Vasp 6.5.1 test issue

[liang-wen_pi]

Hello,everyone!
These are my basic parameters.
1. Rocky / HPC-sdk 25.11 / CUDA 13.0 / intel-mkl 25.2 hdf5 1.14.6 / gcc 8.5.0

2.I choose makefile.include.nvhpc_ompi_mkl_omp_acc to make vasp. At first, i choose wannier90 3.1.0 libxc 6.2.2 and libbeef-master and DFT-D4 3.6.0,t but during the compilation of DFT-D4, an error occurred：

Building Fortran object _deps/multicharge-build/CMakeFiles/multicharge-lib.dir/src/multicharge/model/eeq.f90.o

nvfortran-Info-Switch -Mvect -Mvect forces -O3

NVFORTRAN-S-0074-Illegal number or type of arguments to move_alloc - keyword argument to (/opt/Software/dftd4-v3.6.0/dftd4-main/build/_deps/multicharge-src/src/multicharge/model/eeq.f90: 589)

So i gave up.

3.When compiling vasp6.5,1, another error occurred stating that it could not be compiled together with libxc and libbeef.

NVFORTRAN-S-1061-Procedures called in a compute region must have acc routine information - beefx (xc_driver.f90: 1565

So i gave up libxc and libbeef

4.After a successful installation, I did a small test. Although the program ran through its process, it gave a warning:
Warning: ieee_invalid is signaling
Warning: ieee_divide_by_zero is signaling
Warning: ieee_underflow is signaling
Warning: ieee_inexact is signaling
FORTRAN STOP
And i use the same parameters to run this test on GPU(RTX5090) vasp 6.5.1 and CPU vasp 6.4.2, the output and testing time are very different.

In the attachment, there are my "make.include" file and two different versions of the output data.
I want to know if there exist some mistakes when making Vasp, wish a reply for these questions. THANKS!

[alexey.tal]

Dear liang-wen_pi,

1. We are not involved in the development of DFT-D4 and therefore cannot provide any support for it.

2. The issues related to building libxc and libbeef with the GPU support in VASP were discussed in these topics:
   https://www.vasp.at/forum/viewtopic.php?t=18567
   https://www.vasp.at/forum/viewtopic.php?t=18940

VASP uses OpenACC for offloading, but neither libbeef nor libxc support OpenACC, which unfortunately makes them incompatible for offloading the calculations in VASP.

1. These warnings have been discussed in this topic:
   https://www.vasp.at/forum/viewtopic.php?t=18160

Regarding the difference between your CPU and GPU runs: from a performance point of view, these calculations are very different. You use GPUs in one of the runs but not in the other. Furthermore, the number of CPU cores, nodes, and OpenMP threads differs between the two cases, and even the VASP versions are not the same. Under these circumstances, I would not expect the wall times of these calculations to be very similar.

The final total energies of the two runs differ by 0.00057 eV. Without knowing the convergence criteria you used, it is difficult to draw a definitive conclusion, but this difference does not suggest any underlying problem. To verify that your VASP installation was compiled correctly, I recommend running the VASP testsuite. For more details, see the documentation here: testsuite.

Best regards,
Alexey

[liang-wen_pi]

THANKS!
I did the fast test, and i got a prompt.

==================================================================
SUCCESS: ALL SELECTED TESTS PASSED
==================================================================