Improving HPC Code Generation Capability of LLMs via Online Reinforcement Learning with Real-Machine Benchmark Rewards

TL;DR

This paper introduces an online reinforcement learning method that uses real-machine benchmark rewards to enhance the HPC code generation capabilities of large language models, demonstrating improved performance on matrix multiplication tasks.

Contribution

It presents a novel online RL approach with real-machine feedback and a staged optimization algorithm to improve LLMs for HPC code generation.

Findings

Reinforcement learning with real-machine benchmarks improves code performance.

The staged optimization algorithm enables diverse code improvements.

Experimental results show enhanced HPC code generation capabilities.

Abstract

Large language models (LLMs) have demonstrated strong code generation capabilities, yet the runtime performance of generated code is not guaranteed, and there have been few attempts to train LLMs using runtime performance as a reward in the HPC domain. We propose an online reinforcement learning approach that executes LLM-generated code on a supercomputer and directly feeds back the measured runtime performance (GFLOPS) as a reward. We further introduce a Staged Quality-Diversity (SQD) algorithm that progressively varies the permitted optimization techniques on a per-problem basis, enabling the model to learn code optimization from diverse perspectives. We build a distributed system connecting a GPU training cluster with a CPU benchmarking cluster, and train Qwen2.5 Coder 14B on a double-precision matrix multiplication task using Group Relative Policy Optimization (GRPO). Through two experiments, we show that reinforcement learning combining runtime performance feedback with staged optimization can improve the HPC code generation capability of LLMs.