Gather each dtensor weight one at a time and send it to vllm; repeatedly call the weight update on vllm side