Backends Supported with llama.cpp

[yomultihead]

In addition to BLAS, llama.cpp supports various backends such as Vulkan, Kompute, and SYCL. Wanted to understand which backend is most suitable for Low Precision Inference on CPUs.

• Is there any support for low-precision GEMM?
• What are your thoughts on using BLAS libraries for CPU inferencing?
• Any resources/documents that document and outline the SOTA benchmarked results?