Fangcheng Fu

Yujie Wang, Siwei Chen, Longzan Luo et al.

Reinforcement Learning (RL) has become pivotal for improving model capabilities yet suffers from rollout efficiency bottlenecks due to the long-tail response length distribution. While existing works mitigate the impact of long tails via prompt-level tail scheduling, we focus on the root source of inefficiency: the distribution itself. Specifically, we characterize the long-tail distribution at a finer granularity, identifying intra-prompt long tails, and revealing that they frequently consist of ineffective verbosity. To address this, we propose a novel paradigm of active distribution shaping to shape the rollout distribution towards conciseness and certainty, thereby fundamentally resolving tail-induced overheads. We achieve this through a distribution-aware trajectory sampling mechanism, which selects trajectories from a redundant exploration space for each prompt, and an adaptive redundancy allocation scheme to maximize both shaping effectiveness and system efficiency. Experiments demonstrate significant acceleration over state-of-the-art systems by up to 1.77x without compromising model performance.

Qiang Huang, Xiao Yan, Xin Wang et al.

Temporal graph neural networks (TGNNs) outperform regular GNNs by incorporating time information into graph-based operations. However, TGNNs adopt specialized models (e.g., TGN, TGAT, and APAN ) and require tailored training frameworks (e.g., TGL and ETC). In this paper, we propose TF-TGN, which uses Transformer decoder as the backbone model for TGNN to enjoy Transformer's codebase for efficient training. In particular, Transformer achieves tremendous success for language modeling, and thus the community developed high-performance kernels (e.g., flash-attention and memory-efficient attention) and efficient distributed training schemes (e.g., PyTorch FSDP, DeepSpeed, and Megatron-LM). We observe that TGNN resembles language modeling, i.e., the message aggregation operation between chronologically occurring nodes and their temporal neighbors in TGNNs can be structured as sequence modeling. Beside this similarity, we also incorporate a series of algorithm designs including suffix infilling, temporal graph attention with self-loop, and causal masking self-attention to make TF-TGN work. During training, existing systems are slow in transforming the graph topology and conducting graph sampling. As such, we propose methods to parallelize the CSR format conversion and graph sampling. We also adapt Transformer codebase to train TF-TGN efficiently with multiple GPUs. We experiment with 9 graphs and compare with 2 state-of-the-art TGNN training frameworks. The results show that TF-TGN can accelerate training by over 2.20 while providing comparable or even superior accuracy to existing SOTA TGNNs. TF-TGN is available at https://github.com/qianghuangwhu/TF-TGN.

Ran Yan, Youhe Jiang, Xiaonan Nie et al.

Training large language models (LLMs) is a computationally intensive task, which is typically conducted in data centers with homogeneous high-performance GPUs. In this paper, we explore an alternative approach by deploying training computations across heterogeneous GPUs to enable better flexibility and efficiency for heterogeneous resource utilization. Toward this end, we propose a novel system, HexiScale, that can flexibly support asymmetric partition of training computations in the scope of data-, pipeline-, and tensor model parallelism. We further formalize the allocation of asymmetric partitioned training computations over a set of heterogeneous GPUs as a constrained optimization problem and propose an efficient hierarchical graph partitioning algorithm. Our approach effectively allocates training computations across heterogeneous GPUs, fully leveraging the available computational power. We compare the performance of HexiScale with state-of-the-art homogeneous and heterogeneous training systems. When training LLMs at different scales (from 7B to 30B), empirical results demonstrate that: (i) compared to state-of-the-art homogeneous baselines running over homogeneous GPUs, HexiScale achieves similar performance when running over heterogeneous GPUs with the same theoretical FLOPS; (ii) compared to state-of-the-art heterogeneous baselines running on the same heterogeneous clusters, HexiScale delivers $1.5\times$ to $2.4\times$ higher throughput.

Youhe Jiang, Fangcheng Fu, Taiyi Wang et al.

