3 ms·
Mixtral aka the 8x7B the "sparse mixture of experts" one is not the same as, eg. Mistral-7B which is still very, very good, just not quite hitting the mark on s
by brunooliv 3y ago
Mixtral aka the 8x7B the "sparse mixture of experts" one is not the same as, eg. Mistral-7B which is still very, very good, just not quite hitting the mark on some things.
I still couldn't run Mixtral 8x7B on an M1 Macbook Pro with 32Gb ram, so maybe I am indeed doing it wrong? Or are there better quantized versions available now or..?
- d-z-m 3y ago32 isn't quite enough to run a decent quant of mixtral(on a Macbook). You could try a Q3_K_M, but not sure how lobotomized it would be.
- CuriouslyC 3y agoThat's not true, the GGUF quants aren't great but there are exl2 4bit quantizations floating around that are pretty sweet.
- d-z-m 3y agofair enough, yeah I'm talking about GGUF quants only.
- brucethemoose2 3y agoexl2 is Nvidia/AMD only. But GGUF Mixtral should fit in 32GB... just not with the full 32K context. Long context is very memory intense in llama.cpp, at least until they fully implement flash attention and a quantized cache.