Joachim

**Problem with `cudaStreamSynchronize` ?** I think the problems comes from the way Kokkos and MPI are put together. Somehow, a Cuda stream is synchronized (necessary operation for reduction), then it...

I confirm, just linking with Kokkos (no Kokkos::initialize) creates this behavior, unless I use `hpcbind`. Next week I will try to run on another cluster, see if I can recreate...

Sorry, I have a lot on my plate currently, I haven't had the time to reproduce it. I will come back in January, when I will work again on this.

Hi again, I tried on another cluster (1 node with two Tesla V100), the bug didn't appear. This means that the problem comes from the first cluster. I'm not sure...

And would they be any way to use the parser file/function without forking the project ? Because currently this API is private, which doesn't allow me to implement my own...