Will llama.cpp take advantage of FP4 compute of the new chips?

[Sehyo]

Will llama.cpp take advantage of FP4 compute of the new chips? E.g. Blackwell GPUs.

[ExtReMLapin]

I could be wrong but if it uses MMQ, isn't that automatically using the specialized instructions as soon as you have cuda 12.8 installed ?

[ExtReMLapin]

@JohannesGaessler might be able to enlighten us on that

[JohannesGaessler]

FP4 (hardware) is currently not used at any point in the llama.cpp/GGML code and whether or not the feature is available will make no difference.