While I work on RustiCL... (until Intel finishes the Vulkan driver actually).
I'm trying to build llama.cpp and llama-cpp-python for OneAPI (actually with CLBlast). I ran make LLAMA_CLBLAST=1, with the basekit and opencl ocd installed in Conda. ocl-icd-system and I can say that Intel OpenCL Graphics seems detected by clinfo and I did not set the variable for RustiCL.
make LLAMA_CLBLAST=1 doesn't give acceleration. I always get BLAS=0 and runs exclusively on the CPU with no hint of any layers offloaded to GPU.
cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx is what they said to on the github to build for MKL. But that command ends up with ld cannot find crtbeginS.o, so I copied the /opt oneapi installation version to the /lib and /lib64 to see what happens, but that gives me -lgcc not found after that. So yes, the "proper" CMake way didn't work.
Any ideas on how to compile llama-cpp-python for oobabooga and llama.cpp alone? With Arc Acceleration?
There are some factors that indirectly play into economics.
AKA hiding your trade secrets.