Fix: Issue with CUBLAS compilation error due to missing -fPIC flag

[B1gM8c]

Fix: Issue with CUBLAS compilation error due to missing -fPIC flag

[slaren]

This should have been fixed already with the addition of --forward-unknown-to-host-linker and $(CXXFLAGS) to the command line. Are you sure that it isn't working?

[B1gM8c]

--forward-unknown-to-host-linker

Yes, it does seem like it's not working. I'm getting a similar error message as what was reported in the previous issues when I try to compile it.
#1077

[slaren]

I think I see the issue, can you test if replacing --forward-unknown-to-host-linker with --forward-unknown-to-host-compiler works? We should pass all the C++ flags to the nvcc compiler anyway, only -fPIC is not enough.

[B1gM8c]

I think I see the issue, can you test if replacing --forward-unknown-to-host-linker with --forward-unknown-to-host-compiler works? We should pass all the C++ flags to the nvcc compiler anyway, only -fPIC is not enough.

You're right. I tested it, and modifying it to NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native works.

ifdef LLAMA_CUBLAS
	CFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64
	OBJS      += ggml-cuda.o
	NVCC      = nvcc
	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native
ggml-cuda.o: ggml-cuda.cu ggml-cuda.h

[slaren]

Superseded by #1128. I have kept your commits to give you credit for you work, though.