Han Tian

Liu Yang, Di Chai, Junxue Zhang et al.

Vertical federated learning (VFL) is a promising category of federated learning for the scenario where data is vertically partitioned and distributed among parties. VFL enriches the description of samples using features from different parties to improve model capacity. Compared with horizontal federated learning, in most cases, VFL is applied in the commercial cooperation scenario of companies. Therefore, VFL contains tremendous business values. In the past few years, VFL has attracted more and more attention in both academia and industry. In this paper, we systematically investigate the current work of VFL from a layered perspective. From the hardware layer to the vertical federated system layer, researchers contribute to various aspects of VFL. Moreover, the application of VFL has covered a wide range of areas, e.g., finance, healthcare, etc. At each layer, we categorize the existing work and explore the challenges for the convenience of further research and development of VFL. Especially, we design a novel MOSP tree taxonomy to analyze the core component of VFL, i.e., secure vertical federated machine learning algorithm. Our taxonomy considers four dimensions, i.e., machine learning model (M), protection object (O), security model (S), and privacy-preserving protocol (P), and provides a comprehensive investigation.

Luxuan Chen, Han Tian, Xinran Chen et al.

The dynamic range of activations is a first-order constraint for low-bit quantization, activation scaling, and stable LLM inference. Prior work characterized outlier features and massive activations on pre-2024 LLaMA-style models, and the downstream activation-quantization stack inherits that picture without revisiting it for the post-LLaMA open-model boom. We ask the deployment-oriented question: how large can activations get in modern open LLMs, and how does this magnitude vary across families, generations, and training stages? Under a unified pipeline (5,000-sample multi-domain corpus, family-specific tokenization, identical hooks across embeddings, hidden states, attention, MLP/MoE, SwiGLU gates, and final norm), we measure global and layerwise maxima on 27 checkpoints from 8 open families spanning dense, MoE, vision-language, intermediate-training, and instruction-tuned variants. We find that (i) global maxima span over nearly four orders of magnitude at comparable parameter counts, with Qwen3.5 and MoE checkpoints in the 10^2 to 10^3 range and Gemma3-27B-it reaching ~7 x 10^5; (ii) cross-family and cross-generation comparisons break simple monotonic scaling; and (iii) MoE checkpoints exhibit 14.0-23.4x lower peaks than matched-scale dense counterparts, while the residual stream carries the global maximum in 22/24 checkpoints. A lightweight INT-8 sanity check shows that measured maxima co-vary with low-bit reconstruction error via activation-scale selection. We conclude that maximum activation magnitude is a model property tied to family, architecture, and training stage - not a simple byproduct of size - and should be measured and reported alongside any open-weight release before low-bit deployment. The code is publicly available at https://github.com/clx1415926/Max_act_llm.

Han Tian, Luxuan Chen, Xinran Chen et al.

Extending the context window of large language models typically requires training on sequences at the target length, incurring quadratic memory and computational costs that make long-context adaptation expensive and difficult to reproduce. We propose EndPrompt, a method that achieves effective context extension using only short training sequences. The core insight is that exposing a model to long-range relative positional distances does not require constructing full-length inputs: we preserve the original short context as an intact first segment and append a brief terminal prompt as a second segment, assigning it positional indices near the target context length. This two-segment construction introduces both local and long-range relative distances within a short physical sequence while maintaining the semantic continuity of the training text--a property absent in chunk-based simulation approaches that split contiguous context. We provide a theoretical analysis grounded in Rotary Position Embedding and the Bernstein inequality, showing that position interpolation induces a rigorous smoothness constraint over the attention function, with shared Transformer parameters further suppressing unstable extrapolation to unobserved intermediate distances. Applied to LLaMA-family models extending the context window from 8K to 64K, EndPrompt achieves an average RULER score of 76.03 and the highest average on LongBench, surpassing LCEG (72.24), LongLoRA (72.95), and full-length fine-tuning (69.23) while requiring substantially less computation. These results demonstrate that long-context generalization can be induced from sparse positional supervision, challenging the prevailing assumption that dense long-sequence training is necessary for reliable context-window extension. The code is available at https://github.com/clx1415926/EndPrompt.

Weiyan Wang, Yilun Jin, Yiming Zhang et al.

