Add cmake option to build without CUDA VMM

[WilliamTambellini]

Prerequisites

Please answer the following questions for yourself before submitting an issue.

• [X ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
• [X ] I carefully followed the README.md.
• [X ] I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• [X ] I reviewed the Discussions, and have a new bug or useful enhancement to share.

Feature Description

As today ggml-cuda.cu tries to take advantage of cuda VMM if possible:
https://github.com/ggerganov/llama.cpp/blob/784e11dea1f5ce9638851b2b0dddb107e2a609c8/ggml-cuda.cu#L116
This is not necessarilly possible/desired eg:
https://forums.developer.nvidia.com/t/potential-nvshmem-allocated-memory-performance-issue/275416/4

Would you mind if I add a cmake/make option/define in order to build ggml-cuda with(default) or without VMM support ?

Best
WT

Motivation

In order to build llamacpp/ggml without vmm.

Possible Implementation

Adding a cmake arg 'GGML_USE_CUDA_VMM' (default ON)
and then line 115:

#if !defined(GGML_USE_HIPBLAS) && defined(GGML_USE_CUDA_VMM)
        CUdevice device;
        CU_CHECK(cuDeviceGet(&device, id));
        CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
     ...

[slaren]

Feel free to open a PR if this is useful for you, I never was able to measure a performance difference. However, make the flag negative so that it is still enabled without additional compilation flags, and use the GGML_CUDA prefix (eg. something like GGML_CUDA_NO_VMM).