Recommendation systems based on Large Language Models (LLM4Rec) have emerged as a promising recommendation paradigm. However, efficient deployment of LLM4Rec on resource-constrained environment presents significant challenges in terms of computational efficiency and memory utilization. The LLM4Rec architecture shows unique memory and computation feature compared to traditional large language models. In this paper, we propose LLM4Rec-LIGHTNING, a high-throughput training system for LLM4Rec models. LLM4Rec-LIGHTNING introduces a performance model to help select the memory optimization technology without introducing latency. For offloaded model, LLM4Rec-LIGHTNING designs a novel task scheduler to overlap the model computation and offloading communication, enable efficient training. Experiment results show that LLM4Rec-LIGHTNING achieves up to 1.02 times higher throughput than other widely used methods. LLM4Rec-LIGHTNING can save over 25% communication time, significantly increasing efficiency.