Serving Large Language Models (LLMs) can benefit immensely from parallelizing both the model and input requests across multiple devices, but incoming workloads exhibit substantial spatial and temporal heterogeneity. Spatially, workloads comprise heterogeneous requests with varying compute and memory demands. Temporally, workload composition varies over time. Nevertheless, existing systems typically assume spatially uniform and temporally stable workloads, employing a homogeneous, static model deployment. This mismatch between the assumption and real-world spatial-temporal heterogeneity results in suboptimal performance. We present OServe, an LLM serving system with heterogeneous and flexible model deployment that addresses both spatial and temporal heterogeneity. First, OServe introduces a novel workload-aware scheduling algorithm that optimizes heterogeneous model deployments according to real-time workload characteristics. Second, OServe proposes an efficient workload-adaptive switching method that migrates model deployments in response to predicted workload changes. Experiments on real-world traces show that OServe improves performance by up to 2$\times$ (average: 1.5$\times$) compared to state-of-the-art serving systems.

Yifan Dai, Zhenhua Wu, Bohan Zeng et al.

Joint audio-visual reasoning is essential for omnimodal understanding, yet current multimodal large language models (MLLMs) still struggle when reasoning requires fine-grained evidence from both modalities. A central limitation is that explicit text-based chain-of-thought (CoT) compresses continuous audio-visual signals into discrete tokens, weakening temporal grounding and shifting intermediate reasoning toward language priors. We argue that a unified latent space is a better medium for such reasoning because it preserves dense sensory information while remaining compatible with autoregressive generation. Based on this insight, we propose \textbf{LatentOmni}, a cross-modal reasoning framework that interleaves textual reasoning with audio-visual latent states. LatentOmni introduces feature-level supervision to align latent reasoning states with task-relevant sensory features and uses Omni-Sync Position Embedding (OSPE) to maintain temporal consistency between latent audio and visual states. We further construct \textbf{LatentOmni-Instruct-35K}, a dataset of audio-visual interleaved reasoning trajectories for supervising latent-space reasoning. Comprehensive evaluation across multiple audio-visual reasoning benchmarks demonstrates that LatentOmni achieves the best performance among the evaluated open-source models and consistently outperforms the Explicit Text CoT baseline, supporting latent-space joint reasoning as a promising path toward stronger omnimodal understanding.

Hailin Zhang, Xiaodong Ji, Yilin Chen et al.

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.

Xinyi Liu, Yujie Wang, Fangcheng Fu et al.

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.

Peichao Lai, Zhengfeng Zhang, Wentao Zhang et al. · pku

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.

Fangcheng Fu, Huanran Xue, Yong Cheng et al.

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. To be specific, we introduce BlindFL, a novel framework for VFL training and inference. First, to address the functionality of VFL models, we propose the federated source layers to unite the data from different parties. Various kinds of features can be supported efficiently by the federated source layers, including dense, sparse, numerical, and categorical features. Second, we carefully analyze the security during the federated execution and formalize the privacy requirements. Based on the analysis, we devise secure and accurate algorithm protocols, and further prove the security guarantees under the ideal-real simulation paradigm. Extensive experiments show that BlindFL supports diverse datasets and models efficiently whilst achieves robust privacy guarantees.

Fangcheng Fu, Xupeng Miao, Jiawei Jiang et al.

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 staleness, 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.

Yujie Wang, Youhe Jiang, Xupeng Miao et al.

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.

Xiaonan Nie, Yi Liu, Fangcheng Fu et al.

Recent years have witnessed the unprecedented achievements of large-scale pre-trained models, especially the 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 the fine-grained memory management via the Page abstraction and a unified scheduling method that coordinate the computations, data movements, and communications. Furthermore, Angel-PTM supports extreme model scaling with SSD storage and implements the lock-free updating mechanism to address the SSD I/O bandwidth 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 the strong scalability of Angel-PTM.

Li Zhang, Youhe Jiang, Guoliang He et al.

