Inconsistencies of performance with the official demo

#2
by Jinze - opened

Thank you for your demo, but there seems to be some inconsistency of performance with the official demo( link: https://modelscope.cn/studios/qwen/Qwen-VL-Chat-Demo/summary)

For example:

This demo
20230907112407.jpg
Official demo
20230907112437.jpg

This demo
20230907112807.jpg
Official demo
20230907112919.jpg

I believe this is caused by using the int4 model.

But I'll have to do more research tomorrow to confirm my suspicions. (It's now 02:40)

image.png

Sign up or log in to comment