mosaicml

/

mpt-7b-instruct

like 461

Hi,

I would like to draw everyone's attention to AWQ quantized version of mpt-7b-instruct model at https://huggingface.co/abhinavkulkarni/mpt-7b-instruct-w4-g128-awq.

For more on AWQ, click here.

The quantized model size on disk is 3.76GB vs 13.3GB for the original model. Similar gains could be observed for VRAM usage. The perplexity is only worse by 5%.

Please take a look and give it a try.

Thanks!