Hybrid GNN-Tree Model Proposed to Speed Up XLA Compiler Runtime Predictions

A developer has proposed a hybrid architecture combining a lightweight two-layer Graph Neural Network with gradient-boosted decision trees to predict XLA compiler runtimes more efficiently. The approach replaces deep neural network inference with tree-based regression, aiming to reduce peak RAM usage and accelerate compilation autotuning. Rather than relying on continuous gradient propagation, the system uses closed-form global graph metrics for feature extraction before passing data to a Scikit-Learn HistGradientBoostingRegressor. The design claims to scale cleanly across large NLP and XLA layout workloads without depending on physical hardware sensors. The author has made the codebase and notebooks publicly available and notes potential relevance to large-scale AI infrastructure such as Google's Gemini platform.