[RESOLVED] Issue with hcp_diffusion and --nv


I’m trying to run hcp_diffusion with the latest 0.99.2d singularity container (updated a few days ago), and I met some issues when running with --nv parameter for gpu.

Specifically, I met the following error at the step of freesurfer tkregister2 after eddy is finished.

/opt/freesurfer/freesurfer/tktools/tkregister2.bin: /lib64/libc.so.6: version `GLIBC_ABI_DT_RELR' not found (required by /.singularity.d/libs/libGL.so.1)
/opt/freesurfer/freesurfer/tktools/tkregister2.bin: /lib64/libc.so.6: version `GLIBC_ABI_DT_RELR' not found (required by /.singularity.d/libs/libGLdispatch.so.0)
/opt/freesurfer/freesurfer/tktools/tkregister2.bin: /lib64/libc.so.6: version `GLIBC_2.34' not found (required by /.singularity.d/libs/libGLdispatch.so.0)
/opt/freesurfer/freesurfer/tktools/tkregister2.bin: /lib64/libc.so.6: version `GLIBC_ABI_DT_RELR' not found (required by /.singularity.d/libs/libGLX.so.0)
/opt/freesurfer/freesurfer/tktools/tkregister2.bin: /lib64/libc.so.6: version `GLIBC_2.34' not found (required by /.singularity.d/libs/libGLX.so.0)

I was able to reproduce this by directly calling tkregister2 inside:

singularity exec --cleanenv --nv ${QUNEX_CONTAINER} bash

However, when I run the command without the --nv flag and call tkregister2, the error does not appear.

Could you help with confirming this please? Thanks!

Zhen-Qi Liu


Yes, we are aware of this issue. I assume you are using CUDA 12? The errors you are seeing usually pop up if you have CUDA 12 or newer. The problem is that your system CUDA is too new for the system inside the container. We are actively working on updating the container so this will work. We plan to release 0.100.0 over the next couple of weeks and it should support CUDA 12 as well.

In the meantime, you could try running this with an older CUDA version. At this moment 10.2 is the version that is the most tested for QuNex and tools under the hood (FSL, eddy). We also processed a bunch of DWI things with 11.3.

Best, Jure

Thanks so much! Since I’m on HPC, I was able to switch to CUDA 10.2 and it works nicely.