Expert parallelism is vital for effectively training Mixture-of-Experts (MoE) models, enabling different devices to host distinct experts, with each device processing different input data. However, during expert parallel training, dynamic routing results in significant load imbalance among experts: a handful of overloaded experts hinder overall iteration, emerging as a training bottleneck. In this paper, we introduce LAER-MoE, an efficient MoE training framework. The core of LAER-MoE is a novel parallel paradigm, Fully Sharded Expert Parallel (FSEP), which fully partitions each expert parameter by the number of devices and restores partial experts at expert granularity through All-to-All communication during training. This allows for flexible re-layout of expert parameters during training to enhance load balancing. In particular, we perform fine-grained scheduling of communication operations to minimize communication overhead. Additionally, we develop a load balancing planner to formulate re-layout strategies of experts and routing schemes for tokens during training. We perform experiments on an A100 cluster, and the results indicate that our system achieves up to 1.69x acceleration compared to the current state-of-the-art training systems. Source code available at https://github.com/PKU-DAIR/Hetu-Galvatron/tree/laer-moe.

Distributed deep learning (DL) training faces instability from GPU/node failures of multi-GPU clusters, necessitating robust fault recovery from model checkpoints. However, we find that existing works only considers node failures but fails to handle partial GPU unavailability, and suffers from inefficient model checkpointing saving and loading, particularly when the GPU availability changes.

Advancements in model algorithms, the growth of foundational models, and access to high-quality datasets have propelled the evolution of Artificial Intelligence Generated Content (AIGC). Despite its notable successes, AIGC still faces hurdles such as updating knowledge, handling long-tail data, mitigating data leakage, and managing high training and inference costs. Retrieval-augmented generation (RAG) has recently emerged as a paradigm to address such challenges. In particular, RAG introduces the information retrieval process, which enhances the generation process by retrieving relevant objects from available data stores, leading to higher accuracy and better robustness. In this paper, we comprehensively review existing efforts that integrate RAG techniques into AIGC scenarios. We first classify RAG foundations according to how the retriever augments the generator, distilling the fundamental abstractions of the augmentation methodologies for various retrievers and generators. This unified perspective encompasses all RAG scenarios, illuminating advancements and pivotal technologies that help with potential future progress. We also summarize additional enhancement methods for RAG, facilitating effective engineering and implementation of RAG systems. Then from another view, we survey practical applications of RAG across different modalities and tasks, offering valuable references for researchers and practitioners. Furthermore, we introduce the benchmarks for RAG, discuss the limitations of current RAG systems, and suggest potential directions for future research.

Vertical federated learning (VFL) trains model when the features of data samples are scattered over multiple clients. To improve efficiency, a promising approach is to find a coreset of the data samples and use it as a smaller training set. However, existing methods produce a large coreset when there are many clients and have long running time. To address these problems, we propose HaCore for efficient coreset construction in VFL setting. HaCore first employs locality sensitive hashing (LSH) to map features to bit signatures locally on the clients, and then merges the local signatures for k-medoids clustering. Data samples that correspond to the medoids are added to the coreset. The core idea is that the distance of original data samples can be approximated by the Hamming distance between their LSH-based bit signatures. To accelerate k-medoids, we utilize an inverted index to search the nearest medoid and a bit-counting method to quickly compute the aggregate distance from many signatures to a medoid. We evaluate HaCore on 5 datasets and compare with state-of-the-art coreset construction methods for VFL. The results show that HaCore accelerates the best-performing baseline by over 45x and matches the accuracy of training with all samples.

Recently, using large language models (LLMs) for data augmentation has led to considerable improvements in unsupervised sentence embedding models. However, existing methods encounter two primary challenges: limited data diversity and high data noise. Current approaches often neglect fine-grained knowledge, such as entities and quantities, leading to insufficient diversity. Besides, unsupervised data frequently lacks discriminative information, and the generated synthetic samples may introduce noise. In this paper, we propose a pipeline-based data augmentation method via LLMs and introduce the Gaussian-decayed gradient-assisted Contrastive Sentence Embedding (GCSE) model to enhance unsupervised sentence embeddings. To tackle the issue of low data diversity, our pipeline utilizes knowledge graphs (KGs) to extract entities and quantities, enabling LLMs to generate more diverse samples. To address high data noise, the GCSE model uses a Gaussian-decayed function to limit the impact of false hard negative samples, enhancing the model’s discriminative capability. Experimental results show that our approach achieves state-of-the-art performance in semantic textual similarity (STS) tasks, using fewer data samples and smaller LLMs, demonstrating its efficiency and robustness across various models.

