EnergyLens: Interpretable Closed-Form Energy Models for Multimodal LLM Inference Serving

2026-05-11

AI Summary

Existing approaches to optimizing inference energy consumption in multimodal large language models often misuse latency as a proxy or rely on data-intensive black-box models, resulting in poor generalization and limited interpretability. This work proposes a symbolic regression–based closed-form energy modeling method that explicitly links system-level attributes—such as parallelism degree, batch size, and sequence length—to energy consumption using only 12 parameters. For the first time, it decouples tensor and pipeline parallelism energy costs and separates the prefill and decode phases, yielding a physically interpretable model capable of extrapolating to unseen hardware and batch sizes without architectural modifications. Requiring just 50 profiling runs, the method achieves 88.2% accuracy in selecting the top-1 configuration, substantially outperforming analytical baselines (60.9%) and matching ensemble learning performance with tenfold fewer samples.

Abstract

As large language models span dense, mixture-of-experts, and state-space architectures and are deployed on heterogeneous accelerators under increasingly diverse multimodal workloads, optimising inference energy has become as critical as optimizing latency and throughput. Existing approaches either treat latency as an energy proxy or rely on data-hungry black-box surrogates. Both fail under varying parallelism strategies: latency and energy optima diverge in over 20% of configurations we tested, and black-box surrogates require hundreds of profiling samples to generalize across model families and hardware. We present EnergyLens, which uses symbolic regression as a structure-discovery tool over profiling data to derive a single twelve-parameter closed-form energy model expressed in terms of system properties such as degree of parallelism, batch size, and sequence length. Unlike black-box surrogates, EnergyLens decouples tensor and pipeline parallelism contributions and separates prefill from decode energy, making its predictions physically interpretable and actionable. Fitted from as few as 50 profiling measurements, EnergyLens achieves 88.2% Top-1 configuration selection accuracy across many evaluation scenarios compared to 60.9% for the closest prior analytical baseline, matches the predictive accuracy of ensemble ML methods with 10x fewer profiling samples, and extrapolates reliably to unseen batch sizes and hardware platforms without structural modification, making it a practical, interpretable tool for energy-optimal LLM deployment.

Problem

• energy optimization
• multimodal LLM inference
• closed-form energy model
• heterogeneous accelerators
• parallelism strategies

Innovation

• symbolic regression
• closed-form energy model
• interpretable AI
• LLM inference optimization
• multimodal serving