Cylinder pull geometry crashes the simulation

GROMACS version: 2021.5-EasyBuild-4.5.3
GROMACS modification: Not sure

I’m trying to pull a molecule through a bilayer.
Everything works fine if I use:

pull = yes
pull-print-ref-value = yes
pull-print-components = yes
pull_ncoords = 1
pull_ngroups = 2
pull_group1_name = My_molecule
pull_group2_name = Membrane
pull_coord1_type = umbrella
pull_coord1_geometry = direction
pull-coord1-vec = 0 0 -1
pull_coord1_dim = N N Y
pull_coord1_groups = 1 2
pull_coord1_start = no
pull-coord1-init = 1.92 ; reference distance at t=0.
pull_coord1_rate = 0
pull_coord1_k = 3000 ; kJ mol^-1 nm^-2
pull-pbc-ref-prev-step-com = yes
pull_group2_pbcatom = 32078

But if I change direction to cylinder I get the following error right at the start of the simulation:

GROMACS:      gmx mdrun, version 2021.5-EasyBuild-4.5.3
Executable:   /cluster/software/GROMACS/2021.5-foss-2021b/bin/gmx_mpi
Data prefix:  /cluster/software/GROMACS/2021.5-foss-2021b
Working dir:  /cluster/work/jobs/395132
Command line:
  gmx_mpi mdrun -deffnm umbrella_dmpc_1 -px umbrella_dmpc_1_pullx.xvg -pf umbrella_dmpc_1_pullf.xvg -cpi -pin on -v -maxh 10

Reading file umbrella_dmpc_1.tpr, VERSION 2021.5-EasyBuild-4.5.3 (single precision)
Changing nstlist from 20 to 100, rlist from 1.21 to 1.324

Using 640 MPI processes
Using 2 OpenMP threads per MPI process

Overriding thread affinity set outside gmx mdrun
starting mdrun 'Title in water'
50000000 steps, 100000.0 ps.
step 0

Program:     gmx mdrun, version 2021.5-EasyBuild-4.5.3
Source file: src/gromacs/ewald/pme_redistribute.cpp (line 305)
MPI rank:    29 (out of 640)

Fatal error:
291 particles communicated to PME rank 5 are more than 2/3 times the cut-off
out of the domain decomposition cell of their charge group in dimension y.
This usually means that your system is not well equilibrated.

For more information and tips for troubleshooting, please check the GROMACS
website at

Which is not really helpful since I’m using the same starting configuration in both cases (which is the last frame in a ~1 microsecond standard md simulation).
Does anyone have any idea on what’s going wrong and how to possibly fix it?