Is it possible to use qwen_vl_2.5_32b as clip? Will the performance improve.
#14
by codexq - opened
As the text_encoder folder provide qwen_vl_2.5_7b, will qwen_vl_2.5_32b do a better work?
Since the pre-learning was originally done with "qwen_vl_2.5_7b", there is no reason to change it.
did you find an answer? I'm curious as well but it doesn't seem to work when I just swap the 7b clip for 32b