(OperatorName op_name) {
   std::lock_guard<std::mutex> lock(guard_->mutex);
   auto elt = pythonModulesSingleton().find(op_name);
   if (elt == pythonModulesSingleton().end()) {
@@ -364,7 +362,7 @@ void Dispatcher::throwIfHasPythonModule(const OperatorName& op_name) {
 }
 
 RegistrationHandleRAII Dispatcher::registerImpl(
-  const OperatorName& op_name,
+  OperatorName op_name,
   std::optional<DispatchKey> dispatch_key,
   KernelFunction kernel,
   std::optional<impl::CppSignature> cpp_signature,
@@ -379,7 +377,7 @@ RegistrationHandleRAII Dispatcher::registerImpl(
     *this,
     dispatch_key,
     std::move(kernel),
-    cpp_signature,
+    std::move(cpp_signature),
     std::move(inferred_function_schema),
     std::move(debug)
   );
@@ -408,7 +406,7 @@ void Dispatcher::deregisterImpl_(const OperatorHandle& op, const OperatorName& o
   cleanup(op, op_name);
 }
 
-RegistrationHandleRAII Dispatcher::registerName(const OperatorName& op_name) {
+RegistrationHandleRAII Dispatcher::registerName(OperatorName op_name) {
   std::lock_guard<std::mutex> lock(guard_->mutex);
   auto op = findOrRegisterName_(op_name);
   ++op.operatorDef_->def_and_impl_count;

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

@@ -13,10 +13,15 @@
 #include <condition_variable>
 #include <list>
 #include <mutex>
+#include <type_traits>
 
 #include <ATen/core/enum_tag.h>
 #include <ATen/core/grad_mode.h>
 
+#ifndef NDEBUG
+#include <iostream>
+#endif
+
 namespace c10 {
 
 TORCH_API bool show_dispatch_trace();
@@ -250,7 +255,7 @@ class TORCH_API Dispatcher final {
   // NB: steals the inferred function schema, as we may need to hold on to
   // it for a bit until the real schema turns up
   RegistrationHandleRAII registerImpl(
-      const OperatorName& op_name,
+      OperatorName op_name,
       std::optional<DispatchKey> dispatch_key,
       KernelFunction kernel,
       std::optional<impl::CppSignature> cpp_signature,
@@ -269,15 +274,15 @@ class TORCH_API Dispatcher final {
   /**
    * Given an operator, throws if we have a pystub.
    */
-  void throwIfHasPythonModule(const OperatorName& op_name);
+  void throwIfHasPythonModule(OperatorName op_name);
 
   std::optional<std::pair<const char*, const char*>> getPyStub(
-      const OperatorName& op_name);
+      OperatorName op_name);
 
   /**
    * Register a new operator by name.
    */
-  RegistrationHandleRAII registerName(const OperatorName& op_name);
+  RegistrationHandleRAII registerName(OperatorName op_name);
 
   /**
    * Register a fallback kernel for a backend.
@@ -295,9 +300,7 @@ class TORCH_API Dispatcher final {
    * API.  These invocations are only permitted once per program, so we raise
    * an error if this is called again for the same namespace.
    */
-  RegistrationHandleRAII registerLibrary(
-      const std::string& ns,
-      std::string debug);
+  RegistrationHandleRAII registerLibrary(std::string ns, std::string debug);
 
   // ------------------------------------------------------------------------
   //
@@ -445,12 +448,8 @@ class TORCH_API OperatorHandle {
   OperatorHandle& operator=(OperatorHandle&&) noexcept = default;
   OperatorHandle(const OperatorHandle&) = default;
   OperatorHandle& operator=(const OperatorHandle&) = default;
-#if defined(_WIN32)
   // NOLINTNEXTLINE(performance-trivially-destructible)
   ~OperatorHandle();
-#else
-  ~OperatorHandle() = default;
-#endif
 
   const OperatorName& operator_name() const {
     return operatorDef_->op.operator_name();

aten/src/ATen/core/function_schema.h

@@ -556,7 +556,7 @@ inline std::ostream& operator<<(std::ostream& out, const Argument& arg) {
   // real_type versus fake_type: in order to be compatible with FunctionSchema
   // parser, printing an argument with either MemoryFormat or Layout type should
   // give us the original schema string, hence printing out real_type.
-  const auto& type = arg.real_type();
+  auto type = arg.real_type();
   bool is_opt = type->kind() == OptionalType::Kind;
   auto unopt_type = is_opt ? type->castRaw<OptionalType>()->getElementType() : type;
 

