UCL Error: Error compiling PTX Program

Hi, everyone.
I installed LAMMPS with a GPU accelerator on my machine. I used Opencl API and the installation was successful. But when I run my simulation an error occurs and is pasted below:

mpirun -np 12 /sw/pub/apps/lammps_fortran_mpi_gpu/bin/lmp_fotran_mpi_gpu -sf gpu -pk gpu 1 -in mixturePetro_293.txt

LAMMPS (22 Jul 2025)
OMP_NUM_THREADS environment is not set. Defaulting to 1 thread.
using 1 OpenMP thread(s) per MPI task


UCL Error: Error compiling PTX Program…

I have compatibles nvidia driver and nvcc compiler:
nvidia-smi
Wed Aug 6 13:03:18 2025
±----------------------------------------------------------------------------------------+
| NVIDIA-SMI 570.172.08 Driver Version: 570.172.08 CUDA Version: 12.8

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Fri_Feb_21_20:23:50_PST_2025
Cuda compilation tools, release 12.8, V12.8.93
Build cuda_12.8.r12.8/compiler.35583870_0

My configure:

cmake -D CMAKE_INSTALL_PREFIX=/sw/pub/apps/lammps_fortran_mpi_gpu -D BUILD_MPI=y -D LAMMPS_MACHINE=fortran_mpi_gpu -D CMAKE_C_COMPILER=gcc -D CMAKE_CXX_COMPILER=g++ -D CMAKE_FORTRAN_COMPILER=gfortran -D FFT=fftw3 -D BUILD_TOOLS=no -D PKG_GPU=yes -D GPU_API=CUDA -D GPU_PREC=mixed -D GPU_ARCH=sm_120 -D CUDA_MPS_SUPPORT=y -D PKG_ORIENT=ON -D PKG_OPENMP=ON -D PKG_MEAM=ON -D PKG_MANYBODY=ON -DBIN2C=/usr/local/cuda-12.8/bin/bin2c -D PKG_MOLECULE=on -D PKG_RIGID=on ../cmake

The compilation always work fine. I have used several GPU_ARCH values, from sm_50 to sm_120, and the behavior is the same. During cuda files compilations, there are not error mesages. The lammps executable works for CPU calculation. I think I have tried all possible solutions pointed others users. Anyone can help me?

Thanks

Actually, your compilation uses CUDA and not OpenCL. You should actually try OpenCL.

Your report looks like there is a CUDA driver issue and that could be an installation/configuration issue of you machine, rather than a LAMMPS issue. Setting GPU_ARCH has no impact for CMake builds and recent CUDA versions, because LAMMPS will compile a “fat” GPU binary which can run on all GPU architectures supported by the toolkit.

Ok, Axel. I wiil try the OpenCl.

Thanks