Quantized version of this: https://huggingface.co/ausboss/llama-30b-supercot
GPTQ quantization using https://github.com/0cc4m/GPTQ-for-LLaMa for compatibility with 0cc4m's fork of KoboldAI
This one is without groupsize to save on VRAM, so that you can enjoy the full 2048 max context if you have 24GB VRAM (or at least get a lot closer to it versus the groupsize version)
Command used to quantize:python llama.py c:\llama-30b-supercot c4 --wbits 4 --act-order --true-sequential --save_safetensors 4bit.safetensors
Evaluation & Score (Lower is better):
- WikiText2: 4.66
- PTB: 17.64
- C4: 6.50
Groupsize version is here: https://huggingface.co/tsumeone/llama-30b-supercot-4bit-128g-cuda