Extending the context length (i.e., the maximum supported sequence length) of LLMs is of paramount significance. To facilitate long context training of LLMs, sequence parallelism has emerged as an essential technique, which scatters each input sequence across multiple devices and necessitates communication to process the sequence. In essence, existing sequence parallelism methods assume homogeneous sequence lengths (i.e., all input sequences are equal in length) and therefore leverages a single, static scattering strategy for all input sequences. However, in reality, the sequence lengths in LLM training corpora exhibit substantial variability, often following a long-tail distribution, which leads to workload heterogeneity.

Recent foundation models are capable of handling multiple tasks and multiple data modalities with the unified base model structure and several specialized model components. However, efficient training of such multi-task (MT) multi-modal (MM) models poses significant system challenges due to the sophisticated model architecture and the heterogeneous workloads of different tasks and modalities. In this paper, we propose Spindle, a brand new training system tailored for resource-efficient and high-performance training of MT MM models via wavefront scheduling. The key idea of Spindle is to decompose the model execution into waves and address the joint optimization problem sequentially, including both heterogeneity-aware workload parallelization and dependency-driven execution scheduling. We build our system and evaluate it on various MT MM models. Experiments demonstrate the superior performance and efficiency of Spindle, with speedup ratio up to 71% compared to state-of-the-art training systems.

Sequence labeling remains a significant challenge in low-resource, domain-specific scenarios, particularly for character-dense languages.Existing methods primarily focus on enhancing model comprehension and improving data diversity to boost performance.However, these approaches still struggle with inadequate model applicability and semantic distribution biases in domain-specific contexts.To overcome these limitations, we propose a novel framework that combines an LLM-based knowledge enhancement workflow with a span-based Knowledge Fusion for Rich and Efficient Extraction (KnowFREE) model 1 .Our workflow employs explanation prompts to generate precise contextual interpretations of target entities, effectively mitigating semantic biases and enriching the model's contextual understanding.The KnowFREE model further integrates extension label features, enabling efficient nested entity extraction without relying on external knowledge during inference.Experiments on multiple domain-specific sequence labeling datasets demonstrate that our approach achieves stateof-the-art performance, effectively addressing the challenges posed by low-resource settings.

In recent years, Graph Neural Networks (GNNs) have demonstrated remarkable capabilities in learning from graph-structured data. However, GNNs face challenges when dealing with imperfect graph structures, which often lead to performance degradation due to the underlying message propagation mechanism. In response to this issue, a class of data-centric techniques called Graph Structure Learning (GSL) has emerged, with a focus on improving the quality of graph structures. Our review of the existing GSL literature, combined with empirical studies, reveals two primary limitations: low scalability and low efficiency. To mitigate these limitations, we introduce Random Walk-based Graph Structure Learning (RWGSL), a new GSL method that utilizes random walk strategies and operates in a parameter-free manner. Extensive experiments demonstrate that Rwgsl consistently improves the classification performance of both vanilla GNNs and advanced GSL methods across various graph datasets, and Rwgsl can scale to extremely large graphs (e.g. Ogbn-Products) with acceptable time cost. In particular, the combination of Rwgsl and GCN significantly reduces the run time to approximately 5% of those observed in most GSL methods, while also achieving a superior classification accuracy. These findings validate the high scalability and robustness of Rwgsl.

Due to the rising concerns on privacy protection, how to build machine learning models from distributed databases with privacy guarantees has gained more popularity. Vertical federated learning (VFL) trains machine learning models in a privacy-preserving way when the data features are scattered over distributed databases. We study the participant selection problem (PSP) for VFL, which chooses a given number of participants to conduct training while maximizing model accuracy. Compared to training with all participants, PSP can filter out hitch-riders that contribute marginally to model quality and reduce training time by involving fewer participants. To achieve good model accuracy, we formulate PSP as choosing a set of participants that maximizes the likelihood of the data samples. Then, utilizing the k-nearest neighbors (KNN) classifier as the proxy model, we express the likelihood as a function of the selected participants and prove that the function is sub modular. The submodular property is favorable as it can account for the feature diversity among the participants and allows to greedily select the participant with the maximum gain in each step. However, the selection process requires finding the top-k neighbors of a data sample as the basic operation, which is expensive in VFL setting as it involves encrypted communication. As such, we adapt the Fagin's algorithm, a famous top-k query algorithm, to reduce the amount of encrypted communication. We deploy our solution VFPS-SM across five distributed nodes and conduct experiments with 10 datasets and 3 models to evaluate its performance. The results show that VFPS-SM can reduce the end-to-end running time by up to$35\times$, selection time$365\times$and improve model accuracy by 6.0% compared with state-of-the-art baselines.

