Pixtral-12B-2409: int4 Weight Quant

W4A16 quant of mistral-community/pixtral-12b using kylesayrs/gptq-partition branch of LLM Compressor for optimised inference on VLLM.

vision_tower kept at FP16. language_model weights quantized to 4bit.

Calibrated on 512 flickr samples.

Example VLLM usage

vllm serve nintwentydo/pixtral-12b-2409-W4A16-G128 --max-model-len 131072 --limit-mm-per-prompt 'image=4' 

If you want a more advanced/fully featured chat template you can use this jinja template