Thanks for the information. But it seems not expected. Even if
amdgpu_device_initialize is called multiple times with different fds (returned
by opening the device nodes) in one single process, only one amdgpu_device
should be created, since a hash table is maitained for the created devices. In
this case, the per-device vamgr should only be initialized as well in one
process.I wrote a simple test application (see attached), and it worked for me
quite well. Could you try it on your failing system?