Mixture of Experts (MoE) is a widely used technique to expand model sizes for better model quality while maintaining the computation cost constant. In a nutshell, an MoE model consists of multiple experts in each model layer and routes the training tokens to only a fixed number of experts rather than all. In distributed training, as experts are distributed among different GPUs, All-to-All communication is necessary to exchange the training tokens among the GPUs after each time of expert routing. Due to the frequent and voluminous data exchanges, All-to-All communication has become a notable challenge to training efficiency. In this paper, we manage to accelerate All-to-All communication in MoE models from the training sample perspective, which is unexplored so far. In particular, we put forward the observation that tokens in the same training sample have certain levels of locality in expert routing. Motivated by this, we develop NetMoE, which takes such locality into account and dynamically rearranges the placement of training samples to minimize All-to-All communication costs. Specifically, we model the All-to-All communication given the sample placement and formulate an integer programming problem to deduce the optimal placement in polynomial time. Experiments with 32 GPUs show that NetMoE achieves a maximum efficiency improvement of $1.67 \times$ compared with current MoE training frameworks.

Recent advancements in Large Language Models (LLMs) have led to increasingly diverse requests, accompanied with varying resource (compute and memory) demands to serve them. However, this in turn degrades the cost-efficiency of LLM serving as common practices primarily rely on homogeneous GPU resources. In response to this problem, this work conducts a thorough study about serving LLMs over heterogeneous GPU resources on cloud platforms. The rationale is that different GPU types exhibit distinct compute and memory characteristics, aligning well with the divergent resource demands of diverse requests. Particularly, through comprehensive benchmarking, we discover that the cost-efficiency of LLM serving can be substantially optimized by meticulously determining GPU composition, deployment configurations, and workload assignments. Subsequently, we design a scheduling algorithm via mixed-integer linear programming, aiming at deducing the most cost-efficient serving plan under the constraints of price budget and real-time GPU availability. Remarkably, our approach effectively outperforms homogeneous and heterogeneous baselines under a wide array of scenarios, covering diverse workload traces, varying GPU availablilities, and multi-model serving. This casts new light on more accessible and efficient LLM serving over heterogeneous cloud resources.

Split learning is a prominent framework for vertical federated learning, where multiple clients collaborate with a central server for model training by exchanging intermediate embeddings. Recently, it is shown that an adversarial server can exploit the intermediate embeddings to train surrogate models to replace the bottom models on the clients (i.e., model stealing). The surrogate models can also be used to reconstruct private training data of the clients (i.e., data stealing). To defend against these stealing attacks, we propose Model Rake (i.e., Rake), which runs two bottom models on each client and differentiates their output spaces to make the two models distinct. Rake hinders the stealing attacks because it is difficult for a surrogate model to approximate two distinct bottom models. We prove that, under some assumptions, the surrogate model converges to the average of the two bottom models and thus will be inaccurate. Extensive experiments show that Rake is much more effective than existing methods in defending against both model and data stealing attacks, and the accuracy of normal model training is not affected.

Scaling long-context ability is essential for Large Language Models (LLMs). To amortize the memory consumption across multiple devices in long-context training, inter-data partitioning (a.k.a. Data Parallelism) and intra-data partitioning (a.k.a. Context Parallelism) are commonly used. Current training frameworks predominantly treat the two techniques as orthogonal, and establish static communication groups to organize the devices as a static mesh (e.g., a 2D mesh). However, the sequences for LLM training typically vary in lengths, no matter for texts, multi-modalities or reinforcement learning. The mismatch between data heterogeneity and static mesh causes redundant communication and imbalanced computation, degrading the training efficiency.

As the scale of models and training data continues to grow, there is an expanding reliance on more GPUs to train large-scale models, which inevitably increases the likelihood of encountering dynamic stragglers that some devices lag behind in performance occasionally. However, hybrid parallel training, one of the de facto paradigms to train large models, is typically sensitive to the stragglers. This paper presents Malleus , a straggler-resilient hybrid parallel training framework for large-scale models. Malleus quantifies the stragglers at the nuanced, per-GPU granularity during training, and develops a novel planning algorithm to deduce the optimal parallelization of GPU devices, pipeline stages, model layers, and training data, maximizing training efficiency when stragglers exist. In addition, once a shift in the straggler situation is detected, Malleus adaptively adjusts the parallelization via a re-planning process, and seamlessly and efficiently migrates the model states on the fly, without sacrificing the stability of the training tasks. Empirical results on large language models with up to 110B parameters show that Malleus consistently outperforms existing parallel training frameworks under various straggler situations, delivering on average 2.63-5.28x of efficiency improvement.