Due to high accuracy, BERT-like models have been widely adopted by text mining and web searching. However, large BERT-like models suffer from inefficient online inference, facing the following two problems on GPUs: (1) their high accuracy relies on the large model depth, which linearly increases the sequential computation on GPUs; (2) stochastic and dynamic online workloads cause extra costs from batching and paddings. Therefore, we present \sys for the real-world setting of GPU inference on online workloads. At its core, \sys adopts stacking distillation and boosting ensemble, distilling the original deep model into a group of shallow but virtually stacked student models running in parallel. This enables \sys to achieve a lower model depth (e.g., two layers) than the others and the lowest inference latency while maintaining accuracy. In addition, adaptive student pruning realizes dynamic student numbers according to changing online workloads. Especially for occasional workload bursts, it can temporarily decrease the student number with minimal accuracy loss to improve system throughput. We conduct comprehensive experiments to verify the effectiveness, whose results show that \sys outperforms the baselines by $4.1\times\sim 1.6\times$ in latency while maintaining accuracy and achieves up to $22.27\times$ higher throughput for workload bursts.

Yijun Sun, Xudong Liao, Songrun Xie et al.

Meeting stringent Time-To-First-Token (TTFT) requirements is crucial for LLM applications. To improve efficiency, modern LLM serving systems adopt disaggregated architectures with diverse parallelisms, introducing complex multi-stage workflows involving reusable KV-block retrieval, collective communication, and P2D transfer. Flows from dependent stages overlap within and across requests on shared bottleneck links, making TTFT highly susceptible to network contention and necessitating stage-aware scheduling. Unfortunately, most existing works schedule flows in a stage-agnostic manner, leading to uncoordinated contention that constitutes a primary cause of SLO violations. In this paper, we present MFS, a holistic multi-stage flow scheduling mechanism designed to maximize TTFT SLO attainment. At its core, MFS approximates the Least-Laxity-First (LLF) scheduling policy without requiring precise knowledge of a request's remaining slack. It achieves this through a Defer-and-Promote principle implemented through a Reverse Multi-Level Queue (RMLQ) structure. By dynamically promoting task precedence as effective laxity diminishes, MFS prioritizes flows with less laxity while preventing requests with loose SLOs from prematurely consuming network bandwidth. We implement MFS as a pluggable module integrated into vLLM, and evaluate it on a 8-server, 32-GPU testbed as well as through large-scale simulations. Our results demonstrate that MFS effectively outperforms state-of-the-art baselines, improving the TTFT SLO attainment by 1.2x--2.4x.

Fengfeng Li, Zhou Feng, Hongliang Wu et al.

In well logging operations using the oil-based mud (OBM) microresistivity imager, which employs an interleaved design with upper and lower pad sets, depth misalignment issues persist between the pad images even after velocity correction. This paper presents a depth matching method for borehole images based on the Shape Dynamic Time Warping (ShapeDTW) algorithm. The method extracts local shape features to construct a morphologically sensitive distance matrix, better preserving structural similarity between sequences during alignment. We implement this by employing a combined feature set of the one-dimensional Histogram of Oriented Gradients (HOG1D) and the original signal as the shape descriptor. Field test examples demonstrate that our method achieves precise alignment for images with complex textures, depth shifts, or local scaling. Furthermore, it provides a flexible framework for feature extension, allowing the integration of other descriptors tailored to specific geological features.

Xudong Liao, Han Tian, Chaoliang Zeng et al.

Recent years have witnessed a plethora of learning-based solutions for congestion control (CC) that demonstrate better performance over traditional TCP schemes. However, they fail to provide consistently good convergence properties, including {\em fairness}, {\em fast convergence} and {\em stability}, due to the mismatch between their objective functions and these properties. Despite being intuitive, integrating these properties into existing learning-based CC is challenging, because: 1) their training environments are designed for the performance optimization of single flow but incapable of cooperative multi-flow optimization, and 2) there is no directly measurable metric to represent these properties into the training objective function. We present Astraea, a new learning-based congestion control that ensures fast convergence to fairness with stability. At the heart of Astraea is a multi-agent deep reinforcement learning framework that explicitly optimizes these convergence properties during the training process by enabling the learning of interactive policy between multiple competing flows, while maintaining high performance. We further build a faithful multi-flow environment that emulates the competing behaviors of concurrent flows, explicitly expressing convergence properties to enable their optimization during training. We have fully implemented Astraea and our comprehensive experiments show that Astraea can quickly converge to fairness point and exhibit better stability than its counterparts. For example, \sys achieves near-optimal bandwidth sharing (i.e., fairness) when multiple flows compete for the same bottleneck, delivers up to 8.4$\times$ faster convergence speed and 2.8$\times$ smaller throughput deviation, while achieving comparable or even better performance over prior solutions.

