The paper introduces TokenRec, a novel framework for LLM-based recommender systems (RecSys) that addresses the challenges of tokenizing users and items into discrete tokens compatible with LLMs. TokenRec employs a Masked Vector-Quantized Tokenizer (MQ-Tokenizer) to quantize collaborative representations learned from advanced GNN-based methods, effectively capturing high-order collaborative knowledge. The generative retrieval paradigm in TokenRec efficiently recommends top-K items by generating item representations and retrieving appropriate items, avoiding the time-consuming auto-regressive decoding and beam search processes used by LLMs. Comprehensive experiments on four real-world datasets demonstrate that TokenRec outperforms competitive benchmarks, including traditional and emerging LLM-based RecSys, in terms of recommendation performance and generalization to unseen users/items. The proposed method significantly reduces inference time and computational resources, making it suitable for real-time recommendation scenarios.The paper introduces TokenRec, a novel framework for LLM-based recommender systems (RecSys) that addresses the challenges of tokenizing users and items into discrete tokens compatible with LLMs. TokenRec employs a Masked Vector-Quantized Tokenizer (MQ-Tokenizer) to quantize collaborative representations learned from advanced GNN-based methods, effectively capturing high-order collaborative knowledge. The generative retrieval paradigm in TokenRec efficiently recommends top-K items by generating item representations and retrieving appropriate items, avoiding the time-consuming auto-regressive decoding and beam search processes used by LLMs. Comprehensive experiments on four real-world datasets demonstrate that TokenRec outperforms competitive benchmarks, including traditional and emerging LLM-based RecSys, in terms of recommendation performance and generalization to unseen users/items. The proposed method significantly reduces inference time and computational resources, making it suitable for real-time recommendation scenarios.