To optimize large Transformer model training, both efficient parallel computing and advanced data management are indispensable. However, current methods often assume a stable and uniform training workload, neglecting data-induced imbalances-arising from both sampling and packing processes-which can impede training performance. Specifically, data sampling imbalance arises from uneven sequence length distribution of the training data, while data packing imbalance stems from the discrepancy between the linear memory complexity and quadratic time complexity of the attention mechanism. To address these imbalance issues, we develop Hydraulis, which jointly optimizes the parallel strategies and data assignment. For one thing, we introduce large model training with dynamic heterogeneous parallel strategies in response to the sequence length variations within and across training iterations. For another, we devise a two-stage data assignment approach, which strikes a good balance in terms of the training workloads both within and across model replicas. Empirical results demonstrate that Hydraulis outperforms existing systems by 1.32-2.66×. Our source code is available: https://github.com/PKU-DAIR/Hetu.

As the field of Large Language Models (LLMs) continues to evolve, the context length in inference is steadily growing. Key-Value Cache (KVCache), the intermediate representations of tokens within LLM inference, has now become the primary memory bottleneck due to limited GPU memory. Current methods selectively determine suitable keys and values for self-attention computation in LLMs to address the issue. However, they either fall short in maintaining model quality or result in high serving latency. Drawing inspiration from advanced embedding retrieval techniques prevalent in the data management community, we consider the storage and retrieval of KVCache as a typical embedding retrieval problem. We propose PQCache , which employs Product Quantization (PQ) to manage KVCache, maintaining model quality while ensuring low serving latency. During the prefilling phase, we apply PQ to tokens' keys for each LLM layer and head. During the autoregressive decoding phase, we use PQ codes and centroids to approximately identify important preceding tokens, then fetch the corresponding key-value pairs for self-attention computation. Through meticulous design of overlapping and caching, we minimize any additional computation and communication overhead during both phases. Extensive experiments demonstrate that PQCache achieves both effectiveness and efficiency, with 4.60% score improvement over existing methods on InfiniteBench and low system latency in both prefilling and decoding.

Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing memory reuse across transformer layers. Empirical results demonstrate that MEMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.

With the breakthrough of Transformer-based pre-trained models, the demand for fine-tuning (FT) to adapt the base pre-trained models to downstream applications continues to grow, so it is essential for service providers to reduce the cost of processing FT requests. Low-rank adaption (LoRA) is a widely used FT technique that only trains small-scale adapters and keeps the base model unaltered, conveying the possibility of processing multiple FT tasks by jointly training different LoRA adapters with a shared base model. Nevertheless, through in-depth analysis, we reveal the efficiency of joint FT is dampened by two heterogeneity issues in the training data — the sequence length variation and skewness. To tackle these issues, we develop LobRA, a brand new framework that supports processing multiple FT tasks by jointly training LoRA adapters. Two innovative designs are introduced. Firstly, LobRA deploys the FT replicas (i.e., model replicas for FT) with heterogeneous resource usages and parallel configurations, matching the diverse workloads caused by the sequence length variation. Secondly, for each training step, LobRA takes account of the sequence length skewness and dispatches the training data among the heterogeneous FT replicas to achieve workload balance. We conduct experiments to assess the performance of LobRA, validating that it significantly reduces the GPU seconds required for joint FT by 45.03%-60.67%.

Vertical federated learning (VFL) trains models when multiple databases (a.k.a participants) hold different features of the same set of samples. By quantifying each participant's contribution to model training, data valuation can prevent hitch-riders and reward the instrumental parties. However, vertical federated data valuation (VFDV) is challenging because it needs to be accurate and efficient while protecting participant data privacy. In this paper, we propose a method meeting all three requirements by using projection and sampling for mutual information estimation (thus dubbed PS-MI). In particular, we first show that the utility of a participant set (a.k.a a coalition ) can be expressed as the mutual information (MI) between their features and the target labels. MI is favorable because it does not depend on the model to train (i.e., model-agnostic ) and can be estimated via k -nearest neighbor (KNN). To run KNN, instead of using costly homomorphic encryption to protect data privacy, we apply simple random projection to participant features before distance computation. We prove that random projection ensures differential privacy and preserves unbiased distance estimates. Since the contribution of a participant involves many coalitions, we adopt stratified sampling to reduce the number of coalitions while controlling estimation variance. To further improve efficiency, we incorporate optimizations including using locality sensitive hashing (LSH) to prune kNN candidates, batching kNN candidate checking for multiple coalitions, and adaptive early termination for utility evaluation. We compare PS-MI with 5 state-of-the-art VFDV methods. The results show that PS-MI yields higher accuracy and shorter running time than the baselines, and the maximum speedup can be 592×.

