Presentation Information
[B-6-55]Construction and Evaluation of a Distributed Code Generation System Using Cloud LLM Orchestration with Raspberry Pi
◎△Tomoyuki Kubo1, Daisuke Kondo2, Akihito Kohiga1, Takahiro Koita2 (1. Doshisha Univ., 2. Graduated School of Doshisha Univ.)
Keywords:
Orchestration,Raspberry Pi,Local LLM,Distributed Code Generation,Instruction Format Optimization
In recent years, rising LLM API costs and stricter token limits have made full reliance on cloud LLMs financially burdensome. In this study, we construct a distributed code generation system using Raspberry Pi 5 (AI HAT+ 2) as workers and a cloud LLM (Claude Haiku 4.5) as an orchestrator. The orchestrator generates instruction prompts per task, dispatched to workers running Qwen2.5-Coder-1.5B, which generate method implementations in parallel; results are then integrated via a rule-based pipeline. We evaluate the system on the first 20 samples of ClassEval, a Python class-level code generation benchmark, comparing five instruction format conditions using pass@1 as the metric. Results show that code-oriented formats—structured instruction and Python comment hints—improved pass@1 fourfold over the baseline (0.050 → 0.200). The two formats showed different strengths: structured instruction was more effective for arithmetic-type classes; Python comment hints for state-management-type classes. The upper bound of direct cloud LLM generation was pass@1 = 0.500. Future work will focus on upgrading worker models, dynamic instruction format selection, and evaluation on the full 100-sample dataset.