Does koboldcpp make use of tensor cores?

[VL4DST3R]

After playing a bit with "Chat with RTX" that released recently, I was wondering if/when the use of tensor cores will be added to kobold as well? I've found mentions about this being the case for things like llama.cpp but i've not seen it being mentioned much here except maybe here.

[LostRuins]

It uses them if MMQ is disabled

[VL4DST3R]

Oh so using --usecublas normal already makes use of them? That's neat! Re-reading the wiki page leads me to understand that it's better to stick to regular cublas then, given how most models are K variants? Also in your testing what does "slightly more VRAM" mean in this case? Something palpable/worth considering?

mmq is an upstream feature can be added to --usecublas to use quantized matrix multiplication in CUDA during prompt processing, instead of using cuBLAS for matrix multiplication. Experimentally this uses slightly less memory, and is slightly faster for Q4_0 but slower for K-quants. However, for long prompts on new GPUs, cuBLAS is generally faster at the cost of slightly more VRAM (MMQ off).