Motif detection is a graph algorithm that detects certain local structures in a graph. Although network motif has been studied in graph analytics, e.g., social network and biological network, it is yet unclear whether network motif is useful for analyzingonline transaction networkthat is generated in applications such as instant messaging and e-commerce. In an online transaction network, each vertex represents a user’s account and each edge represents a money transaction between two users. In this work, we try to analyze online transaction networks with network motifs. We design motif-based vertex embedding that integrates motif counts and centrality measurements. Furthermore, we design a distributed framework to detect motifs in large-scale online transaction networks. Our framework obtains the edge directions using a bi-directional tagging method and avoids redundant detection with a reduced view of neighboring vertices. We implement the proposed framework under the parameter server architecture. In the evaluation, we analyze different kinds of online transaction networks w.r.t the distribution of motifs and evaluate the effectiveness of motif-based embedding in downstream graph analytical tasks. The experimental results also show that our proposed motif detection framework can efficiently handle large-scale graphs.

Due to the recent success of diffusion models, text-to-image generation is becoming increasingly popular and achieves a wide range of applications. Among them, text-to-image editing, or continuous text-to-image generation, attracts lots of attention and can potentially improve the quality of generated images. It's common to see that users may want to slightly edit the generated image by making minor modifications to their input textual descriptions for several rounds of diffusion inference. However, such an image editing process suffers from the low inference efficiency of many existing diffusion models even using GPU accelerators. To solve this problem, we introduce Fast Image Semantically Edit (FISEdit), a cached-enabled sparse diffusion model inference engine for efficient text-to-image editing. The key intuition behind our approach is to utilize the semantic mapping between the minor modifications on the input text and the affected regions on the output image. For each text editing step, FISEdit can 1) automatically identify the affected image regions and 2) utilize the cached unchanged regions' feature map to accelerate the inference process. For the former, we measure the differences between cached and ad hoc feature maps given the modified textual description, extract the region with significant differences, and capture the affected region by masks. For the latter, we develop an efficient sparse diffusion inference engine that only computes the feature maps for the affected region while reusing the cached statistics for the rest of the image. Finally, extensive empirical results show that FISEdit can be 3.4 times and 4.4 times faster than existing methods on NVIDIA TITAN RTX and A100 GPUs respectively, and even generates more satisfactory images.

For graph self-supervised learning (GSSL), masked autoencoder (MAE) follows the generative paradigm and learns to reconstruct masked graph edges or node features while contrastive learning (CL) maximizes the similarity between augmented views of the same graph. Existing works utilize MAE and CL separately but we observe that the MAE and CL paradigms are complementary and propose the graph contrastive masked autoencoder (GCMAE) framework to unify them. Specifically, by focusing on local edges or node features, MAE cannot capture global information of the graph and is sensitive to particular edges and features. On the contrary, CL excels in extracting global information because it considers the relation between graphs. As such, we equip GCMAE with an MAE branch and a CL branch, and the two branches share a common encoder, which allows the MAE branch to exploit the global information extracted by the CL branch. To force GCMAE to capture global graph structures, we train it to reconstruct the entire adjacency matrix instead of only the masked edges as in existing works. Moreover, a discrimination loss is proposed for feature reconstruction, which improves the disparity between node embeddings rather than reducing the reconstruction error to tackle the feature smoothing problem of MAE. We evaluate GCMAE on four popular graph tasks (i.e., node classification, node clustering, link prediction, and graph classification) and compare it with 14 state-of-the-art baselines. The results show that GCMAE consistently provides good accuracy across these tasks, and the maximum accuracy improvement is up to 3.2% compared with the best-performing baseline.

Larger transformer models perform better on various downstream tasks but require more cost to scale up the model size. To efficiently enlarge models, the Mixture-of-Expert (MoE) architecture is widely adopted, which consists of a gate network and a series of experts and keep the training cost constant by routing the input data to a fixed number of experts instead of all. In existing large-scale MoE training systems, experts would be distributed among different GPUs for parallelization, and thus input data requires additional all-to-all communication to access the target expert and conduct corresponding computation. However, upon evaluating the training process of three mainstream MoE models on commonly used GPU clusters, we found that the all-to-all communication ratio averaged around 45\%, which significantly hinders the training efficiency and scalability of MoE models. In this paper, we propose LSH-MoE, a communication-efficient MoE training framework using locality-sensitive hashing (LSH). We first present the problems of scaling MoE training in existing systems and highlight the potential of exploiting token similarity to facilitate data compression. Then, we introduce an efficient LSH-based compression technique, which utilizes the cross-polytope hashing for rapid clustering and implements a residual-based error compensation scheme to alleviate the adverse impact of compression. To verify the effectiveness of our methods, we conduct experiments on both language models (e.g., RoBERTa, GPT, and T5) and vision models (e.g., Swin) for both pre-training and fine-tuning tasks. The results demonstrate that our method substantially outperforms its counterparts across different tasks by 1.28-2.2$\times$ of speedup.

