Generative concurrent data structures

Abstract

The design of efficient concurrent data structures in shared-memory multiprocessor sys tems is a critical yet challenging task. Developers must optimize synchronization and communication among threads, often requiring specialized skills in concurrent program ming. While large language models (LLMs) like GPT-4 and LLaMA exhibit remarkable performance in various domains, their high computational demands and environmental impact—emitting approximately 4.32 grams of CO2 per query—underscore the need for sustainable alternatives. To address these challenges, we propose fine-tuning smaller, task-specific models like GPT-2 and StarCoderBase 1B for concurrent programming. By leveraging techniques such as LoRA (Low-Rank Adaptation) and 8-bit optimization, we aim to maximize performance while minimizing computational overhead. Our approach involves creating a highly curated, domain-specific dataset in Java and C++ that includes synchronization code, shared-memory algorithms, and thread-safe data structure implementations. This dataset serves as the foundation for fine-tuning StarCoderBase 1B, a model with an architecture optimized for code-related tasks, in cluding auto-completion and algorithm optimization. Through lightweight fine-tuning methods like LoRA, combined with 8-bit quantization, we significantly reduce the mem ory and compute requirements of the model. This enables us to achieve comparable performance to larger LLMs in specialized tasks while drastically reducing environmental costs. Our work demonstrates the potential of smaller, efficient models in addressing the unique challenges of concurrent programming.