cuda: remove linking to cublasLt

[yeahdongcn]

Make sure to read the contributing guidelines before submitting a PR

The new MUSA SDK now includes mublasLt (the equivalent of cublasLt). However, I found that llama.cpp doesn't use any of the cublasLt* APIs. Therefore, this PR removes cublasLt from the build link dependencies.

# ldd /usr/local/lib/python3.12/dist-packages/nvidia/cublas/lib/libcublas.so.12
        linux-vdso.so.1 (0x00007ffc7b9d8000)
        libcublasLt.so.12 => /usr/local/lib/python3.12/dist-packages/nvidia/cublas/lib/libcublasLt.so.12 (0x00007fdf7cc00000)
        librt.so.1 => /usr/lib/x86_64-linux-gnu/librt.so.1 (0x00007fdfa1857000)
        libpthread.so.0 => /usr/lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fdfa1852000)
        libdl.so.2 => /usr/lib/x86_64-linux-gnu/libdl.so.2 (0x00007fdfa184d000)
        libm.so.6 => /usr/lib/x86_64-linux-gnu/libm.so.6 (0x00007fdfa1764000)
        libgcc_s.so.1 => /usr/lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007fdfa1744000)
        libc.so.6 => /usr/lib/x86_64-linux-gnu/libc.so.6 (0x00007fdf7c9d7000)
        /lib64/ld-linux-x86-64.so.2 (0x00007fdfa1865000)

[slaren]

I am pretty sure that cublas depends on cublasLt, and building without it with the Makefile resulted in link errors. cmake adds the dependency automatically when linking to cublas, so removing it there is probably correct: https://gitlab.kitware.com/cmake/cmake/-/blob/master/Modules/FindCUDAToolkit.cmake#L1282. The Makefile is broken at this point, so I could not test it.

[yeahdongcn]

I am pretty sure that cublas depends on cublasLt, and building without it with the Makefile resulted in link errors. cmake adds the dependency automatically when linking to cublas, so removing it there is probably correct: https://gitlab.kitware.com/cmake/cmake/-/blob/master/Modules/FindCUDAToolkit.cmake#L1282. The Makefile is broken at this point, so I could not test it.

Let me revert the change to Makefile.

[JohannesGaessler]

From the CUDA documentation:

The cuBLASLt library is a new lightweight library dedicated to GEneral Matrix-to-matrix Multiply (GEMM) operations with a new flexible API.

We don't use cuBLASLt directly, but I guess they may be mapping the old cuBLAS API to cuBLASLt?

[yeahdongcn]

We don't use cuBLASLt directly, but I guess they may be mapping the old cuBLAS API to cuBLASLt?

In my understanding (though I’m not certain), cuBLAS might have a mechanism to automatically dispatch or fall back to cuBLASLt depending on the data types, memory layouts, or hardware.