Mixed-precision inference techniques reduce the memory and computational demands of Large Language Models (LLMs) by applying hybrid precision formats to model weights, activations, and KV caches. However, existing systems struggle to (i) automatically generalize across diverse hardware architectures and precision formats, often requiring fragmented, hand-tuned kernels, and (ii) fully exploit available memory and compute resources, often causing performance bottlenecks. To address these problems, we propose TurboMind, a generalizable and efficient mixed-precision LLM inference engine of LMDeploy. TurboMind is built around two hardware-aware mixed-precision pipelines: A General Matrix Multiply (GEMM) pipeline that optimizes matrix operations through offline weight packing and online acceleration, and an attention pipeline that enables efficient attention computation with different Query, Key, and Value precision combinations. These pipelines are enabled by four key techniques: (i) Hardware-aware weight packing and (ii) adaptive head alignment for generalizability, and (iii) instruction-level parallelism and (iv) a KV memory loading pipeline for efficiency. We conduct comprehensive evaluations of LMDeploy powered by TurboMind across sixteen popular LLMs and four representative GPU architectures. Results demonstrate that LMDeploy achieves up to 61% lower serving latency (30% on average) and up to 156% higher throughput (58% on average) in mixed-precision workloads compared to existing mixed-precision frameworks, establishing consistent performance improvements across all tested configurations and hardware types. This work is open-sourced and publicly available at https://github.com/InternLM/lmdeploy.

Penghao Zhao, Hailin Zhang, Qinhan Yu et al.

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 technique 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 enhancements methods for RAG, facilitating effective engineering and implementation of RAG systems. Then from another view, we survey on 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. Github: https://github.com/PKU-DAIR/RAG-Survey.

Yuxiang Wang, Xiao Yan, Chuang Hu et al.

For graph self-supervised learning (GSSL), masked autoencoder (MAE) follows the generative paradigm and learns to reconstruct masked graph edges or node features. Contrastive Learning (CL) maximizes the similarity between augmented views of the same graph and is widely used for GSSL. However, MAE and CL are considered separately in existing works for GSSL. 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 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.

Yujie Wang, Shenhan Zhu, Fangcheng Fu et al.

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.

Pinxue Zhao, Hailin Zhang, Fangcheng Fu et al.

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%.

Hao Liang, Zhengyang Zhao, Meiyi Qiang et al.

Data-centric training has emerged as a promising direction for improving large language models (LLMs) by optimizing not only model parameters but also the selection, composition, and weighting of training data during optimization. However, existing approaches to data selection, data mixture optimization, and data reweighting are often developed in isolated codebases with inconsistent interfaces, hindering reproducibility, fair comparison, and practical integration. In this paper, we present DataFlex, a unified data-centric dynamic training framework built upon LLaMA-Factory. DataFlex supports three major paradigms of dynamic data optimization: sample selection, domain mixture adjustment, and sample reweighting, while remaining fully compatible with the original training workflow. It provides extensible trainer abstractions and modular components, enabling a drop-in replacement for standard LLM training, and unifies key model-dependent operations such as embedding extraction, inference, and gradient computation, with support for large-scale settings including DeepSpeed ZeRO-3. We conduct comprehensive experiments across multiple data-centric methods. Dynamic data selection consistently outperforms static full-data training on MMLU across both Mistral-7B and Llama-3.2-3B. For data mixture, DoReMi and ODM improve both MMLU accuracy and corpus-level perplexity over default proportions when pretraining Qwen2.5-1.5B on SlimPajama at 6B and 30B token scales. DataFlex also achieves consistent runtime improvements over original implementations. These results demonstrate that DataFlex provides an effective, efficient, and reproducible infrastructure for data-centric dynamic training of LLMs.

Xinyi Liu, Yujie Wang, Shenhan Zhu et al.

Galvatron is a distributed system for efficiently training large-scale Foundation Models. It overcomes the complexities of selecting optimal parallelism strategies by automatically identifying the most efficient hybrid strategy, incorporating data, tensor, pipeline, sharded data, and sequence parallelism, along with recomputation. The system's architecture includes a profiler for hardware and model analysis, a search engine for strategy optimization using decision trees and dynamic programming, and a runtime for executing these strategies efficiently. Benchmarking on various clusters demonstrates Galvatron's superior throughput compared to existing frameworks. This open-source system offers user-friendly interfaces and comprehensive documentation, making complex distributed training accessible and efficient. The source code of Galvatron is available at https://github.com/PKU-DAIR/Hetu-Galvatron.

