Is it possible to map device allocated memory to system memory?

while this was not possible in 2017, on the GH200 (Grace Hopper) and most likely on the future GB200 platform (Grace Blackwell) platforms this is fully possible.

The device memory shows up as a NUMA region, and both the CPU can access GPU memory and the GPU can access host memory seamlessly. This is because on the GH200 platform the CPU is connected to the GPU via a coherent off package chip-to-chip interconnect (NVLink-C2C).