GPU 服务双塔模型用于轻量级广告互动预测

Yuanlu Bai | Machine Learning Engineer II, L1 Conversion and Shopping Modeling; Yao Cheng | Sr. Machine Learning Engineer, L1 Conversion and Shopping Modeling; Xiao Yang | Sr. Staff Machine Learning Engineer, Ads Lightweight Ranking; Zhaohong Han | Manager II, Ads Lightweight Ranking; Jinfeng Zhuang | Sr. Manager, Ads Ranking

Introduction

引言

Lightweight ranking plays a crucial role as an intermediate stage in Pinterest’s ads recommendation system. Its main purpose is to efficiently narrow down the set of candidate ads before passing them to downstream, more complex ranking models. By doing so, it ensures that only the most relevant candidates move forward, improving both the efficiency and quality of our ads recommendations.

轻量级排名在 Pinterest 的广告推荐系统中扮演着关键的中间阶段角色。其主要目的是在将候选广告传递给下游更复杂的排名模型之前，有效缩小候选广告集。这样做确保只有最相关的候选者继续前进，从而提升广告推荐的效率和质量。

To balance model performance and serving latency, we adopted a classic two-tower paradigm. In this design, the Pin (ad) tower calculates Pin embeddings via offline batch updates, while the query (user) tower generates real-time embeddings. The prediction score is computed as the sigmoid of the dot product between the Pin and query embeddings. Previously, all two-tower models were served on CPUs. In 2025, we launched our first GPU-serving model for engagement prediction, which was an important milestone in the roadmap for next-generation infrastructure and model architecture.

为了平衡模型性能和服务延迟，我们采用了经典的 two-tower 范式。在这种设计中，Pin (ad) tower 通过离线批量更新计算 Pin embeddings，而 query (user) tower 生成实时 embeddings。预测分数计算为 Pin 和 query embeddings 点积的 sigmoid。此前，所有 two-tower 模型均在 CPU 上服务。2025 年，我们推出了首个用于 engagement prediction 的 GPU-serving 模型，这是下一代基础设施和模型架构路线图上的重要里程碑。

The new model arch...