OpGPT: LLM-driven Operator Generation Prompt Framework for Domain-Specific Architectures

To address the growing need for operator libraries to leverage domain-specific architectures (DSAs) in the age of large language models (LLM), we propose OpGPT, an LLM-driven prompt framework for DSA operator library generation. By integrating methodologies including chain of thought, iterative refinement, and structural prompts, OpGPT generates operators with significantly reduced manual effort. We customize OpGPT for different DSAs by adding domain-specific information and enabling easy portability. Experiments show that OpGPT reduces manual tokens by 2%–30% compared to generic prompts while achieving up to 7× pass rate, demonstrating its efficiency. OpGPT’s code is available at https://anonymous.4open.science/r/OpGPT-5F24 .