Feiyuan Zhang, Dezhi Zhu, James Ming et al.

Retrieval-Augmented Generation (RAG) systems have shown substantial benefits in applications such as question answering and multi-turn dialogue \citep{lewis2020retrieval}. However, traditional RAG methods, while leveraging static knowledge bases, often overlook the potential of dynamic historical information in ongoing conversations. To bridge this gap, we introduce DH-RAG, a Dynamic Historical Context-Powered Retrieval-Augmented Generation Method for Multi-Turn Dialogue. DH-RAG is inspired by human cognitive processes that utilize both long-term memory and immediate historical context in conversational responses \citep{stafford1987conversational}. DH-RAG is structured around two principal components: a History-Learning based Query Reconstruction Module, designed to generate effective queries by synthesizing current and prior interactions, and a Dynamic History Information Updating Module, which continually refreshes historical context throughout the dialogue. The center of DH-RAG is a Dynamic Historical Information database, which is further refined by three strategies within the Query Reconstruction Module: Historical Query Clustering, Hierarchical Matching, and Chain of Thought Tracking. Experimental evaluations show that DH-RAG significantly surpasses conventional models on several benchmarks, enhancing response relevance, coherence, and dialogue quality.

Xudong Liao, Yijun Sun, Han Tian et al.

Mixture-of-Expert (MoE) models outperform conventional models by selectively activating different subnets, named experts, on a per-token basis. This gated computation generates dynamic communications that cannot be determined beforehand, challenging the existing GPU interconnects that remain static during the distributed training process. In this paper, we advocate for a first-of-its-kind system, called MixNet, that unlocks topology reconfiguration during distributed MoE training. Towards this vision, we first perform a production measurement study and show that the MoE dynamic communication pattern has strong locality, alleviating the requirement of global reconfiguration. Based on this, we design and implement a regionally reconfigurable high-bandwidth domain on top of existing electrical interconnects using optical circuit switching (OCS), achieving scalability while maintaining rapid adaptability. We have built a fully functional MixNet prototype with commodity hardware and a customized collective communication runtime that trains state-of-the-art MoE models with in-training topology reconfiguration across 32 A100 GPUs. Large-scale packet-level simulations show that MixNet delivers comparable performance as the non-blocking fat-tree fabric while boosting the training cost efficiency (e.g., performance per dollar) of four representative MoE models by 1.2x-1.5x and 1.9x-2.3x at 100 Gbps and 400 Gbps link bandwidths, respectively.

Di Chai, Pengbo Li, Feiyuan Zhang et al.

Token filtering has been proposed to enhance utility of large language models (LLMs) by eliminating inconsequential tokens during training. While using fewer tokens should reduce computational workloads, existing studies have not succeeded in achieving higher efficiency. This is primarily due to the insufficient sparsity caused by filtering tokens only in the output layers, as well as inefficient sparse GEMM (General Matrix Multiplication), even when having sufficient sparsity. This paper presents Collider, a system unleashing the full efficiency of token filtering in LLM training. At its core, Collider filters activations of inconsequential tokens across all layers to maintain sparsity. Additionally, it features an automatic workflow that transforms sparse GEMM into dimension-reduced dense GEMM for optimized efficiency. Evaluations on three LLMs-TinyLlama-1.1B, Qwen2.5-1.5B, and Phi1.5-1.4B-demonstrate that Collider reduces backpropagation time by up to 35.1% and end-to-end training time by up to 22.0% when filtering 40% of tokens. Utility assessments of training TinyLlama on 15B tokens indicate that Collider sustains the utility advancements of token filtering by relatively improving model utility by 16.3% comparing to regular training, and reduces training time from 4.7 days to 3.5 days using 8 GPUs. Collider is designed for easy integration into existing LLM training frameworks, allowing systems already using token filtering to accelerate training with just one line of code.

Liu Yang, Shuowei Cai, Di Chai et al.

