GPU version does not refine unary potential, but CPU version does

[PhiAbs]

I managed to implement the crf-rnn layer into my custom model. When using the CPU version (master branch), the crf-rnn layer nicely refines my segmentation mask, which is great!

However, when I tried to use the GPU version, the output of the crf-rnn layer was still the same as its input, meaning that the segmentation map did not get refined. I used the same parameters as for the CPU version.

Did anyone encounter the same issue? Is there a fix for this problem?