Measure framework overhead independently of attribution computation

Measure the overhead introduced by LumiXAI’s scheduling, serialization, and result-retrieval components separately from model inference and attribution computations.

Background

LumiXAI supports attribution across multiple model families and methods through a client–server architecture with asynchronous jobs, accelerator scheduling, serialization, and persistent result retrieval. The paper states that runtime is primarily determined by model inference, repeated forward or backward passes, diffusion, or generation steps, and expects the framework-specific components not to dominate latency.

A dedicated measurement remains unresolved because the reported system description does not quantify the time or relative cost attributable specifically to LumiXAI’s orchestration and infrastructure components. Such a benchmark would clarify the framework’s systems overhead independently of the external models and attribution methods it executes.