With the remarkable achievements of large language models (LLMs), the demand for fine-tuning and deploying LLMs in various downstream tasks has garnered widespread interest. Parameter-efficient fine-tuning techniques represented by LoRA and model quantization techniques represented by GPTQ and AWQ are of paramount significance. However, although these techniques have been widely adopted in single-task scenarios, research is scarce in multi-task scenarios. To be specific, we find that mainstream quantization methods would prevent the base LLM from being shared among tasks, so current LLM serving systems are infeasible to integrate LLM quantization with multiple LoRA adapters to achieve memory-efficient multi-task serving. Moreover, existing LLM serving systems lack support for dynamic task addition and overlook the workload differences among tasks, leading to inefficiencies in multi-task scenarios. This work proposes LoRA-Inlaid, an efficient multi-task LLM serving system. On the one hand, LoRA-Inlaid designs a flexible and efficient multi-task quantization algorithm (MLGPTQ) that facilitates the sharing of a single quantized model for multiple LoRA adapters, which significantly reduces the memory consumption for model deployment. Meanwhile, it supports adding LoRA adapters for new tasks on the fly, without sacrificing the stability of online services. On the other hand, LoRA-Inlaid develops a novel multi-task scheduling algorithm guided by output length prediction and grouping among different tasks, which effectively shrinks the memory consumption and avoids frequent switching of LoRA adapters. Empirical results verify that LoRA-Inlaid outperforms existing state-of-the-art LLM serving systems by up to 1.58 times in terms of throughput, 1.76 times in terms of average latency, 2 times in terms of job completion time, and 10 times in terms of SLO Attainment, while maintaining the same level of model quality.

Training of large-scale deep learning models necessitates parallelizing the model and data across numerous devices, and the choice of parallelism strategy substantially depends on the training workloads such as memory consumption, computation cost, and communication cost. Current approaches generally assume uniform training workloads across samples in a given task. Thus, existing systems are designed to adopt a static parallelism strategy throughout one training process. Nevertheless, when training models with sequence inputs, this assumption fails due to the sequence length variation across samples. Consequently, training with a static parallelism strategy would result in sub-optimal performance.

One of the paradigms under which split learning (SL) is used is for the vertical federated learning (VFL) setting, where two or more parties build models over feature-partitioned data. However, to protect the private labels of one party, random noises are needed to perturb the backward derivatives (i.e., gradients w.r.t. forward activations), which incurs the privacy-utility tradeoff. In this work, we introduce PROJPERT,, a novel algorithm that searches for the optimal “perturbation knobs” for label protection in SL-based VFL. We formulate the problem of perturbation searching as how to minimize the impact on model quality given the desired privacy guarantee. Based on the problem, two solutions are introduced, where the first obtains the optimal perturbation via a simple but effective binary searching scheme, and the second heuristically approximates the optimality within a negligible error bound. Empirical results demonstrate that both our solutions are more effective in protecting the labels and achieve significantly better privacy-utility tradeoffs than state-of-the-art perturbation-based label protection methods. Furthermore, our heuristic solution is very efficient and incurs almost zero extra overhead in the overall running time, improving the usability in real-world applications.

Transformer models have emerged as the leading approach for achieving state-of-the-art performance across various application domains, serving as the foundation for advanced large-scale deep learning (DL) models. However, efficiently training these models across multiple GPUs remains a complex challenge due to the abundance of parallelism options. Existing DL systems either require manual efforts to design distributed training plans or limit parallelism combinations to a constrained search space. In this paper, we present Galvatron-BMW, a novel system framework that integrates multiple prevalent parallelism dimensions and automatically identifies the most efficient hybrid parallelism strategy. To effectively navigate this vast search space, we employ a decision tree approach for decomposition and pruning based on intuitive insights. We further utilize a dynamic programming search algorithm to derive the optimal plan. Moreover, to improve resource utilization and enhance system efficiency, we propose a bi-objective optimization workflow that focuses on workload balance. Our evaluations on different Transformer models demonstrate the capabilities of Galvatron-BMW in automating distributed training under varying GPU memory constraints. Across all tested scenarios, Galvatron-BMW consistently achieves superior system throughput, surpassing previous approaches that rely on limited parallelism strategies.

