Design and optimization of online assisted programming system based on SSA fine-tuning model

This paper proposes an online programming system design and optimization scheme based on Shifted Sparse Attention (SSA) fine-tuning model. By simultaneously applying SSA and Low-Rank adaptation (LoRA) in the self-attention layer, Under the framework of low-rank fine-tuning with local sparse attention transfer, the fine-grained fine-tuning of the large model's ability to expand the code context is realized, and the long sequence code understanding and code generation in programming scenarios are realized. The parameters of the fine-tuning model were optimized, the Normal Float (NF4), Normal Float (NF8) quantization and page attention were introduced, the performance and evaluation indicators of the system were optimized, which improved the real-time response ability and cost effectiveness of the online assistant programming system.