QSIPrep 26.0.0 Singularity image: eddy_cuda10.2 missing and DSI Studio fails with GLIBC_2.38 under --nv

[bkossows]

I am trying to run pennlinc/qsiprep:26.0.0 converted to a Singularity image with GPU support:

singularity run --nv qsiprep-26.0.0.simg ...

Two issues appear.

First, QSIPrep selects CUDA eddy but looks specifically for eddy_cuda10.2:

Using CUDA and 1 threads in eddy
Cannot run QSIPrep. Missing dependencies:
* eddy_cuda10.2 (Interface: ExtendedEddy)

However, the image contains:

/app/.pixi/envs/qsiprep/bin/eddy_cpu
/app/.pixi/envs/qsiprep/bin/eddy_cuda11.0

A workaround is to bind a wrapper/symlink named eddy_cuda10.2 pointing to eddy_cuda11.0.

Second, after fixing the eddy name issue, DSI Studio fails when running with --nv:

dsi_studio: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.38' not found
(required by /.singularity.d/libs/libGLdispatch.so.0)

dsi_studio: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.38' not found
(required by /.singularity.d/libs/libGLX.so.0)

This happens for commands such as:

dsi_studio --action=src ...

The container uses /opt/dsi-studio/dsi_studio, so this does not seem to be caused by a host-side dsi_studio shadowing the container binary.

Environment:

QSIPrep image: pennlinc/qsiprep:26.0.0 converted to Singularity .simg
Runtime: Singularity with --nv
Command includes: --output-resolution 1.5 --nthreads 16 --omp-nthreads 8 --skip-bids-validation

[tsalo]

Can you try using the unstable Docker image? I believe I fixed this bug recently, but we haven't made a new release since that PR was merged.

[samuelholly]

Hello, I ran into the same problem with eddy_cuda today. I can confirm that the bug is resolved in the unstable version, eddy runs fine.

I tried it with Docker on Windows 11.

Hope this helps.

Best regards

Samuel

[tsalo]

Sorry, I'm a little confused by your first two sentences. Does unstable work, or are you experiencing a different issue with eddy_cuda?

[samuelholly]

Sorry for the confusion. What I meant is this: I encountered the same issue as the original poster when running the latest version (26.0.0):

Cannot run QSIPrep. Missing dependencies:
* eddy_cuda10.2 (Interface: ExtendedEddy)

Then as per your suggestion I tried to run the latest unstable version and this worked fine.

[bkossows]

CUDA works fine in the "unstable" version, but some a several times dsi_studio has failed to execute throwing errors like this below:

Cmdline:
        dsi_studio --action=src  --bval=/home/jovyan/shared_storage/QSIPREP/bids-dir/sub-KJ023/ses-01/dwi/sub-KJ023_ses-01_dir-AP_dwi.bval --bvec=/home/jovyan/shared_storage/QSIPREP/bids-dir/sub-KJ023/ses-01/dwi/sub-KJ023_ses-01_dir-AP_dwi.bvec --source=/home/jovyan/qsiprep-work/sub-KJ023/qsiprep_26_1_wf/sub_KJ023_ses_01_wf/dwi_preproc_ses_01_dir_AP_wf/pre_hmc_wf/merge_and_denoise_wf/raw_merge/sub-KJ023_ses-01_dir-AP_dwi_merged.nii.gz --thread_count=8 --output=/home/jovyan/qsiprep-work/sub-KJ023/qsiprep_26_1_wf/sub_KJ023_ses_01_wf/dwi_preproc_ses_01_dir_AP_wf/pre_hmc_wf/merge_and_denoise_wf/dwi_qc_wf/raw_src/sub-KJ023_ses-01_dir-AP_dwi_merged.src.gz
Stdout:

Stderr:
        dsi_studio: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.38' not found (required by /.singularity.d/libs/libGLdispatch.so.0)
        dsi_studio: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.38' not found (required by /.singularity.d/libs/libGLX.so.0)
Traceback:
        RuntimeError: subprocess exited with code 1.

As suggested by AI copilot I commented 4 libraries in /etc/apptainer/nvliblist.conf and the workflow completed successfully. Is that the right way to use qsiprep container in singularity right now?

sudo sed -i.bak \
  -e 's/^\(libGLdispatch\.so\)/#\1/' \
  -e 's/^\(libGLX_installertest\.so\)/#\1/' \
  -e 's/^\(libGLX_nvidia\.so\)/#\1/' \
  -e 's/^\(libGLX\.so\)/#\1/' \
  /etc/apptainer/nvliblist.conf