Inferlock

$ cat CURRENT_FOCUS.md

Current focus

Keeping GPU coordination conservative, inspectable and small enough to fit local inference stacks.

What is established

• Exclusive leases prevent cooperating engines from fighting over one GPU.
• The API is intentionally narrower than a general scheduler.
• Designed around llama.cpp, vLLM and local operational constraints.

Known constraints

• Cooperating clients are required; Inferlock does not police arbitrary GPU processes.