having trouble with making cuda backend work

[flyingshan]

trafficstars

I have tested with different backends provided by the repo. Most results are fine but only the cuda backend can not work as intended. When I run the cli command, the cuda backend printed the info of the gpus, then after a long-time computation, it output a wrong blending result (abs error [0, 0, 0]). Do you have any clue on how to solve this problem? Thank you!