Multiple GPU-Worker Protocol Issue

[Lihengwannafly]

We first test the star_server protocol on the CPU machine, and the training task runs normally. Now, we want to switch to the GPU machine. The cluster info is 2 PS node and 2 GPU-Worker node.
When in star_server protocol, the training task is failed with the ERROR /job:worker/replica:0/task:0/device:GPU:0 unknown device . But when in grpc++ and grpc, the training task runs normally.