Large language models have demonstrated promising capabilities upon scaling up parameters. However, serving large language models incurs substantial computation and memory movement costs due to their large scale. Quantization methods have been employed to reduce service costs and latency. Nevertheless, outliers in activations hinder the development of INT4 weight-activation quantization. Existing approaches separate outliers and normal values into two matrices or migrate outliers from activations to weights, suffering from high latency or accuracy degradation. Based on observing activations from large language models, outliers can be classified into channel-wise and spike outliers. In this work, we propose Rotated Runtime Smooth (**RRS**), a plug-and-play activation smoother for quantization, consisting of Runtime Smooth and the Rotation operation. Runtime Smooth (**RS**) is introduced to eliminate **channel-wise outliers** by smoothing activations with channel-wise maximums during runtime. The Rotation operation can narrow the gap between **spike outliers** and normal values, alleviating the effect of victims caused by channel-wise smoothing. The proposed method outperforms the state-of-the-art method in the LLaMA and Qwen families and improves WikiText-2 perplexity from 57.33 to 6.66 for INT4 inference.

Large language models (LLMs) have demonstrated strong capabilities in handling long-context tasks, but processing such long contexts remains challenging due to the substantial memory requirements and inference latency. In this work, we discover that certain attention heads exhibit sequential consistency in their attention patterns, which can be persistently identified using a coefficient-of-variation-based algorithm. Inspired by this observation, we propose CateKV, a hybrid KV cache method that retains only critical token information for consistent heads, thereby reducing KV cache size and computational overhead, while preserving the majority of KV pairs in adaptive heads to ensure high accuracy. We show the unique characteristics of our algorithm and its extension with existing acceleration methods. Comprehensive evaluations on long-context benchmarks show that, while maintaining accuracy comparable to full attention, CateKV reduces memory usage by up to $2.72\times$ and accelerates decoding by $2.18\times$ in single-sample inputs, and boosts throughput by $3.96\times$ in batch scenarios.

Influenced by the great success of deep learning via cloud computing and the rapid development of edge chips, research in artificial intelligence (AI) has shifted to both of the computing paradigms, i.e., cloud computing and edge computing. In recent years, we have witnessed significant progress in developing more advanced AI models on cloud servers that surpass traditional deep learning models owing to model innovations (e.g., Transformers, Pretrained families), explosion of training data and soaring computing capabilities. However, edge computing, especially edge and cloud collaborative computing, are still in its infancy to announce their success due to the resource-constrained IoT scenarios with very limited algorithms deployed. In this survey, we conduct a systematic review for both cloud and edge AI. Specifically, we are the first to set up the collaborative learning mechanism for cloud and edge modeling with a thorough review of the architectures that enable such mechanism. We also discuss potentials and practical experiences of some on-going advanced edge AI topics including pretraining models, graph neural networks and reinforcement learning. Finally, we discuss the promising directions and challenges in this field.

Large-scale pretrained AI models have shown state-of-the-art accuracy in a series of important applications. As the size of pretrained AI models grows dramatically each year in an effort to achieve higher accuracy, training such models requires massive computing and memory capabilities, which accelerates the convergence of AI and HPC. However, there are still gaps in deploying AI applications on HPC systems, which need application and system co-design based on specific hardware features.

Sequential recommendation has become increasingly essential in various online services. It aims to model the dynamic preferences of users from their historical interactions and predict their next items. The accumulated user behavior records on real systems could be very long. This rich data brings opportunities to track actual interests of users. Prior efforts mainly focus on making recommendations based on relatively recent behaviors. However, the overall sequential data may not be effectively utilized, as early interactions might affect users' current choices. Also, it has become intolerable to scan the entire behavior sequence when performing inference for each user, since real-world system requires short response time. To bridge the gap, we propose a novel long sequential recommendation model, called Dynamic Memory-based Attention Network (DMAN). It segments the overall long behavior sequence into a series of sub-sequences, then trains the model and maintains a set of memory blocks to preserve long-term interests of users. To improve memory fidelity, DMAN dynamically abstracts each user's long-term interest into its own memory blocks by minimizing an auxiliary reconstruction loss. Based on the dynamic memory, the user's short-term and long-term interests can be explicitly extracted and combined for efficient joint recommendation. Empirical results over four benchmark datasets demonstrate the superiority of our model in capturing long-term dependency over various state-of-the-art sequential models.

Deep candidate generation (DCG) that narrows down the collection of relevant items from billions to hundreds via representation learning has become prevalent in industrial recommender systems. Standard approaches approximate maximum likelihood estimation (MLE) through sampling for better scalability and address the problem of DCG in a way similar to language modeling. However, live recommender systems face severe exposure bias and have a vocabulary several orders of magnitude larger than that of natural language, implying that MLE will preserve and even exacerbate the exposure bias in the long run in order to faithfully fit the observed samples. In this paper, we theoretically prove that a popular choice of contrastive loss is equivalent to reducing the exposure bias via inverse propensity weighting, which provides a new perspective for understanding the effectiveness of contrastive learning. Based on the theoretical discovery, we design CLRec, a contrastive learning method to improve DCG in terms of fairness, effectiveness and efficiency in recommender systems with extremely large candidate size. We further improve upon CLRec and propose Multi-CLRec, for accurate multi-intention aware bias reduction. Our methods have been successfully deployed in Taobao, where at least four-month online A/B tests and offline analyses demonstrate its substantial improvements, including a dramatic reduction in the Matthew effect.

