The sympy --analytic option to NMODL produces code that crashes on GPU.
The same models pass on GPU if this option is not passed.
This option is enabled by the +sympy variant of the CoreNEURON Spack recipe.

Reproducing

One way is to use the channel-benchmark tests as added in neuronsimulator/nrn#1439.

-DNRN_ENABLE_CORENEURON=ON \
-DCORENRN_ENABLE_GPU=ON \
-DCORENRN_ENABLE_NMODL=ON \
-DCORENRN_NMODL_FLAGS=" sympy --analytic"

and build (with NVHPC compilers etc.). Two (three, counting the results comparison) tests should fail:
channel_benchmark_Hippo::coreneuron_gpu and channel_benchmark_Hippo::coreneuron_gpu_offline.