Compile without cublas dlls?

[vladfaust]

Is it possible to compile a llama binary without it requiring cublas64_12.dll and cublasLt64_12.dll in runtime? cudart64_12.dll is tiny, but cublas is around half a gig! I don't want to ship it with my app neither I want to make users install CUDA toolkit (cublas is not found when installing usual Nvidia drivers).

I tried setting -DGGML_CUDA_FORCE_MMQ=ON, but it still crashes because it can't find cublas64_12.dll in runtime.

[vladfaust]

FYI: https://siboehm.com/articles/22/CUDA-MMM.
Just saying that's theoretically possible to get rid of cublas in favor of custom kernels!