aten/src/ATen/core/jit_type.h

@@ -232,7 +232,7 @@ struct TORCH_API OptionalType : public UnionType {
   static TypePtr ofTensor();
   //
   // global singleton
-  static TypePtr get(const TypePtr& inner);
+  static TypePtr get(TypePtr inner);
 
  private:
   explicit OptionalType(const TypePtr& contained);
@@ -895,7 +895,7 @@ struct TORCH_API ListType
   // the type List<T>.
   // The extra "identifier" argument is needed beccause we have multiple container types
   // that all re-use this function (List<T>, array<T, N>, etc.)
-  static TypePtr get(const std::string& identifier, const TypePtr& inner);
+  static TypePtr get(const std::string& identifier, TypePtr inner);
 
   // common cast List[Tensor]
   static ListTypePtr ofTensors();

aten/src/ATen/core/type.cpp

@@ -274,7 +274,7 @@ ListTypePtr ListType::ofNumbers() {
   return value;
 }
 
-TypePtr OptionalType::get(const TypePtr& inner) {
+TypePtr OptionalType::get(TypePtr inner) {
   static ska::flat_hash_map<TypePtr, TypePtr> containerTypePtrs;
   static std::mutex mutex;
   // Perf from the lock is ok because this function is guarded behind
@@ -287,7 +287,7 @@ TypePtr OptionalType::get(const TypePtr& inner) {
   return containerTypePtrs[inner];
 }
 
-TypePtr ListType::get(const std::string& identifier, const TypePtr& inner) {
+TypePtr ListType::get(const std::string& identifier, TypePtr inner) {
   static ska::flat_hash_map<std::tuple<std::string, TypePtr>, TypePtr> containerTypePtrs;
   static std::mutex mutex;
   // Perf from the lock is ok because this function is guarded behind

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,
@@ -1954,8 +1950,8 @@ void scaled_gemm(
   #if ROCM_VERSION >= 70000
             if (at::detail::getCUDAHooks().isGPUArch({"gfx950"})) {
                 // TODO: add constraints based on hipblaslt internals
-                TORCH_CHECK((m % 16 == 0) && (n % 16 == 0) && (k % 128 == 0),
-                           "M, N must be multiples of 16 and K should be multiple of 128 for MX format. "
+                TORCH_CHECK((m % 32 == 0) && (n % 32 == 0) && (k % 32 == 0),
+                           "Matrix dimensions must be multiples of 32 for MX format. "
                            "Got m=", m, ", n=", n, ", k=", k);
             }
   #endif

aten/src/ATen/cuda/CUDAGeneratorImpl.cpp

@@ -266,14 +266,11 @@ CUDAGeneratorImpl::CUDAGeneratorImpl(
  * See Note [Acquire lock when using random generators]
  */
 void CUDAGeneratorImpl::set_current_seed(uint64_t seed) {
-  if (C10_LIKELY(at::cuda::currentStreamCaptureStatus() == at::cuda::CaptureStatus::None)) {
-    state_->seed_ = seed;
-    state_->philox_offset_per_thread_ = 0;
-    no_reset_rnn_state_.clear();
-  } else {
-    TORCH_CHECK(state_->seed_ == seed, "CUDAGeneratorImpl::set_current_seed can be called during stream capture only if new seed is the same as the original seed.");
-    // no-op case
-  }
+  at::cuda::assertNotCapturing(
+      "Cannot call CUDAGeneratorImpl::set_current_seed");
+  state_->seed_ = seed;
+  state_->philox_offset_per_thread_ = 0;
+  no_reset_rnn_state_.clear();
 }
 
 /**
@@ -302,6 +299,9 @@ uint64_t CUDAGeneratorImpl::get_offset() const {
  * Gets the current seed of CUDAGeneratorImpl.
  */
 uint64_t CUDAGeneratorImpl::current_seed() const {
+  // Debatable if current_seed() should be allowed in captured regions.
+  // Conservatively disallow it for now.
+  at::cuda::assertNotCapturing("Cannot call CUDAGeneratorImpl::current_seed");
   return state_->seed_;
 }
 
