Model hallucinating on easy tasks

#3
by johnlockejrr - opened

Statistic Info, num_tokens=2273; generate_time(s)=13.6282; tps=166.7868; forward_step=425; num_boxes=378; bps=27.7366; prefill_time=2.3443; switch_to_ar=13

image

NVIDIA org

How about the slow mode? We find our model indeed performing poorly on OCR tasks when using the hybrid mode. We will fix this in the next version. @johnlockejrr

Sign up or log in to comment