In global data analysis, the central server needs the global statistic of the user data stored in local clients. In such cases, an Honest-but-Curious central server might put user privacy at risk in trying to collect individual statistics of each user. In response, the secure aggregation provides a solution for calculating global statistics without revealing users’ privacy data. However, existing secure aggregation protocols only focus on the data in the form of vectors or common sets, which limits their application scope. We formalize a general problem—key-value set secure aggregation—that not only includes secure vector aggregation and private set union but also supports more applications. To address the proposed problem, we devise our solution (called the KVSAgg framework) that promises satisfactory performance in security, efficiency, and accuracy. Our key technique is a homomorphic transform algorithm (called HyperIBLT) that is not only capable of bidirectionally transforming data between key-value sets and vectors, but also able to transform sum operation of sets to addition of vectors. We implement KVSAgg on both CPU and GPU platforms and perform the evaluation on three use cases including federated learning, distributed data counting, and finding global hot items. Compared with our baselines, KVSAgg simultaneously achieves the best security, efficiency higher by orders of magnitude, and zero-error in nearly all cases. All codes are open-source anonymously.

Large-scale deep learning models contribute to significant performance improvements on varieties of downstream tasks. Current data and model parallelism approaches utilize model replication and partition techniques to support the distributed training of ultra-large models. However, directly deploying these systems often leads to sub-optimal training efficiency due to the complex model architectures and the strict device memory constraints. In this paper, we propose Optimal Sharded Data Parallel (OSDP), an automated parallel training system that combines the advantages from both data and model parallelism. Given the model description and the device information, OSDP makes trade-offs between the memory consumption and the hardware utilization, thus automatically generates the distributed computation graph and maximizes the overall system throughput. In addition, OSDP introduces operator splitting to further alleviate peak memory footprints during training with negligible overheads, which enables the trainability of larger models as well as the higher throughput. Extensive experimental results of OSDP on multiple different kinds of large-scale models demonstrate that the proposed strategy outperforms the state-of-the-art in multiple regards.

Recent years have witnessed the unprecedented achievements of large-scale pre-trained models, especially Transformer models. Many products and services in Tencent Inc., such as WeChat, QQ, and Tencent Advertisement, have been opted in to gain the power of pre-trained models. In this work, we present Angel-PTM, a productive deep learning system designed for pre-training and fine-tuning Transformer models. Angel-PTM can train extremely large-scale models with hierarchical memory efficiently. The key designs of Angel-PTM are a fine-grained memory management via the Page abstraction and a unified scheduling method that coordinates computations, data movements, and communications. Furthermore, Angel-PTM supports extreme model scaling with SSD storage and implements a lock-free updating mechanism to address the SSD I/O bottlenecks. Experimental results demonstrate that Angel-PTM outperforms existing systems by up to 114.8% in terms of maximum model scale as well as up to 88.9% in terms of training throughput. Additionally, experiments on GPT3-175B and T5-MoE-1.2T models utilizing hundreds of GPUs verify our strong scalability.

Network motif is a kind of frequently occurring subgraph that reflects local topology in graphs. Although network motif has been studied in graph analytics, e.g., social network and biological network, it is yet unclear whether network motif is useful for analyzing online transaction network that is generated in applications such as instant messaging and e-commerce. In this work, we analyze online transaction networks from the perspective of network motif. We define vertex features based on size-2 and size-3 motifs, and introduce motif-based centrality measurements. We further design motif-based vertex embedding that integrates weighted motif counts and centrality measurements. Afterward, we implement a distributed framework for motif detection in large-scale online transaction networks. To understand the effectiveness of motif for analyzing online transaction network, we study the statistical distribution of motifs in various kinds of graphs in Tencent and assess the benefit of motif-based embedding in a range of downstream graph analytical tasks. Empirical results show that our proposed method can efficiently find motifs in large-scale graphs, help interpretability, and benefit downstream tasks.

Vertical Federated Learning (VFL), that trains federated models over vertically partitioned data, has emerged as an important learning paradigm. However, existing VFL methods are facing two challenges: (1) scalability when # participants grows to even modest scale and (2) diminishing return w.r.t. # participants: not all participants are equally important and many will not introduce quality improvement in a large consortium. Inspired by these two challenges, in this paper, we ask: How can we select l out of m participants, where l ≪ m, that are most important?We call this problem Vertically Federated Participant Selection, and model it with a principled mutual information-based view. Our first technical contribution is VF-MINE—a Vertically Federated Mutual INformation Estimator—that uses one of the most celebrated algorithms in database theory—Fagin’s algorithm as a building block. Our second contribution is to further optimize VF-MINE to enable VF-PS, a group testing-based participant selection framework. We empirically show that vertically federated participation selection can be orders of magnitude faster than training a full-fledged VFL model, while being able to identify the most important subset of participants that often lead to a VFL model of similar quality.

Due to the rising concerns on privacy protection, how to build machine learning (ML) models over different data sources with security guarantees is gaining more popularity. Vertical federated learning (VFL) describes such a case where ML models are built upon the private data of different participated parties that own disjoint features for the same set of instances, which fits many real-world collaborative tasks. Nevertheless, we find that existing solutions for VFL either support limited kinds of input features or suffer from potential data leakage during the federated execution. To this end, this paper aims to investigate both the functionality and security of ML modes in the VFL scenario.

