From Algorithm to System: Integrated Design of Compiler and Toolchain for Large Model Inference Optimization

With the rapid growth of deep learning model scales, especially large models such as Transformers and GPT, efficient inference has become a critical challenge due to increasing computational and memory demands. This paper proposes an integrated optimization framework that unifies algorithmic simplifications, compiler transformations, and system-level scheduling to enhance large model inference performance. By tightly coupling quantization, pruning, operator fusion, memory reuse, and automated heterogeneous hardware scheduling, the framework achieves significant improvements in computation reduction, memory efficiency, and parallel execution. Theoretical analysis and design considerations demonstrate the framework's potential for predictable performance gains and scalability across diverse hardware platforms. Future work will focus on extending hardware support, distributed inference, and adaptive optimization strategies. This integrated approach lays a foundation for efficient, scalable, and accurate large model deployment in practical AI applications.