Enabling Performant and Flexible Model-Internal Observability for LLM Inference

DMI-Lib decouples model-internal observability from inference hot path via asynchronous GPU-CPU memory abstraction (Ring²) and policy-controlled host backend for capturing internal LLM states.