LLM-Encoded Neural Architecture Descriptions for Cross-Model GPU Memory Prediction

Abstract

Accurate GPU memory prediction supports admission control, device selection, and capacity planning for deep learning workloads. Underestimation can cause out-of-memory failures after a job has already occupied an accelerator, whereas excessive safety margins reduce cluster utilization. This study evaluates an architecture-description approach on the GPUMemNet MLP, CNN, and Transformer training datasets. Each architecture is rendered as text from filename fields, scalar configuration variables, and its ordered layer inventory. A deterministic target-free architecture-language embedding is then fused with numeric and categorical design features. After removing 80 CNN rows with zero recorded target memory, the analysis contains 16,931 measured configurations. On a family-stratified 80/20 split, fused ExtraTrees achieved a mean absolute error of 1,059.9 MiB, a root mean square error of 2,335.0 MiB, an R2 of 0.9753, and 82.23% exact 4 GiB bucket accuracy. A numeric decision tree produced the lowest mixed-split MAE at 975.3 MiB, while fusion improved the ExtraTrees numeric baseline from 1,145.9 MiB to 1,059.9 MiB. Within-family fused ExtraTrees reached 12.8 MiB MAE for MLP, 858.3 MiB for CNN, and 914.8 MiB for Transformer. Leave-one-family-out performance was substantially weaker because the three families occupy different memory ranges and expose different allocation mechanisms. An auxiliary fused-feature OOM classifier reached 0.9702 F1. The findings show that architecture text is a useful semantic complement to numeric design variables, but dependable cross-family prediction still requires target-family observations and hardware-aware calibration.