Handle split GGUF sets end to end

Reclaim the GPU when a client stops listening; guidance on Thinking vs Instruct

Stop the engine when the client disconnects

AILOFlow: local LLM runtime with storage streaming

initial commit