Serving long context to many users in llama.cpp is not simple

With llama.cpp, a key question is whether an LLM with a 128k context can give that full length to every user at the same time. The example is a service with up to 8 parallel users.

The concern is that llama.cpp appears to share the 128k context across users instead of giving each user a separate 128k window. That means a system may need a different setup if every user must keep the full long-context capability during simultaneous use.