OpenROAD-Assistant: An Open-Source Large Language Model for Physical Design Tasks

This paper presents OpenROAD-Assistant, an open-source chatbot designed for OpenROAD, leveraging only public data to respond to queries in prose or Python script using the OpenROAD APIs. OpenROAD-Assistant uses the foundational Llama3-8B model, enhanced with retrieval-aware fine-tuning (RAFT) for physical design-specific applications. OpenROAD-Assistant, its RAG database, and all related scripts are available on Zinodo and GitHub. It includes the prompt-script and question-answer adaptors and the associated codes to train the models and perform inference. The minimal hardware requirements are 2 CPU cores, 32GB of RAM, and 4 NVIDIA RTX A5500 GPUs for the Script Adaptor or 1 NVIDIA V100 GPU for the QA Adaptor. The software requirements are Python version >= 3.8 but <3.11.