Enabling Performant and Flexible Model-Internal Observability for LLM Inference

TL;DR

DMI-Lib is a high-speed, flexible system for internal observability of large language models during inference, significantly reducing latency overhead while maintaining detailed internal state access.

Contribution

It introduces DMI-Lib, a novel asynchronous observability system that decouples internal state monitoring from inference, enabling efficient, detailed internal signals collection.

Findings

Incurs only 0.4%-6.8% overhead in offline inference

Achieves an average of 6% overhead in online serving

Reduces latency overhead by 2x-15x compared to baselines

Abstract

Today's inference-time workloads increasingly depend on timely access to a model's internal states. We present DMI-Lib, a high-speed deep model inspector that treats internal observability as a first-class systems primitive, decoupling it from the inference hot path via an asynchronous observability substrate built from Ring^2, a GPU-CPU memory abstraction for capturing and staging tensors, and a policy-controlled host backend that exports them. DMI-Lib enables the placement of observation points across a rich space of internal signals and diverse inference backends while preserving serving optimizations and adhering to tight GPU memory budgets. Our experiments demonstrate that DMI-Lib incurs only 0.4%--6.8% overhead in offline batch inference and an average of 6% in moderate online serving, reducing latency overhead by 2x-15x compared to existing baselines with similar observability features. DMI-Lib is open-sourced at https://github.com/ProjectDMX/DMI.