@@ -346,6 +346,8 @@ c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
  * and size of the internal state.
  */
 void CUDAGeneratorImpl::set_state(const c10::TensorImpl& new_state) {
+  at::cuda::assertNotCapturing(
+      "Please ensure to utilize the CUDAGeneratorImpl::set_state_index method during capturing.");
   static const size_t seed_size = sizeof(uint64_t);
   static const size_t offset_size = sizeof(int64_t);
   static const size_t total_size = seed_size + offset_size;
@@ -400,27 +402,15 @@ c10::intrusive_ptr<c10::GeneratorImpl> CUDAGeneratorImpl::graphsafe_get_state()
  */
 void CUDAGeneratorImpl::set_philox_offset_per_thread(uint64_t offset) {
   // see Note [Why enforce RNG offset % 4 == 0?]
-
-  // Note: If you use CUDNN RNN's, calling
-  // set_philox_offset_per_thread instead of set_offset will cause the
-  // cudnn RNN rng state to become stale.
   TORCH_CHECK(offset % 4 == 0, "offset must be a multiple of 4");
-  if (C10_LIKELY(at::cuda::currentStreamCaptureStatus() == at::cuda::CaptureStatus::None)) {
-    state_->philox_offset_per_thread_ = offset;
-  } else {
-    state_->offset_intragraph_ = offset;
-  }
+  state_->philox_offset_per_thread_ = offset;
 }
 
 /**
  * Gets the current philox_offset_per_thread_ of CUDAGeneratorImpl.
  */
 uint64_t CUDAGeneratorImpl::philox_offset_per_thread() const {
-  if (C10_LIKELY(at::cuda::currentStreamCaptureStatus() == at::cuda::CaptureStatus::None)) {
-    return state_->philox_offset_per_thread_;
-  } else {
-    return state_->offset_intragraph_;
-  }
+  return state_->philox_offset_per_thread_;
 }
 
 /**

aten/src/ATen/cuda/detail/DeviceThreadHandles.h

@@ -122,7 +122,7 @@ struct DeviceThreadHandlePool : public std::enable_shared_from_this<DeviceThread
 
     // Called by the destructor.  Releases this thread's handles back into the pool.
     void release() {
-        if(!my_handles.empty()) {
+        if(my_handles.size() > 0) {
             auto parent = weak_parent.lock();
             if (!parent) {
                 // If this thread exits after atexit handlers have completed, the

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/dlpack.h

@@ -19,7 +19,7 @@
 #define DLPACK_MAJOR_VERSION 1
 
 /*! \brief The current minor version of dlpack */
-#define DLPACK_MINOR_VERSION 1
+#define DLPACK_MINOR_VERSION 0
 
 /*! \brief DLPACK_DLL prefix for windows */
 #ifdef _WIN32
@@ -32,7 +32,9 @@
 #define DLPACK_DLL
 #endif
 
+// NOLINTNEXTLINE(modernize-deprecated-headers)
 #include <stdint.h>
+// NOLINTNEXTLINE(modernize-deprecated-headers)
 #include <stddef.h>
 
 #ifdef __cplusplus
@@ -157,26 +159,6 @@ typedef enum {
   kDLComplex = 5U,
   /*! \brief boolean */
   kDLBool = 6U,
-  /*! \brief FP8 data types */
-  kDLFloat8_e3m4 = 7U,
-  kDLFloat8_e4m3 = 8U,
-  kDLFloat8_e4m3b11fnuz = 9U,
-  kDLFloat8_e4m3fn = 10U,
-  kDLFloat8_e4m3fnuz = 11U,
-  kDLFloat8_e5m2 = 12U,
-  kDLFloat8_e5m2fnuz = 13U,
-  kDLFloat8_e8m0fnu = 14U,
-  /*! \brief FP6 data types
-   * Setting bits != 6 is currently unspecified, and the producer must ensure it is set
-   * while the consumer must stop importing if the value is unexpected.
-   */
-  kDLFloat6_e2m3fn = 15U,
-  kDLFloat6_e3m2fn = 16U,
-  /*! \brief FP4 data types
-   * Setting bits != 4 is currently unspecified, and the producer must ensure it is set
-   * while the consumer must stop importing if the value is unexpected.
-   */
-  kDLFloat4_e2m1fn = 17U,
 } DLDataTypeCode;
 
