tsumeone/llama-30b-supercot-3bit-128g-cuda

3bit quantized version of this: https://huggingface.co./ausboss/llama-30b-supercot

GPTQ quantization using https://github.com/0cc4m/GPTQ-for-LLaMa

Made at the request of someone that wanted a 3bit version. The file is 17% smaller than 4bit non-groupsize, but the wikitext2 ppl is 12% worse. I don't have a functioning Ooba install so I can't test this myself.

Command used to quantize:
python llama.py c:\llama-30b-supercot c4 --wbits 3 --true-sequential --groupsize 128 --save_safetensors 4bit-128g.safetensors

Evaluation & Score (Lower is better):

• WikiText2: 5.22 (12% worse than 4bit non-groupsize)
• PTB: 19.63 (11% worse than 4bit non-groupsize)
• C4: 6.93 (7% worse than 4bit non-groupsize)

4bit non-groupsize version is here: https://huggingface.co./tsumeone/llama-30b-supercot-4bit-cuda

4bit 128 groupsize version is here: https://huggingface.co./tsumeone/llama-30b-supercot-4bit-128g-cuda