lightblue/Karasu-Mixtral-8x22B-v0.1-gguf

This is the Q3_K_M GGUF port of the lightblue/Karasu-Mixtral-8x22B-v0.1 model.

How to use

The way to run this directly are using the llama.cpp package.

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make
huggingface-cli download lightblue/Karasu-Mixtral-8x22B-v0.1-gguf  --local-dir /some/folder/
./main -m /some/folder/Karasu-Mixtral-8x22B-v0.1-Q3_K_M-00001-of-00005.gguf -p "<s>[INST] Tell me a really funny joke. No puns! [/INST]" -n 256 -e

If you would like a nice easy GUI and have >64GB of RAM, then you could also run this using LM Studio and search for this model on the search bar.

Commands to make this:

cd llama.cpp

./convert-hf-to-gguf.py --outfile  /workspace/Karasu-Mixtral-8x22B-v0.1.gguf --outtype f16 /workspace/llm_training/axolotl/mixtral_8x22B_training/merged_model_multiling

./quantize /workspace/Karasu-Mixtral-8x22B-v0.1.gguf /workspace/Karasu-Mixtral-8x22B-v0.1-Q3_K_M.gguf Q3_K_M

./gguf-split  --split --split-max-tensors 128  /workspace/Karasu-Mixtral-8x22B-v0.1-Q3_K_M.gguf /workspace/split_gguf_q3km/Karasu-Mixtral-8x22B-v0.1-Q3_K_M