4-bit quant of llama part of llava https://github.com/haotian-liu/LLaVA https://huggingface.co/liuhaotian/LLaVA-13b-delta-v0
quantized by:
CUDA_VISIBLE_DEVICES=0 python llama.py /workspace/LLaVA-13B-v0/ c4 --wbits 4 --true-sequential --groupsize 128 --save_safetensors llava-13b-v0-4bit-128g.safetensors
on https://github.com/oobabooga/GPTQ-for-LLaMa CUDA branch of GPTQ (commit 57a2629)
see: https://github.com/oobabooga/text-generation-webui/issues/1355#issuecomment-1518116906 and a PR which enables inference with images in webui: https://github.com/oobabooga/text-generation-webui/pull/1487