Vertical federated learning (VFL) is an emerging paradigm that allows different parties (e.g., organizations or enterprises) to collaboratively build machine learning models with privacy protection. In the training phase, VFL only exchanges the intermediate statistics, i.e., forward activations and backward derivatives, across parties to compute model gradients. Nevertheless, due to its geo-distributed nature, VFL training usually suffers from the low WAN bandwidth. In this paper, we introduce CELU-VFL, a novel and efficient VFL training framework that exploits the local update technique to reduce the cross-party communication rounds. CELU-VFL caches the stale statistics and reuses them to estimate model gradients without exchanging the ad hoc statistics. Significant techniques are proposed to improve the convergence performance. First, to handle the stochastic variance problem, we propose a uniform sampling strategy to fairly choose the stale statistics for local updates. Second, to harness the errors brought by the stateness, we devise an instance weighting mechanism that measures the reliability of the estimated gradients. Theoretical analysis proves that CELU-VFL achieves a similar sub-linear convergence rate as vanilla VFL training but requires much fewer communication rounds. Empirical results on both public and real-world workloads validate that CELU-VFL can be up to six times faster than the existing works.

With the ever-evolving concerns on privacy protection, vertical federated learning (FL), where participants own non-overlapping features for the same set of instances, is becoming a heated topic since it enables multiple enterprises to strengthen the machine learning models collaboratively with privacy guarantees. Nevertheless, to achieve privacy preservation, vertical FL algorithms involve complicated training routines and time-consuming cryptography operations, leading to slow training speed.

Recent years have witnessed intensive research interests on training deep neural networks (DNNs) more efficiently by quantization-based compression methods, which facilitate DNNs training in two ways: (1) activations are quantized to shrink the memory consumption, and (2) gradients are quantized to decrease the communication cost. However, existing methods mostly use a uniform mechanism that quantizes the values evenly. Such a scheme may cause a large quantization variance and slow down the convergence in practice. In this work, we introduce TinyScript, which applies a non-uniform quantization algorithm to both activations and gradients. TinyScript models the original values by a family of Weibull distributions and searches for ”quantization knobs” that minimize quantization variance. We also discuss the convergence of the non-uniform quantization algorithm on DNNs with varying depths, shedding light on the number of bits required for convergence. Experiments show that TinyScript always obtains lower quantization variance, and achieves comparable model qualities against full precision training using 1-2 bits less than the uniform-based counterpart.

Gradient boosting decision tree (GBDT) is a widely-used machine learning algorithm in both data analytic competitions and real-world industrial applications. Further, driven by the rapid increase in data volume, efforts have been made to train GBDT in a distributed setting to support large-scale workloads. However, we find it surprising that the existing systems manage the training dataset in different ways, but none of them have studied the impact of data management. To that end, this paper aims to study the pros and cons of different data management methods regarding the performance of distributed GBDT. We first introduce a quadrant categorization of data management policies based on data partitioning and data storage. Then we conduct an in-depth systematic analysis and summarize the advantageous scenarios of the quadrants. Based on the analysis, we further propose a novel distributed GBDT system named Vero, which adopts the unexplored composition of vertical partitioning and row-store and suits for many large-scale cases. To validate our analysis empirically, we implement different quadrants in the same code base and compare them under extensive workloads, and finally compare Vero with other state-of-the-art systems over a wide range of datasets. Our theoretical and experimental results provide a guideline on choosing a proper data management policy for a given workload.

Gradient boosting decision tree (GBDT) is one of the most popular machine learning models widely used in both academia and industry. Although GBDT has been widely supported by existing systems such as XGBoost, LightGBM, and MLlib, one system bottleneck appears when the dimensionality of the data becomes high. As a result, when we tried to support our industrial partner on datasets of the dimension up to 330K, we observed suboptimal performance for all these aforementioned systems. In this paper, we ask "Can we build a scalable GBDT training system whose performance scales better with respect to dimensionality of the data?"

To address the challenge of explosive big data, distributed machine learning (ML) has drawn the interests of many researchers. Since many distributed ML algorithms trained by stochastic gradient descent (SGD) involve communicating gradients through the network, it is important to compress the transferred gradient. A category of low-precision algorithms can significantly reduce the size of gradients, at the expense of some precision loss. However, existing low-precision methods are not suitable for many cases where the gradients are sparse and nonuniformly distributed. In this paper, we study is there a compression method that can efficiently handle a sparse and nonuniform gradient consisting of key-value pairs?