I built NIBIA Fabric, an open-source distributed LLM inference system that pools CPU and RAM across macOS, Linux, and Windows machines.
It builds on llama.cpp, with capacity-aware scheduling, adaptive memory reservation, persistent tensor caching, and an OpenAI-compatible API.
I’ve validated the current alpha on three physical machines across several models, including a 30B-class Qwen model, as well as GPT-OSS 20B. Feedback on the architecture and use cases is very welcome.
It builds on llama.cpp, with capacity-aware scheduling, adaptive memory reservation, persistent tensor caching, and an OpenAI-compatible API.
I’ve validated the current alpha on three physical machines across several models, including a 30B-class Qwen model, as well as GPT-OSS 20B. Feedback on the architecture and use cases is very welcome.