Untitled

Since we introduce many new modules and processes, users may need to relearn the flow and often make iterative modifications to correctly run \textit{Open3DFlow} using TCL commands. To address this, we do another experiment for LLM-driven TCL script generation.

We extend the Keep Pre-Training (KPT) of the \textit{Llama-3}\cite{touvron2023llama} 8B model with a large corpus of TCL scripts to enhance its knowledge of IC physical design script language. Additionally, we integrate the documentation of \textit{Open3DFlow} as supplementary training material. During the KPT phase, we set the learning rate to 1.0e-5 and conducted training over 5 epochs, using a cutoff length of 1024. The training dataset was split in a 9:1 ratio for training and validation, and we utilized a cosine learning rate scheduler without warmup. Then we collaborate with experienced TCL script developers to design hundreds of QA pairs to perform Supervised Fine-Tuning (SFT) to align the model for our usage. For this SFT phase, we reduced the learning rate to 1.0e-6 and extended the training to 10 epochs, again implementing a cosine learning rate scheduler but this time with a 10% warmup period. The completion of these two stages resulted in the development of our effective TCL Generator, which is finely tuned to meet the specific needs of the Open3DFlow framework.

we use \textit{BLEU}\cite{papineni2002bleu} and \textit{ROUGE}\cite{lin2005recall} to evaluate our TCL quality. Their main purpose is to measure the similarity between the generated results and the reference summaries, so as to determine whether the generated text contains the right information. As shown in Figure \ref{fig:llm-eval}, our TCL generator achieved a very significant improvement in performance compared to the original Llama-3 8B model. This can significantly reduce the learning curve and debugging effort for users, allowing them to focus on optimization without being bogged down by scripting nuances.