Haoyu Zheng, Fangcheng Fu, Jia Wu et al.

LLM-based workflows compose specialized agents to execute complex tasks, and these agents usually share substantial context, allowing KV-Cache reuse to save computation. Existing approaches either manage KV-Cache at agent level and fail to exploit the reuse opportunities within workflows, or manage cache at the workflow level but assume that each workflow calls a static sequence of agents. However, practical workflows are typically dynamic, where the sequence of invoked agents and thus induced cache reuse opportunities depend on the context of each task. To serve such dynamic workflows efficiently, we build a system dubbed PBKV (\textbf{P}rediction-\textbf{B}ased \textbf{KV}-Cache Management). For each workflow, PBKV predicts the agent invocations in several future steps by fusing the guidance from historical workflows and context of the target workflow. Based on the predictions, PBKV estimates the reuse potential of cache entries and keeps the high-potential entries in GPU memory. To be robust to prediction errors, PBKV utilizes the predictions conservatively during both cache eviction and prefetching. Experiments on three workflow benchmarks show that PBKV achieves up to $1.85\times$ speedup over LRU on dynamic workflows, and up to $1.26\times$ speedup over the SOTA baseline KVFlow on the static workflow.

Haoyu Zheng, Yongqiang Zhang, Fangcheng Fu et al.

To schedule LLM inference, the \textit{shortest job first} (SJF) principle is favorable by prioritizing requests with short output lengths to avoid head-of-line (HOL) blocking. Existing methods usually predict a single output length for each request to facilitate scheduling. We argue that such a \textit{point estimate} does not match the \textit{stochastic} decoding process of LLM inference, where output length is \textit{uncertain} by nature and determined by when the end-of-sequence (EOS) token is sampled. Hence, the output length of each request should be fitted with a distribution rather than a single value. With an in-depth analysis of empirical data and the stochastic decoding process, we observe that output length follows a heavy-tailed distribution and can be fitted with the log-t distribution. On this basis, we propose a simple metric called Tail Inflated Expectation (TIE) to replace the output length in SJF scheduling, which adjusts the expectation of a log-t distribution with its tail probabilities to account for the risk that a request generates long outputs. To evaluate our TIE scheduler, we compare it with three strong baselines, and the results show that TIE reduces the per-token latency by $2.31\times$ for online inference and improves throughput by $1.42\times$ for offline data generation.

Yifei Xia, Suhan Ling, Fangcheng Fu et al.

Generating high-fidelity long videos with Diffusion Transformers (DiTs) is often hindered by significant latency, primarily due to the computational demands of attention mechanisms. For instance, generating an 8-second 720p video (110K tokens) with HunyuanVideo takes about 600 PFLOPs, with around 500 PFLOPs consumed by attention computations. To address this issue, we propose AdaSpa, the first Dynamic Pattern and Online Precise Search sparse attention method. Firstly, to realize the Dynamic Pattern, we introduce a blockified pattern to efficiently capture the hierarchical sparsity inherent in DiTs. This is based on our observation that sparse characteristics of DiTs exhibit hierarchical and blockified structures between and within different modalities. This blockified approach significantly reduces the complexity of attention computation while maintaining high fidelity in the generated videos. Secondly, to enable Online Precise Search, we propose the Fused LSE-Cached Search with Head-adaptive Hierarchical Block Sparse Attention. This method is motivated by our finding that DiTs' sparse pattern and LSE vary w.r.t. inputs, layers, and heads, but remain invariant across denoising steps. By leveraging this invariance across denoising steps, it adapts to the dynamic nature of DiTs and allows for precise, real-time identification of sparse indices with minimal overhead. AdaSpa is implemented as an adaptive, plug-and-play solution and can be integrated seamlessly with existing DiTs, requiring neither additional fine-tuning nor a dataset-dependent profiling. Extensive experiments validate that AdaSpa delivers substantial acceleration across various models while preserving video quality, establishing itself as a robust and scalable approach to efficient video generation.

