link_cuda_dso: warning: dlopen() isn't supported on this platform: failed to load library
vlawhern opened this issue · comments
Hello,
I'm trying to get some models working with GPU support on RHEL 8 and am running into the above error. This error doesn't show up on my other system (Ubuntu 22.04) and on that system I'm able to use the GPU just fine. I suspect this issue might be caused by/related to SELinux (I'm not a sys-admin so I'm not able to change anything related to SELinux).
Just wondering if there's any advice on how to proceed here.
Thanks.
After searching around a bit I'm guessing this issue is due to static linking not being supported in RHEL 8 (see https://access.redhat.com/articles/rhel8-abi-compatibility)... the relevant text is below:
Static Linking with the C/C++ Runtime
Static linking with the C/C++ runtime is not supported. This includes linking with any files that are part of the glibc-static or libstdc++-static packages. You may choose to link statically, but the resulting application binaries may fail to operate if any package in the installation is changed.
If my understanding is correct, how would you advise getting llamafile to work on these kinds of systems?
llamafile is a static binary. If your OS enforces a policy that prevents foreign programs from linking GPU libraries, then there's not a whole lot you can do besides switching to a more permissive operating system. Contributions are welcome if you know how to help us help users to use the hardware resources they own. But if your OS says it's not supported then we can't help you.