• Bug#1105066: pytorch-cuda: FTBFS: ld.lld: error: undefined reference: s

    From Santiago Vila@21:1/5 to All on Sat May 10 20:10:01 2025
    Package: src:pytorch-cuda
    Version: 2.6.0+dfsg-7
    Severity: serious
    Tags: ftbfs trixie sid

    Dear maintainer:

    During a rebuild of all packages in unstable, your package failed to build:

    --------------------------------------------------------------------------------
    [...]
    [2465/2591] : && /usr/bin/cuda-g++ -Wdate-time -D_FORTIFY_SO [too-long-redacted] urce=bin/c10_tempfile_test && :
    [2466/2591] : && /usr/bin/cuda-g++ -Wdate-time -D_FORTIFY_SO [too-long-redacted] source=bin/c10_typeid_test && :
    [2467/2591] : && /usr/bin/cuda-g++ -Wdate-time -D_FORTIFY_SO [too-long-redacted] 10_intrusive_ptr_benchmark && :
    FAILED: bin/c10_intrusive_ptr_benchmark
    : && /usr/bin/cuda-g++ -Wdate-time -D_FORTIFY_SOURCE=2 -g -O [too-long-redacted] 10_intrusive_ptr_benchmark && :
    ld.lld: error: undefined reference: std::ios_base_library_init()@GLIBCXX_3.4.32 >>> referenced by /usr/lib/x86_64-linux-gnu/libbenchmark.so (disallowed by --no-allow-shlib-undefined)
    collect2: error: ld returned 1 exit status
    [2468/2591] /usr/bin/cuda-g++ -DDNNL_EXPERIMENTAL_UKERNEL -D [too-long-redacted] test/cpp/lazy/test_lazy_ops.cpp
    ninja: build stopped: subcommand failed.
    Building wheel torch-2.6.0+debian
    -- Building version 2.6.0+debian
    cmake -GNinja -DBLAS=Generic -DBUILD_PYTHON=True -DBUILD_SPLIT_CUDA=ON -DBUILD_TEST=True -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_INSTALL_PREFIX=/<<PKGBUILDDIR>>/torch -DCMAKE_PREFIX_PATH=/usr/local/lib/python3.13/dist-packages -DONNX_NAMESPACE=onnx -
    DPython_EXECUTABLE=/usr/bin/python3.13 -DTORCH_BUILD_VERSION=2.6.0+debian -DTORCH_CUDA_ARCH_LIST=5.0;6.0;7.0;7.5;8.0;8.6;9.0+PTX -DUSE_CUDA=ON -DUSE_CUDNN=ON -DUSE_DISTRIBUTED=ON -DUSE_FBGEMM=OFF -DUSE_GFLAGS=OFF -DUSE_GLOG=OFF -DUSE_GLOO=ON -DUSE_ITT=
    OFF -DUSE_KINETO=ON -DUSE_LLVM=/usr/lib/llvm-19 -DUSE_MAGMA=OFF -DUSE_MIOPEN=OFF -DUSE_MKLDNN=ON -DUSE_MPI=OFF -DUSE_NNPACK=OFF -DUSE_NUMPY=True -DUSE_PYTORCH_QNNPACK=ON -DUSE_ROCM=OFF -DUSE_SYSTEM_LIBS=ON -DUSE_SYSTEM_NCCL=ON -DUSE_TENSORPIPE=ON -DUSE_
    VULKAN=1 -DUSE_XNNPACK=ON /<<PKGBUILDDIR>>
    cmake --build . --target install --config RelWithDebInfo -- -j 2 -v
    make[1]: *** [debian/rules:199: override_dh_auto_install] Error 1
    make[1]: Leaving directory '/<<PKGBUILDDIR>>'
    make: *** [debian/rules:162: binary] Error 2
    dpkg-buildpackage: error: debian/rules binary subprocess returned exit status 2 --------------------------------------------------------------------------------

    The above is just how the build ends and not necessarily the most relevant part.
    If required, the full build log is available here:

    https://people.debian.org/~sanvila/build-logs/202505/

    About the archive rebuild: The build was made on virtual machines from AWS, using sbuild and a reduced chroot with only build-essential packages.

    If you could not reproduce the bug please contact me privately, as I
    am willing to provide ssh access to a virtual machine where the bug is
    fully reproducible.

    If this is really a bug in one of the build-depends, please use
    reassign and add an affects on src:pytorch-cuda, so that this is still
    visible in the BTS web page for this package.

    Thanks.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)