The kernel calling subsystem uses the Boost.Bind library extensively. For this reason and due to C++ limitations, invoke functions take their arguments by reference and cannot, therefore, accept non-const temporaries or literal constants. This limitation in the current C++ standard is called the forwarding problem.
The MGPU configuration provides two defines, MGPU_MAX_DEVICES and MGPU_NUM_DEVICES. The former is the number of devices the library supports at most. The latter is a number determined for the current platform at compile-time and is not accurate or reliable binaries are moved between systems with different device configurations.
A device thread is bound to a physical device id (mgpu::dev_id_t). The device bound to device thread 0, which must not be the device with id 0 is the device rank 0 (mgpu::dev_rank_t).
All types that are only concerned with one device (all non-segmented types) are unaware of where they are located. They just exist somewhere, mainly in the context they were created in. The only exception is the device pointer. You may ask the device id from a device pointer (mgpu::dev_ptr::dev_id()). Segmented types always have rank/location information.