Keke Zhai

Bi Xue, Hong Wu, Lei Chen et al.

Serving deep learning based recommendation models (DLRM) at scale is challenging. Existing approaches rely on dedicated ANN indexing and filtering services on CPUs, suffering from non-negligible costs and missing co-design opportunities. Such inefficiency makes them difficult to support complex model architectures, such as learned similarities and multi-task retrieval. In this paper, we present SilverTorch, a model-based serving system that brings all components into one unified model. It unifies model serving by replacing standalone indexing and filtering services with model layers. We propose a model-based GPU Bloom index for feature filtering and a fused Int8 ANN kernel for nearest neighbor search. Through co-design of the ANN search and feature filtering, we reduce GPU memory usage and eliminate computation. Benefiting from this design, we scale up retrieval by introducing an OverArch scoring layer and a multi-task retrieval with a Value Model to aggregate scores. These advancements improve the retrieval accuracy and enable future studies for serving more complex models. Our evaluation on industry-scale datasets show that SilverTorch achieves up to 23.7\times higher throughput compared to the state-of-the-art approaches. We also demonstrate that SilverTorch solution is 13.35\times more cost-efficient than CPU-based solution while improving accuracy via serving more complex models. SilverTorch is deployed at scale, serving hundreds of models online and supporting recommendation for diverse applications.

Liang Guo, Wei Li, Lucy Liao et al.

Deep Learning Recommendation Models (DLRMs) represent one of the largest machine learning applications on the planet. Industry-scale DLRMs are trained with petabytes of recommendation data to serve billions of users every day. To utilize the rich user signals in the long user history, DLRMs have been scaled up to unprecedented complexity, up to trillions of floating-point operations (TFLOPs) per example. This scale, coupled with the huge amount of training data, necessitates new storage and training algorithms to efficiently improve the quality of these complex recommendation systems. In this paper, we present a Request-Only Optimizations (ROO) training and modeling paradigm. ROO simultaneously improves the storage and training efficiency as well as the model quality of recommendation systems. We holistically approach this challenge through co-designing data (i.e., request-only data), infrastructure (i.e., request-only based data processing pipeline), and model architecture (i.e., request-only neural architectures). Our ROO training and modeling paradigm treats a user request as a unit of the training data. Compared with the established practice of treating a user impression as a unit, our new design achieves native feature deduplication in data logging, consequently saving data storage. Second, by de-duplicating computations and communications across multiple impressions in a request, this new paradigm enables highly scaled-up neural network architectures to better capture user interest signals, such as Generative Recommenders (GRs) and other request-only friendly architectures.

Keke Zhai

This paper delves into the continuous post-training optimization methods for small language models, and proposes a continuous post-training alignment data construction method for small language models. The core of this method is based on the data guidance of large models, optimizing the diversity and accuracy of alignment data. In addition, to verify the effectiveness of the methods in this paper, we used Qwen2-0.5B-Instruct model as the baseline model for small language models, using the alignment dataset constructed by our proposed method, we trained and compared several groups of experiments, including SFT (Supervised Fine Tuning) post-training experiment and KTO (Kahneman Tversky optimization) post-training experiment, as well as SFT-KTO two-stage post-training experiment and model weight fusion experiment. Finally, we evaluated and analyzed the performance of post-training models, and confirmed that the continuous post-training optimization method proposed by us can significantly improve the performance of small language models.

Keke Zhai

Currently, large models are prone to generating harmful content when faced with complex attack instructions, significantly reducing their defensive capabilities. To address this issue, this paper proposes a method based on constructing data aligned with multi-dimensional attack defense to enhance the generative security of large models. The core of our method lies in improving the effectiveness of safe alignment learning for large models by innova-tively increasing the diversity of attack instruction dimensions and the accuracy of generat-ing safe responses. To validate the effectiveness of our method, beyond existing security evaluation benchmarks, we additionally designed new security evaluation benchmarks and conducted comparative experiments using Llama3.2 as the baseline model. The final ex-perimental results demonstrate that our method can significantly improve the generative security of large models under complex instructional attacks, while also maintaining and enhancing the models' general capabilities.

Keke Zhai, Pan He, Tania Banerjee et al.

We propose SparsePipe, an efficient and asynchronous parallelism approach for handling 3D point clouds with multi-GPU training. SparsePipe is built to support 3D sparse data such as point clouds. It achieves this by adopting generalized convolutions with sparse tensor representation to build expressive high-dimensional convolutional neural networks. Compared to dense solutions, the new models can efficiently process irregular point clouds without densely sliding over the entire space, significantly reducing the memory requirements and allowing higher resolutions of the underlying 3D volumes for better performance. SparsePipe exploits intra-batch parallelism that partitions input data into multiple processors and further improves the training throughput with inter-batch pipelining to overlap communication and computing. Besides, it suitably partitions the model when the GPUs are heterogeneous such that the computing is load-balanced with reduced communication overhead. Using experimental results on an eight-GPU platform, we show that SparsePipe can parallelize effectively and obtain better performance on current point cloud benchmarks for both training and inference, compared to its dense solutions.