Actually it turns out the flag isn't wired correctly in the backend. If you use the --preemptible-worker-accelerator flag, it doesn't attach GPUs to the secondary workers. The flag in gcloud was prematurely made public before the feature was ready in the backend.
So for now, GPUs are not supported on preemptible workers. You can still use GPUs on primary workers. Sorry for the confusion.