Optimizing scientific software is a difficult task because codebases are
often large and complex, and performance can depend upon several factors
including the algorithm, its implementation, and hardware among others. Causes
of poor performance can originate from disparate sources and be difficult to
diagnose. Recent years have seen a multitude of work that use large language
models (LLMs) to assist in software development tasks. However, these tools are
trained to model the distribution of code as text, and are not specifically
designed to understand performance aspects of code. In this work, we introduce
a reinforcement learning based methodology to align the outputs of code LLMs
with performance. This allows us to build upon the current code modeling
capabilities of LLMs and extend them to generate better performing code. We
demonstrate that our fine-tuned model improves the expected speedup of
generated code over base models for a set of benchmark tasks from 0.9 to 1.6
for serial code and 1.9 to 4.5 for OpenMP code.

通过引入强化学习的方法，将代码 LMM 的输出与性能对齐，提高生成代码的期望加速比，并在一组基准任务中展示了 0.9 至 1.6 倍的串行代码速度提升和 1.9 至 4.5 倍的 OpenMP 代码速度提升。

性能对齐的 LLM 用于生成高效代码

Performance-Aligned LLMs for Generating Fast Code

Large language models (LLMs) are playing an increasingly important role in
science and engineering. For example, their ability to parse and understand
human and computer languages makes them powerful interpreters and their use in
applications like code generation are well-documented. We explore the ability
of the GPT-4 LLM to ameliorate two major challenges in computational materials
science: i) the high barriers for adoption of scientific software associated
with the use of custom input languages, and ii) the poor reproducibility of
published results due to insufficient details in the description of simulation
methods. We focus on a widely used software for molecular dynamics simulations,
the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS), and
quantify the usefulness of input files generated by GPT-4 from task
descriptions in English and its ability to generate detailed descriptions of
computational tasks from input files. We find that GPT-4 can generate correct
and ready-to-use input files for relatively simple tasks and useful starting
points for more complex, multi-step simulations. In addition, GPT-4's
description of computational tasks from input files can be tuned from a
detailed set of step-by-step instructions to a summary description appropriate
for publications. Our results show that GPT-4 can reduce the number of routine
tasks performed by researchers, accelerate the training of new users, and
enhance reproducibility.

GPT-4 在计算材料科学中应用可以解决科学软件采用自定义输入语言的挑战以及由于描述模拟方法不足导致的研究结果复现性差的问题，通过生成正确的输入文件和详细的计算任务描述，减少研究者常规任务、加速新用户培训并提高结果可复现性。