Launch the program

To launch a CUDA job, select CUDA on the Run dialog before you click Run/Submit. You can also enable memory debugging for CUDA programs in the CUDA section. See CUDA memory debugging for details.

It is not possible to attach to running CUDA programs if the program has already initialized the driver in some way, for example through having executed any kernel or called any functions from the CUDA library.

For MPI applications it is essential to place all CUDA initialization after the MPI_Init call.