Large Language Models for EDA Cloud Job Resource and Lifetime Prediction

TL;DR

Predicting EDA cloud job resource usage and lifetime is challenging due to highly heterogeneous, semi-structured configurations. The authors cast this as a text-to-text regression problem and fine-tune decoder-only LLMs on serialized job configurations, optimizing the conditional distribution $P_{\theta}(\mathbf{y} \mid \mathbf{x}_{\text{job}})$. Their contributions include representing numerical outputs in fixed scientific notation, applying constrained decoding with deterministic prefix bypass, and performing full-attention fine-tuning to boost generation accuracy, validated on real-world DV datasets. The approach yields improved accuracy, robustness to temporal drift, and notable inference efficiency gains, offering a practical baseline for EDA cloud scheduling and capacity planning.

Abstract

The rapid growth of cloud computing in the Electronic Design Automation (EDA) industry has created a critical need for resource and job lifetime prediction to achieve optimal scheduling. Traditional machine learning methods often struggle with the complexity and heterogeneity of EDA workloads, requiring extensive feature engineering and domain expertise. We propose a novel framework that fine-tunes Large Language Models (LLMs) to address this challenge through text-to-text regression. We introduce the scientific notation and prefix filling to constrain the LLM, significantly improving output format reliability. Moreover, we found that full-attention finetuning and inference improves the prediction accuracy of sliding-window-attention LLMs. We demonstrate the effectiveness of our proposed framework on real-world cloud datasets, setting a new baseline for performance prediction in the EDA domain.

Figures (6)

• Figure 1: Architecture of LLM for EDA workload analysis. The diagram illustrates the system's four main components: (a) The data flow and conversational example of EDA jobs. (b) Supervised fine-tuning LLM with full attention mechanism to minimize CE loss on ground truth tokens. (c) Vanilla LLM decoding methods and Examples of Wrong generations. (d) The proposed constrained decoding method.
• Figure 2: Histograms illustrating the distribution of key resource metrics from dataset 1. All distributions are heavily right-skewed, indicating a long tail of resource-intensive jobs.
• Figure 3: True vs. predicted resource values from the Gemma-3-12B model on the test dataset.
• Figure 4: Comparison of inference times across Gemma-3 models.
• Figure 5: Ablation study comparing the impact of context visibility on performance across Gemma 3 models.