This behavior has been observed consistently with recent OpenCL runtimes.
The OpenCL infrastructure in the GPU package kernels is somewhat outdated and also the OpenCL loader code in the Windows binaries has not been updated in a long time.
Unfortunately, none of the LAMMPS developers currently have time or access to resources suitable to debugging and improving this situation.
The only recommendation we can currently give is to use a Linux machine or install a Linux partition on your Windows machine and compile and run LAMMPS with CUDA support natively on Linux.
Axel.