Implement safe concurrency for a Faiss database with GPU

Question

I'm trying to setup a Faiss database for which many clients might ask to a single server. According to the docs https://github.com/facebookresearch/faiss/wiki/Threads-and-asynchronous-calls

it looks like in a cpu implementation, write operations would need to implement some sort of mutual exclusion.

However, in a gpu approach, it seems that even the read operations are not safe.

I'm working with Python. How could I implement something like a Semaphore for GPUs that will work for read/write operations, or at least read ones?