ggml : add support for dynamic loading of backends

[slaren]

Adds support for loading backends dynamically at load time, without needing to link to them in the build.

• Building the backends as dynamic libraries can be enabled by building with cmake with GGML_BACKEND_DL enabled
• Adds the function ggml_backend_load(const char * path) to load a backend dynamically
• Adds the convenience function ggml_backend_load_all(void) to load all the known backends
• Adds the function ggml_backend_unload(ggml_backend_reg_t reg) to unregister and unload a backend
• Adds the optional function ggml_backend_get_features to obtain a list of flags of a backend. This replaces the calls to the ggml_cpu_has_xx functions from the CPU backend in llama.cpp
• In addition to the CPU backend, the CUDA backend also implements ggml_backend_get_features, which returns the list of archs included in the build and the build flags used such as GGML_CUDA_FORCE_MMQ. Other backends should also implement this function to report compile-time flags and features.

TODO

• Version checking to avoid loading incompatible backends
• Fix ggml_backend_load_all search paths

[slaren]

GGML_USE_CPU now needs to be defined to use the CPU backend with the backend registry. This is necessary because the CPU backend now may be loaded dynamically, so it cannot be assumed that it is linked in the build. This may break other build scripts.

In Linux, it may also be necessary to link to dl for dlopen.

[Vali-98]

May want to mention this change in: #9289

I spent a few hours scratching my head on why I had no devices.

On the side, when no devices are loaded, this causes a segfault due to cpu_dev being a nullptr:

https://github.com/ggerganov/llama.cpp/blob/c9b00a70b080d5c0668608024afc3e0e2fed822f/src/llama.cpp#L7291-L7292

We probably should assert or something here, or perhaps anywhere when 0 devices are present. Let the user know something is wrong.

[MaggotHATE]

Is ggml_backend_load_all() supposed to be called in static builds too? If I don't use it, there's a noticeable reduction in quality of generated answers. When used, it tries to load any backend .dll it can find, which probably shouldn't happen on a static build (especially with a backend - OPENBLAS, for example). Am I doing something wrong?

[slaren]

It is ok to call ggml_backend_load_all even on static builds since it allows loading external backends. If it cannot find dynamically loadable backends present in the search paths, it won't do anything. The reduction in quality that you are observing is not likely to be caused by this.

[MaggotHATE]

it allows loading external backends.

It does, but allocates memory again, essentially duplicating total memory usage. I suppose it's a mistake on my end? It shouldn't behave like that on a combination of a static build with a dynamic backend?

[slaren]

It could happen if you have a static backend and the same backend as a dynamic backend, but that does not happen normally, because backends build without GGML_BACKEND_DL enabled cannot be loaded dynamically.

[paoletto]

is this exposed in llama-server?