ggml : fix race-condition in ggml-rpc

[gkpln3]

No description provided.

[rgerganov]

The RPC backend is always called in a single thread. What kind of crashes do you observe?

[gkpln3]

I'm adding RPC support to Ollama, which depends on this feature. In Ollama, it's called from multiple Go threads, leading to race conditions and crashes. I assumed ggml-rpc.cpp was intended to be thread-safe, given the use of mutexes elsewhere in the file.

[rgerganov]

While we use std::mutex in few places, the RPC backend is not considered thread-safe. I think that ggml backends in general are not required to be thread-safe by design, @slaren please correct me if I am wrong.

I'd suggest to add proper synchronization in the Go code.

[slaren]

Most ggml-backend objects are not intended to be thread safe, however it should still be possible to use different objects in different threads simultaneously. The RPC client reuses the same socket for multiple objects, so the synchronization here seems necessary.