As an essential tool of secure distributed machine learning, vertical federated learning (VFL) based on homomorphic encryption (HE) suffers from severe efficiency problems due to data inflation and time-consuming operations. To this core, we propose PackVFL, an efficient VFL framework based on packed HE (PackedHE), to accelerate the existing HE-based VFL algorithms. PackVFL packs multiple cleartexts into one ciphertext and supports single-instruction-multiple-data (SIMD)-style parallelism. We focus on designing a high-performant matrix multiplication (MatMult) method since it takes up most of the ciphertext computation time in HE-based VFL. Besides, devising the MatMult method is also challenging for PackedHE because a slight difference in the packing way could predominantly affect its computation and communication costs. Without domain-specific design, directly applying SOTA MatMult methods is hard to achieve optimal. Therefore, we make a three-fold design: 1) we systematically explore the current design space of MatMult and quantify the complexity of existing approaches to provide guidance; 2) we propose a hybrid MatMult method according to the unique characteristics of VFL; 3) we adaptively apply our hybrid method in representative VFL algorithms, leveraging distinctive algorithmic properties to further improve efficiency. As the batch size, feature dimension and model size of VFL scale up to large sizes, PackVFL consistently delivers enhanced performance. Empirically, PackVFL propels existing VFL algorithms to new heights, achieving up to a 51.52X end-to-end speedup. This represents a substantial 34.51X greater speedup compared to the direct application of SOTA MatMult methods.

Hao Wang, Jingrong Chen, Xinchen Wan et al.

Communication overhead poses an important obstacle to distributed DNN training and draws increasing attention in recent years. Despite continuous efforts, prior solutions such as gradient compression/reduction, compute/communication overlapping and layer-wise flow scheduling, etc., are still coarse-grained and insufficient for an efficient distributed training especially when the network is under pressure. We present DLCP, a novel solution exploiting the domain-specific properties of deep learning to optimize communication overhead of DNN training in a fine-grained manner. At its heart, DLCP comprises of several key innovations beyond prior work: e.g., it exploits {\em bounded loss tolerance} of SGD-based training to improve tail communication latency which cannot be avoided purely through gradient compression. It then performs fine-grained packet-level prioritization and dropping, as opposed to flow-level scheduling, based on layers and magnitudes of gradients to further speedup model convergence without affecting accuracy. In addition, it leverages inter-packet order-independency to perform per-packet load balancing without causing classical re-ordering issues. DLCP works with both Parameter Server and collective communication routines. We have implemented DLCP with commodity switches, integrated it with various training frameworks including TensorFlow, MXNet and PyTorch, and deployed it in our small-scale testbed with 10 Nvidia V100 GPUs. Our testbed experiments and large-scale simulations show that DLCP delivers up to $84.3\%$ additional training acceleration over the best existing solutions.

Qinghe Jing, Weiyan Wang, Junxue Zhang et al.

The scarcity of data and isolated data islands encourage different organizations to share data with each other to train machine learning models. However, there are increasing concerns on the problems of data privacy and security, which urges people to seek a solution like Federated Transfer Learning (FTL) to share training data without violating data privacy. FTL leverages transfer learning techniques to utilize data from different sources for training, while achieving data privacy protection without significant accuracy loss. However, the benefits come with a cost of extra computation and communication consumption, resulting in efficiency problems. In order to efficiently deploy and scale up FTL solutions in practice, we need a deep understanding on how the infrastructure affects the efficiency of FTL. Our paper tries to answer this question by quantitatively measuring a real-world FTL implementation FATE on Google Cloud. According to the results of carefully designed experiments, we verified that the following bottlenecks can be further optimized: 1) Inter-process communication is the major bottleneck; 2) Data encryption adds considerable computation overhead; 3) The Internet networking condition affects the performance a lot when the model is large.

Han Tian, Hankz Hankui Zhuo

Due to the availability of references of research papers and the rich information contained in papers, various citation analysis approaches have been proposed to identify similar documents for scholar recommendation. Despite of the success of previous approaches, they are, however, based on co-occurrence of items. Once there are no co-occurrence items available in documents, they will not work well. Inspired by distributed representations of words in the literature of natural language processing, we propose a novel approach to measuring the similarity of papers based on distributed representations learned from the citation context of papers. We view the set of papers as the vocabulary, define the weighted citation context of papers, and convert it to weight matrix similar to the word-word cooccurrence matrix in natural language processing. After that we explore a variant of matrix factorization approach to train distributed representations of papers on the matrix, and leverage the distributed representations to measure similarities of papers. In the experiment, we exhibit that our approach outperforms state-of-theart citation-based approaches by 25%, and better than other distributed representation based methods.