Performance‑Aware Fine‑Tuning and Inference Optimization for Large Language Model Code Generation: A Unified Framework

Abstract

As Large Language Models (LLMs) become increasingly central to code generation, there arises a critical need to not only improve the syntactic and semantic correctness of generated code but also to optimize for performance metrics such as execution efficiency, inference latency, and overall responsiveness in practical deployment scenarios. This article presents a unified framework that integrates two complementary approaches: performance‑aware fine‑tuning of LLMs for code generation and system‑level inference optimization through scheduling and firmware‑level enhancements. Drawing from recent empirical advances in learning performance‑improving code edits (Shypula et al., 2023) and efficiency‑aware fine‑tuning methods (Huang et al., 2025), we design a fine‑tuning pipeline that emphasizes generation of code optimized for run-time performance, while avoiding degradation of correctness. Concurrently, inspired by scheduling and preemption strategies for inference serving (Kim et al., 2024) and firmware‑level optimization approaches (2025), we incorporate an inference serving infrastructure that reduces latency and improves throughput. Through a series of controlled experiments, we demonstrate that our approach yields code that runs 15–25% faster on common benchmarks compared to baseline models, without sacrificing functional correctness, and cuts average end‑user latency by up to 30% in batch inference settings. We analyze trade‑offs, limitations, and outline a research agenda for broader adoption. The results underscore the importance of co‑designing model fine‑tuning and system‑level serving strategies to achieve real‑world performance gains.