Llama-swap Matrix Enables Concurrent Models

This is a practical infrastructure upgrade, not a flashy feature: it turns llama-swap from “one model at a time” into a small local model scheduler with explicit resource rules. If you’re running OpenWebUI plus separate llama-server instances today, Matrix is probably the missing piece that lets you simplify the stack.

• –The README now calls out `matrix` as a custom DSL for running concurrent models, with control over how system resources are used.
• –That means you may not need separate always-on servers for every auxiliary task if the models can coexist in VRAM/RAM.
• –The tradeoff is complexity: Matrix helps when you understand your memory budget and traffic patterns, but it is not a magic concurrency switch.
• –For local stacks, this is most useful when you want a few models warm at the same time, not when you want to ignore hardware limits.
• –The feature also fits llama-swap’s core value prop: one OpenAI-compatible front door, with model loading policy pushed into config instead of manual process management.