Recent methods in sequential recommendation focus on learning an overall embedding vector from a user's behavior sequence for the next-item recommendation. However, from empirical analysis, we discovered that a user's behavior sequence often contains multiple conceptually distinct items, while a unified embedding vector is primarily affected by one's most recent frequent actions. Thus, it may fail to infer the next preferred item if conceptually similar items are not dominant in recent interactions. To this end, an alternative solution is to represent each user with multiple embedding vectors encoding different aspects of the user's intentions. Nevertheless, recent work on multi-interest embedding usually considers a small number of concepts discovered via clustering, which may not be comparable to the large pool of item categories in real systems. It is a non-trivial task to effectively model a large number of diverse conceptual prototypes, as items are often not conceptually well clustered in fine granularity. Besides, an individual usually interacts with only a sparse set of concepts. In light of this, we propose a novel Sparse Interest NEtwork (SINE) for sequential recommendation. Our sparse-interest module can adaptively infer a sparse set of concepts for each user from the large concept pool and output multiple embeddings accordingly. Given multiple interest embeddings, we develop an interest aggregation module to actively predict the user's current intention and then use it to explicitly model multiple interests for next-item prediction. Empirical results on several public benchmark datasets and one large-scale industrial dataset demonstrate that SINE can achieve substantial improvement over state-of-the-art methods.

Recently, neural networks have been widely used in e-commerce recommender systems, owing to the rapid development of deep learning. We formalize the recommender system as a sequential recommendation problem, intending to predict the next items that the user might be interacted with. Recent works usually give an overall embedding from a user's behavior sequence. However, a unified user embedding cannot reflect the user's multiple interests during a period. In this paper, we propose a novel controllable multi-interest framework for the sequential recommendation, called ComiRec. Our multi-interest module captures multiple interests from user behavior sequences, which can be exploited for retrieving candidate items from the large-scale item pool. These items are then fed into an aggregation module to obtain the overall recommendation. The aggregation module leverages a controllable factor to balance the recommendation accuracy and diversity. We conduct experiments for the sequential recommendation on two real-world datasets, Amazon and Taobao. Experimental results demonstrate that our framework achieves significant improvements over state-of-the-art models. Our framework has also been successfully deployed on the offline Alibaba distributed cloud platform.

Network embedding (or graph embedding) has been widely used in many real-world applications. However, existing methods mainly focus on networks with single-typed nodes/edges and cannot scale well to handle large networks. Many real-world networks consist of billions of nodes and edges of multiple types, and each node is associated with different attributes. In this paper, we formalize the problem of embedding learning for the Attributed Multiplex Heterogeneous Network and propose a unified framework to address this problem. The framework supports both transductive and inductive learning. We also give the theoretical analysis of the proposed framework, showing its connection with previous works and proving its better expressiveness. We conduct systematical evaluations for the proposed framework on four different genres of challenging datasets: Amazon, YouTube, Twitter, and Alibaba. Experimental results demonstrate that with the learned embeddings from the proposed framework, we can achieve statistically significant improvements (e.g., 5.99-28.23% lift by F1 scores; p<<0.01, t-test) over previous state-of-the-art methods for link prediction. The framework has also been successfully deployed on the recommendation system of a worldwide leading e-commerce company, Alibaba Group. Results of the offline A/B tests on product recommendation further confirm the effectiveness and efficiency of the framework in practice.

Attributed network embedding focuses on learning low-dimensional latent representations of nodes which can well preserve the original topological and node attributed proximity at the same time. Existing works usually assume that nodes with similar topology or similar attributes should also be close in the embedding space. This assumption ignores the phenomenon of partial correlation between network topological and node attributed similarities i.e. nodes with similar topology may be dissimilar in their attributes and vice versa. Partial correlation between the two information sources should be considered especially when there exist fraudulent edges (i.e., information from one source is vague) or unbalanced data distributions (i.e, topology structure similarity and node attribute similarity have different distributions). However, it is very challenging to consider the partial correlation between topology and attributes due to the heterogeneity of these two information sources. In this paper, we take partial correlation between topology and attributes into account and propose the Personalized Relation Ranking Embedding (PRRE) method for attributed networks which is capable of exploiting the partial correlation between node topology and attributes. The proposed PRRE model utilizes two thresholds to define different node relations and employs the Expectation-Maximization (EM) algorithm to learn these thresholds as well as other embedding parameters. Extensive experiments results on multiple real-world datasets show that the proposed PRRE model significantly outperforms the state-of-the-art methods in terms of various evaluation metrics.

Network representation learning (RL) aims to transform the nodes in a network into low-dimensional vector spaces while preserving the inherent properties of the network. Though network RL has been intensively studied, most existing works focus on either network structure or node attribute information. In this paper, we propose a novel framework, named ANRL, to incorporate both the network structure and node attribute information in a principled way. Specifically, we propose a neighbor enhancement autoencoder to model the node attribute information, which reconstructs its target neighbors instead of itself. To capture the network structure, attribute-aware skip-gram model is designed based on the attribute encoder to formulate the correlations between each node and its direct or indirect neighbors. We conduct extensive experiments on six real-world networks, including two social networks, two citation networks and two user behavior networks. The results empirically show that ANRL can achieve relatively significant gains in node classification and link prediction tasks.