llama_cpp_python-0.3.16+cuda12.1.sm86.ampere-cp312-cp312-win_amd64.whl wheel is built against cuda 13
module fails to load .dll ggml-cuda.dll, and according to ntldd (from MSYS2), it searches for cublas64_13.dll
$ ntldd ggml-cuda.dll
ggml-base.dll (0x000001e615c90000)
cublas64_13.dll => not found
KERNEL32.dll => C:\Windows\SYSTEM32\KERNEL32.dll (0x000001e615e10000)
MSVCP140.dll => C:\Windows\SYSTEM32\MSVCP140.dll (0x000001e615e10000)
VCRUNTIME140.dll => C:\Windows\SYSTEM32\VCRUNTIME140.dll (0x000001e6138b0000)
VCRUNTIME140_1.dll => C:\Windows\SYSTEM32\VCRUNTIME140_1.dll (0x000001e6138b0000)