metadata
license: apache-2.0
Experimental quants of 4 expert MoE mixtrals in various GGUF formats.
Goal is to have the best performing MoE < 10gb
Experimental q8 and q4 files for training/finetuning too.
- No sparsity tricks yet.
8.4gb custom 2bit quant works ok up until 512 token length then starts looping.
Install llama.cpp from github and run it:
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make -j
wget https://huggingface.co./nisten/quad-mixtrals-gguf/resolve/main/4mixq2.gguf
./server -m 4mixq2.gguf --host "ec2-3-99-206-122.ca-central-1.compute.amazonaws.com" -c 512
limit output to 500 tokens