 /*!
@@ -190,12 +172,6 @@ typedef enum {
  *   - int8: type_code = 0, bits = 8, lanes = 1
  *   - std::complex<float>: type_code = 5, bits = 64, lanes = 1
  *   - bool: type_code = 6, bits = 8, lanes = 1 (as per common array library convention, the underlying storage size of bool is 8 bits)
- *   - float8_e4m3: type_code = 8, bits = 8, lanes = 1 (packed in memory)
- *   - float6_e3m2fn: type_code = 16, bits = 6, lanes = 1 (packed in memory)
- *   - float4_e2m1fn: type_code = 17, bits = 4, lanes = 1 (packed in memory)
- *
- *  When a sub-byte type is packed, DLPack requires the data to be in little bit-endian, i.e.,
- *  for a packed data set D ((D >> (i * bits)) && bit_mask) stores the i-th element.
  */
 typedef struct {
   /*!
@@ -253,12 +229,12 @@ typedef struct {
   /*! \brief The data type of the pointer*/
   DLDataType dtype;
   /*! \brief The shape of the tensor */
-  int64_t* shape;
+  const int64_t* shape;
   /*!
    * \brief strides of the tensor (in number of elements, not bytes)
    *  can be NULL, indicating tensor is compact and row-majored.
    */
-  int64_t* strides;
+  const int64_t* strides;
   /*! \brief The offset in bytes to the beginning pointer to data */
   uint64_t byte_offset;
 } DLTensor;
@@ -293,7 +269,7 @@ typedef struct DLManagedTensor {
   void (*deleter)(struct DLManagedTensor * self);
 } DLManagedTensor;
 
-// bit masks used in the DLManagedTensorVersioned
+// bit masks used in in the DLManagedTensorVersioned
 
 /*! \brief bit mask to indicate that the tensor is read only. */
 #define DLPACK_FLAG_BITMASK_READ_ONLY (1UL << 0UL)
@@ -306,14 +282,6 @@ typedef struct DLManagedTensor {
  */
 #define DLPACK_FLAG_BITMASK_IS_COPIED (1UL << 1UL)
 
-/*
- * \brief bit mask to indicate that whether a sub-byte type is packed or padded.
- *
- * The default for sub-byte types (ex: fp4/fp6) is assumed packed. This flag can
- * be set by the producer to signal that a tensor of sub-byte type is padded.
- */
-#define DLPACK_FLAG_BITMASK_IS_SUBBYTE_TYPE_PADDED (1UL << 2UL)
-
 /*!
  * \brief A versioned and managed C Tensor object, manage memory of DLTensor.
  *

aten/src/ATen/functorch/ADInterpreters.cpp

@@ -139,7 +139,7 @@ static void autogradBasedTransformSendToNext(
   std::bitset<default_bitset_size> outputs_aliasing_immutable; // set = 1 for all bits
   if(!grad_special_case) {
     for (auto idx = stack->size() - args_size; idx < stack->size(); idx++) {
-      const auto& ivalue = (*stack)[idx];
+      const auto ivalue = (*stack)[idx];
       if (!ivalue.isTensor()) {
         continue; // only input that can be aliased is a tensor, not a tensor list (expect in ops without returns)
       }

aten/src/ATen/functorch/BatchRulesLinearAlgebra.cpp

@@ -6,8 +6,6 @@
 
 #include <ATen/functorch/BatchRulesHelper.h>
 
-#include <algorithm>
-
 namespace at::functorch {
 
 typedef std::tuple<Tensor, std::optional<int64_t>> oneOutput;
@@ -317,7 +315,7 @@ oneOutput linalg_lu_solve_batch_rule(
   const auto LU_num_batch_dims = rankWithoutBatchDim(LU_, LU_bdim) - LU_min_rank;
   const auto pivots_num_batch_dims = rankWithoutBatchDim(pivots_, pivots_bdim) - pivots_min_rank;
   const auto B_num_batch_dims = rankWithoutBatchDim(B_, B_bdim) - B_min_rank;
-  const auto max_num_batch_dims = std::max({LU_num_batch_dims, pivots_num_batch_dims, B_num_batch_dims});
+  const auto max_num_batch_dims = std::max(std::max(LU_num_batch_dims, pivots_num_batch_dims), B_num_batch_dims);
 
