nvcc fatal : Unsupported gpu architecture 'compute_NATIVE'

[newjavaer]

When I run ./build.sh with branch 22.12 ,the erro print log is nvcc fatal : Unsupported gpu architecture 'compute_NATIVE'
How can I deal the erro ,thank you.

[bdice]

@newjavaer What version of CMake are you using? Try updating to 3.24.0 or 3.25.1 (there is a bug in 3.25.0). See info about the native CUDA architecture option here: https://cmake.org/cmake/help/latest/prop_tgt/CUDA_ARCHITECTURES.html

If you can also discuss why you chose to build from source instead of using a conda package or RAPIDS Docker container, that would be helpful to know!

[newjavaer]

Thank you for your reply.I confirm my CMake version is 25.1 again.But there still have this bug .I'm trying to figure out how nvcomp is used in the cudf library.It's important for me .It would be even better if you help me figure out which interfaces or functions of NVCOMP are used in CUDF。Thanks

[bdice]

If you could provide more information about your system, the commit of cudf you're trying to build, any options you're passing to the build script, and the full output/error traceback, that would be helpful for diagnosing the problem.

@vuule Would you be able to help with the questions about nvcomp?

[vuule]

All nvCOMP calls are made from https://github.com/rapidsai/cudf/blob/branch-22.12/cpp/src/io/comp/nvcomp_adapter.cpp
Please let me know if you have any further questions about the nvCOMP use.

[GregoryKimball]

Thank you @newjavaer for sharing your experience with libcudf and nvCOMP. Please let us know if you would like to discuss further.