CUDA out of memory, but still have free memory

after following the course Fastai v3. I have discover that, when I use learn.fit_one_cycle, I get OOM, but there is still a lot of memory left. I try reinstall conda, fastai, os, cuda and driver still not work.I use GTX 1080 Ti 11Gb, but it seem that the pytorch can only occupied 4Gb then OOM occurred.I try using custom code that occupied all of the vram and it worked, so that I think it might be some thing wrong with the fastai library because pytorch work fine.If you have any idea let me know my senior project deadline is near. :joy:
Here some screenshot:

Custom code:


Try decreasing your batch size from 64 to 32.

Thank for the reply, I try this also. I went as low as bs=8 for it to work, but problem still remain because it use within 4Gb budget any further allocation will cause OOM.