   LU_ = maybePadToLogicalRank(LU_, LU_bdim, max_num_batch_dims + LU_min_rank);
   pivots_ = maybePadToLogicalRank(pivots_, pivots_bdim, max_num_batch_dims + pivots_min_rank);

aten/src/ATen/functorch/BatchRulesUnaryOps.cpp

@@ -171,8 +171,6 @@ TORCH_LIBRARY_IMPL(aten, FuncTorchBatched, m) {
 
   POINTWISE_BOXED(fill_.Scalar);
   POINTWISE_BOXED(zero_);
-  // This is special because this op doesn't return anything
-  m.impl("_assert_tensor_metadata", native::_assert_tensor_metadata);
 
 #undef UNARY_POINTWISE
 #undef UNARY_POINTWISE_ALL

aten/src/ATen/native/BinaryOps.cpp

@@ -897,11 +897,11 @@ Tensor& div_(Tensor& self, const Scalar& other) {
 }
 
 Tensor div(const Tensor& self, const Scalar& other, std::optional<std::string_view> rounding_mode) {
-  return self.div(wrapped_scalar_tensor(other), rounding_mode); // redispatch!
+  return self.div(wrapped_scalar_tensor(other), std::move(rounding_mode)); // redispatch!
 }
 
 Tensor& div_(Tensor& self, const Scalar& other, std::optional<std::string_view> rounding_mode) {
-  return self.div_(wrapped_scalar_tensor(other), rounding_mode); // redispatch!
+  return self.div_(wrapped_scalar_tensor(other), std::move(rounding_mode)); // redispatch!
 }
 
 // divide, alias for div
@@ -926,23 +926,23 @@ Tensor& divide_(Tensor& self, const Scalar& other) {
 }
 
 Tensor& divide_out(const Tensor& self, const Tensor& other, std::optional<std::string_view> rounding_mode, Tensor& result) {
-  return at::div_out(result, self, other, rounding_mode);
+  return at::div_out(result, self, other, std::move(rounding_mode));
 }
 
 Tensor divide(const Tensor& self, const Tensor& other, std::optional<std::string_view> rounding_mode) {
-  return self.div(other, rounding_mode);
+  return self.div(other, std::move(rounding_mode));
 }
 
 Tensor& divide_(Tensor& self, const Tensor& other, std::optional<std::string_view> rounding_mode) {
-  return self.div_(other, rounding_mode);
+  return self.div_(other, std::move(rounding_mode));
 }
 
 Tensor divide(const Tensor& self, const Scalar& other, std::optional<std::string_view> rounding_mode) {
-  return self.div(other, rounding_mode);
+  return self.div(other, std::move(rounding_mode));
 }
 
 Tensor& divide_(Tensor& self, const Scalar& other, std::optional<std::string_view> rounding_mode) {
-  return self.div_(other, rounding_mode);
+  return self.div_(other, std::move(rounding_mode));
 }
 
 // true_divide, an alias for div

aten/src/ATen/native/CPUBlas.cpp

@@ -457,9 +457,24 @@ void gemm(
     return;
   }
 #endif
+  // for the fallback path, first compute gemm with beta = 0,
+  // and then add c in full precision.
+  int64_t c_size = n * m;
+  std::vector<float> float_c(c_size, 0.f);
   gemm_no_downcast_stub(
       at::kCPU, at::kBFloat16,
-      transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+      transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
+  for (const auto j : c10::irange(n)) {
+    for (const auto i : c10::irange(m)) {
+      auto offset = j * ldc + i;
+      // beta == 0 won't propagate NaN from C
+      if (beta == 0.f) {
+        c[offset] = float_c[j * m + i];
+      } else {
+        c[offset] = beta * c[offset] + float_c[j * m + i];
+      }
+    }
+  }
 }
 
 void gemm(
@@ -478,9 +493,24 @@ void gemm(
     return;
   }
 #endif
+  // for the fallback path, first compute gemm with beta = 0,
+  // and then add c in full precision.
+  int64_t c_size = n * m;
+  std::vector<float> float_c(c_size, 0.f);
   gemm_no_downcast_stub(
       at::kCPU, at::kHalf,
-      transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+      transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
+  for (const auto j : c10::irange(n)) {
+    for (const auto i : c10::irange(m)) {
+      auto offset = j * ldc + i;
+      // beta == 0 won't propagate NaN from C
+      if (beta == 0.f) {
+        c[offset] = float_c[j * m + i];
+      } else {
+        c[offset] = beta * c[offset] + float_c[j * m + i];
+      }
+    }
+  }
 }
 
 void gemm(