nvjpeg patch version affects a lot.

I build a face detecting program with TensorRT and nvjpeg whose lib is libnvjpeg.10.0 which links to libnvjpeg.10.0.130, works perfectly in my machine. good picture, good faces.

Then I pull a cuda10.0 docker image, and run and guess what?
Nothing good. Nothing bad. No errors, no faces.

It cost me 2 whole days to find out that the problem is in docker image, libnvjpeg.10.0 is a link to libnvjpeg.10.0.318 which i guess comes with the image.

Change the link, and it works!

But I feel disappointed.

Anybody have any idea about what happened?

What’s the different between .130 and .318?

Moving this to the CUDA Programming and Performance forum so the CUDA team can take a look.