Sage fails during GPU setup stage

53 views
Skip to first unread message

Nathan Brown

unread,
Oct 11, 2025, 10:14:44 PMOct 11
to sage-devel
This occurs on a HP z840 with RHEL 10 installed. The GPU is an Nvidia RTX 5060 Ti that is recognized by the operating system.

One small thing I noticed that is almost definitely irrelevant, but might hint at a deeper issue is that the config log shows that gfortran is noticed but that the suitesparce log says there is no fortran compiler installed. Perhaps some recent change has broken some of the defaults and lead to this error on fairly new hardware.

Command:
/path/to/sage/sage -i notebook

Terminal printout:
[suitesparse-7.8.0] [spkg-install] cd /path/to/sage/local/var/tmp/sage/build/suitesparse-7.8.0/src/CHOLMOD && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DBLAS_OpenBLAS -DCHOLMOD_EXPORTS --options-file CMakeFiles/CHOLMOD.dir/includes_CUDA.rsp -O3 -DNDEBUG -std=c++11 "--generate-code=arch=compute_52,code=[compute_52,sm_52]" "--generate-code=arch=compute_75,code=[compute_75,sm_75]" "--generate-code=arch=compute_80,code=[compute_80,sm_80]" -Xcompiler=-fPIC -MD -MT CHOLMOD/CMakeFiles/CHOLMOD.dir/GPU/cholmod_gpu_kernels.cu.o -MF CMakeFiles/CHOLMOD.dir/GPU/cholmod_gpu_kernels.cu.o.d -x cu -rdc=true -c /path/to/sage/local/var/tmp/sage/build/suitesparse-7.8.0/src/CHOLMOD/GPU/cholmod_gpu_kernels.cu -o CMakeFiles/CHOLMOD.dir/GPU/cholmod_gpu_kernels.cu.o
[suitesparse-7.8.0] [spkg-install] nvcc fatal   : Unsupported gpu architecture 'compute_52'
[suitesparse-7.8.0] [spkg-install] make[6]: *** [CHOLMOD/CMakeFiles/CHOLMOD.dir/build.make:3048: CHOLMOD/CMakeFiles/CHOLMOD.dir/GPU/cholmod_gpu_kernels.cu.o] Error 1
[suitesparse-7.8.0] [spkg-install] make[5]: *** [CMakeFiles/Makefile2:423: CHOLMOD/CMakeFiles/CHOLMOD.dir/all] Error 2
[suitesparse-7.8.0] [spkg-install] make[4]: *** [Makefile:149: all] Error 2
[suitesparse-7.8.0] [spkg-install] Error installing suitesparse-7.8.0
[suitesparse-7.8.0] Error installing package suitesparse-7.8.0

suitesparse-7.8.0.log
config.log

Dima Pasechnik

unread,
Oct 12, 2025, 2:28:15 PMOct 12
to sage-...@googlegroups.com
It seems to me that one ought to use CUDA-aware fortran, nvfortran (https://developer.nvidia.com/cuda-fortran) for the whole thing.
That is, uninstall gfortran you for some reason (I suppose Rehl 10 does has a good enough gfortran package provided, which you can install)
built in Sage (make fortran-clean) and run 

FC=nvfortran ./configure 

to configure nvfortran to be used. Not sure if this will work, nobody tried as far as I know.

Failing this, your best bet might be to use a prebuilt suitesparse.
I suppose you will also need a CUDA-enabled blas/lapack.
(I suppose it comes with nvfortran)

HTH
Dima







--
You received this message because you are subscribed to the Google Groups "sage-devel" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sage-devel+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/sage-devel/286d1547-9c2b-4c0a-aafc-5e32a57657ean%40googlegroups.com.
Reply all
Reply to author
Forward
0 new messages