New Cuda-10.1 run installer completely broken

This is on RHEL 7.6.

$ sh cuda_10.1.105_418.39_linux.run --silent --toolkit --toolkitpath=$HOME/tkit
/usr/lib64/libcublasLt.so.10.1.0.105 can't be opened
/usr/lib64/libcublas.so.10.1.0.105 can't be opened
/usr/lib64/libnvblas.so.10.1.0.105 can't be opened

The installation is grossly incomplete!

$ ls tkit/include/
cusolver_common.h  cusolverSp_LOWLEVEL_PREVIEW.h  nvToolsExtMeta.h
cusolverDn.h       nvToolsExtCuda.h               nvToolsExtSync.h
cusolverRf.h       nvToolsExtCudaRt.h             nvtx3
cusolverSp.h       nvToolsExt.h

$ ls tkit/lib64/
libaccinj64.so           liblapack_static.a      libnppisu.so.10
libaccinj64.so.10.1      libmetis_static.a       libnppisu.so.10.1.105
libaccinj64.so.10.1.105  libnppc.so.10           libnppitc.so.10
libcudart.so.10.1        libnppc.so.10.1.105     libnppitc.so.10.1.105
libcudart.so.10.1.105    libnppial.so.10         libnpps.so.10
libcufft.so.10           libnppial.so.10.1.105   libnpps.so.10.1.105
libcufft.so.10.1.105     libnppicc.so.10         libnvgraph.so.10
libcufftw.so.10          libnppicc.so.10.1.105   libnvgraph.so.10.1.105
libcufftw.so.10.1.105    libnppicom.so.10        libnvjpeg.so.10
libcuinj64.so            libnppicom.so.10.1.105  libnvjpeg.so.10.1.105
libcuinj64.so.10.1       libnppidei.so.10        libnvrtc-builtins.so
libcuinj64.so.10.1.105   libnppidei.so.10.1.105  libnvrtc-builtins.so.10.1
libcurand.so.10          libnppif.so.10          libnvrtc-builtins.so.10.1.105
libcurand.so.10.1.105    libnppif.so.10.1.105    libnvrtc.so.10.1
libcusolver.so           libnppig.so.10          libnvrtc.so.10.1.105
libcusolver.so.10        libnppig.so.10.1.105    libnvToolsExt.so
libcusolver.so.10.1.105  libnppim.so.10          libnvToolsExt.so.1
libcusolver_static.a     libnppim.so.10.1.105    libnvToolsExt.so.1.0.0
libcusparse.so.10        libnppist.so.10         stubs
libcusparse.so.10.1.105  libnppist.so.10.1.105

cuda-installer.log (1.53 MB)

Who decided that cublas must be unconditionally installed to /usr/lib64 and /usr/include and that the installer has to quit midway with an incomplete installation if /usr isn’t writeable? Are you serious?

This affect also to creation of the .pc files, the .desktop files and the manuals.

i’ve discover throught strings command and some grep, the hiden option ‘–defaultroot’, but still unable to install these files

I’ve opened a bug into nvidia bugtracker (since cuda 8 is released. same bug, but with the older perl scripts installation). But seems is not a devs priority…

Greetings and Good Luck

Why not use the rpms? They have rhel 7.x specific ones and you know that rpm installation isn’t going to leave you with a mess you can’t easily clean up.

Langdalepl, Chemal is doing a user install which is a task for the .run installer and the rpms would be of no help for that, obviously.

I was told in another forum over here, that there is an undocumented installer option ‘–defaultroot’.

So to get back the old layout, one types:

sh cuda_10.1.105_418.39_linux.run --silent --toolkit --toolkitpath=$HOME/tkit --defaultroot=$HOME/tkit --samples --samplespath=$HOME/tkit/samples

.pc files are still missing, but everything vital is there.