This is the gptq 4bit quantization of this model: https://huggingface.co/jondurbin/airoboros-13b-gpt4 This quantization was made by using this repository: https://github.com/qwopqwop200/GPTQ-for-LLaMa/tree/triton And I used the triton branch with all the gptq implementations available (true_sequential + act_order + groupsize 128) CUDA_VISIBLE_DEVICES=0 python llama.py ./airoboros-13b-gpt4-TRITON c4 --wbits 4 --true-sequential --act-order --groupsize 128 --save_safetensors airoboros-13b-gpt4-128g-ts-ao.safetensors ``` Airoboros 13b gpt4 TRITON (g128 - ts - ao) PPL: 5.480927467346191 max memory(MiB): 8590.25 Airoboros 13b gpt4 CUDA (g128 - ts) PPL: 5.535770893096924 max memory(MiB): 8750.4697265625 ```