---
id: 20261120-T0-07
title: "生成式推荐推理提速：先优化Beam再优化Kernel"
title_en: "Optimizing Beam Search for Generative Recommendation Inference"
url: https://ai.daily.yangsir.net/daily/20261120-T0-07
issue_date: 2026-11-20
publish_date: 2026-11-19T05:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2611.23339
---

# 生成式推荐推理提速：先优化Beam再优化Kernel

生成式推荐（GR）推理与常规大模型服务差异明显：输入序列长、beam搜索宽、输出token数量固定且很少。现有vLLM等推理引擎把各beam当作独立批次处理，导致效率不佳。研究者提出先优化beam搜索、再优化底层kernel的思路，针对GR推理的负载特征重新设计beam调度方式。该方法可提升生成式推荐在线推理的吞吐和资源利用率，对部署大规模推荐系统的团队有直接参考价值。

## English Version

**Optimizing Beam Search for Generative Recommendation Inference**

Generative recommendation (GR) inference differs sharply from conventional LLM serving: long input sequences, wide beam search, and only a few fixed output tokens. Existing engines such as vLLM process beams as batched independent sequences, which is inefficient for this workload. The researchers propose optimizing beam search first and kernels second, redesigning beam scheduling around GR's specific load characteristics. The approach can improve throughput and resource utilization for online GR inference, offering practical guidance for teams deploying large-scale recommendation systems.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2611.23339)

**详情页**：https://ai.daily.yangsir.net/daily/20261120-T0-07

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*