Arch Linux User Repository

what link should i use to try it, i see there the latest jaxlib 0.4.32:

https://github.com/jax-ml/jax/tags

@actionless does it work for the latest version (0.4.35)?

yup, i've posted a patch below

but it probably need some additional bazel command for downloading the files without build - otherwise it would be nothing to patch yet (so build would succeed only after restart after 1st fail)

@medaminezghal I will try one more time but last time it seems impossible without patching bazel's BUILD files (see discussion)

@daskol could you update this package to the latest version (0.4.35)?

@truncs Yes, there is a building issues with clang18 and local cuda.

UPD It turns out that there are multiple issues related to building jaxlib with CUDA support (see jax-ml/jax#23689).

1. For some reason local CUPTI does not materialize in targte build directory (xla/tsl).
2. An old issue between libc++ and libstdc++ when inaccurate ordering of system search paths (-isystem option) leads to broken #include_next inclusion.

Seems like jaxlib and jax versions are not in sync.

  3  aur/python-jax                       0.4.31-1        -> 0.4.33-1
  2  aur/python-orbax-checkpoint          0.5.20-1        -> 0.6.4-1
  1  aur/yay                              12.4.1-1        -> 12.4.2-1
==> Packages to exclude: (eg: "1 2 3", "1-3", "^4" or repo name)
 -> Excluding packages may cause partial upgrades and break systems
==> 
 -> could not find all required packages: python-jaxlib >=0.4.33

i've finally solved previously mentioned problem by adding this before the last line of build() func:

    sed -i \
        -e 's|/opt/cuda/targets/x86_64-linux/include/cudnn_adv.h|/usr/include/cudnn_adv.h|g' \
        -e 's|/opt/cuda/targets/x86_64-linux/include/cudnn_backend.h|/usr/include/cudnn_backend.h|g' \
        -e 's|/opt/cuda/targets/x86_64-linux/include/cudnn_cnn.h|/usr/include/cudnn_cnn.h|g' \
        -e 's|/opt/cuda/targets/x86_64-linux/include/cudnn_graph.h|/usr/include/cudnn_graph.h|g' \
        -e 's|/opt/cuda/targets/x86_64-linux/include/cudnn_ops.h|/usr/include/cudnn_ops.h|g' \
        -e 's|/opt/cuda/targets/x86_64-linux/include/cudnn_version.h|/usr/include/cudnn_version.h|g' \
        bazel-jax-jaxlib-v0.4.31/external/local_config_cuda/cuda/BUILD

i'm getting this error during the build:

ERROR: /home/lie/.cache/pikaur/build/python-jaxlib-cuda/src/bazel/905847f791db658de2b823a610f64a15/external/local_config_cuda/cuda/BUIL
D:4241:8: Executing genrule @local_config_cuda//cuda:cudnn-include failed: (Exit 1): bash failed: error executing command (from target
@local_config_cuda//cuda:cudnn-include)
  (cd /home/lie/.cache/pikaur/build/python-jaxlib-cuda/src/bazel/905847f791db658de2b823a610f64a15/execroot/__main__ && \
  exec env - \
    CUDA_TOOLKIT_PATH=/opt/cuda \
    GCC_HOST_COMPILER_PATH=/usr/bin/gcc-13 \
    JAXLIB_RELEASE=0.4.30 \
    NCCL_INSTALL_PATH=/usr \
    PATH=/home/lie/bin:/usr/sbin:/sbin:/usr/condabin:/usr/local/sbin:/usr/local/bin:/usr/bin:/opt/cuda/bin:/opt/cuda/nsight_compute:/op
t/cuda/nsight_systems/bin:/opt/intel/mediasdk/bin:/usr/lib/jvm/default/bin:/usr/bin/site_perl:/usr/bin/vendor_perl:/usr/bin/core_perl \
    TF_CUDA_COMPUTE_CAPABILITIES=sm_50,sm_60,sm_70,sm_80,compute_90 \
    TF_CUDA_PATHS=/opt/cuda,/usr/lib,/usr \
  /bin/bash -c 'source external/bazel_tools/tools/genrule/genrule-setup.sh; cp -f "/opt/cuda/targets/x86_64-linux/include/cudnn.h" "baz
el-out/k8-opt/bin/external/local_config_cuda/cuda/cudnn/include/cudnn.h" && cp -f "/opt/cuda/targets/x86_64-linux/include/cudnn_adv.h"
"bazel-out/k8-opt/bin/external/local_config_cuda/cuda/cudnn/include/cudnn_adv.h" && cp -f "/opt/cuda/targets/x86_64-linux/include/cudnn
_backend.h" "bazel-out/k8-opt/bin/external/local_config_cuda/cuda/cudnn/include/cudnn_backend.h" && cp -f "/opt/cuda/targets/x86_64-lin
ux/include/cudnn_cnn.h" "bazel-out/k8-opt/bin/external/local_config_cuda/cuda/cudnn/include/cudnn_cnn.h" && cp -f "/opt/cuda/targets/x8
6_64-linux/include/cudnn_graph.h" "bazel-out/k8-opt/bin/external/local_config_cuda/cuda/cudnn/include/cudnn_graph.h" && cp -f "/opt/cud
a/targets/x86_64-linux/include/cudnn_ops.h" "bazel-out/k8-opt/bin/external/local_config_cuda/cuda/cudnn/include/cudnn_ops.h" && cp -f "
/opt/cuda/targets/x86_64-linux/include/cudnn_version.h" "bazel-out/k8-opt/bin/external/local_config_cuda/cuda/cudnn/include/cudnn_versi
on.h" ')
# Configuration: 1f82cdfaad48bddc42a9a9db6afc8d7da3bd013e73745c6926c6e665ef1885b7
# Execution platform: @local_execution_config_platform//:platform
cp: cannot stat '/opt/cuda/targets/x86_64-linux/include/cudnn_adv.h': No such file or directory

for some reason it's looking for that lib inside cuda's install dir (/opt/cuda) instead of the correct path:

$ locate cudnn_adv.h
/usr/include/cudnn_adv.h

$ pacman -Qo /usr/include/cudnn_adv.h
/usr/include/cudnn_adv.h is owned by cudnn 9.2.1.18-1

Does this really need to download an outdated bazel blob and install a doubly outdated gcc (that has since been dropped from the Arch repositories and needs to be built from source)?