Yujie Wang, Shiju Wang, Shenhan Zhu et al.

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. In this paper, we show that employing a single, static strategy results in inefficiency and resource under-utilization, highlighting the need for adaptive approaches to handle the heterogeneous workloads across sequences. To address this, we propose a heterogeneity-adaptive sequence parallelism method. For each training step, our approach captures the variability in sequence lengths and assigns the optimal combination of scattering strategies based on workload characteristics. We model this problem as a linear programming optimization and design an efficient and effective solver to find the optimal solution. Furthermore, we implement our method in a high-performance system that supports adaptive parallelization in distributed LLM training. Experimental results demonstrate that our system outperforms state-of-the-art training frameworks by up to 1.98x.

Hao Ge, Junda Feng, Qi Huang et al.

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. In this work, we introduce ByteScale, an efficient, flexible, and scalable LLM training framework for large-scale mixed training of long and short sequences. The core of ByteScale is a novel parallelism strategy, namely Hybrid Data Parallelism (HDP), which unifies the inter- and intra-data partitioning with a dynamic mesh design. In particular, we build a communication optimizer, which eliminates the redundant communication for short sequences by data-aware sharding and dynamic communication, and further compresses the communication cost for long sequences by selective offloading. Besides, we also develop a balance scheduler to mitigate the imbalanced computation by parallelism-aware data assignment. We evaluate ByteScale with the model sizes ranging from 7B to 141B, context lengths from 256K to 2048K, on a production cluster with more than 12,000 GPUs. Experiment results show that ByteScale outperforms the state-of-the-art training system by up to 7.89x.

Youhe Jiang, Ran Yan, You Peng et al.

Modern Large Language Model (LLM) serving operates in highly volatile environments characterized by severe runtime dynamics, such as workload fluctuations and elastic cluster autoscaling. Traditional serving systems rely on static, human-engineered serving policies (e.g., scheduling algorithms and rescheduling strategies) to manage these dynamics. However, these policies must navigate deeply intertwined runtime trade-offs (e.g., scheduling overhead vs. execution efficiency, rescheduling frequency vs. reconfiguration overhead), whose optimal balance is workload-specific and shifts continuously as runtime conditions evolve, rendering any fixed policy fundamentally unable to adapt. We propose Autopoiesis, a novel online self-evolving system that shifts LLM serving from static policy deployment to continuous online policy evolution. First, Autopoiesis introduces an LLM-driven program synthesis workflow to evolve serving policies with respect to real-time observed dynamics, where the evolved policies reflect the optimal decision in navigating the complex, multi-dimensional trade-off space. Second, Autopoiesis enables this synthesis process to operate continuously during serving, observing real-world system behavior, and rewriting the policy code as runtime trade-offs shift, thereby transforming policy design from a one-time offline endeavor into an ongoing system component, enabling autonomous adaptation to evolving runtime conditions. Together, we establish a new paradigm: Serving policies are no longer static artifacts designed by humans before deployment, but living code that LLMs continuously evolve throughout deployment to navigate runtime trade-offs beyond human design. We evaluate Autopoiesis across diverse runtime dynamics and show up to 53% and on average 34% improvements over state-of-the-art LLM serving systems.

Yuhang Wang, Youhe Jiang, Bin Cui et al.

Recent advances in test-time scaling suggest that Large Language Models (LLMs) can gain better capabilities by generating Chain-of-Thought reasoning (analogous to human thinking) to respond a given request, and meanwhile exploring more reasoning branches (i.e., generating multiple responses and ensembling them) can improve the final output quality. However, when incorporating the two scaling dimensions, we find that the system efficiency is dampened significantly for two reasons. Firstly, the time cost to generate the final output increases substantially as many reasoning branches would be trapped in the over-thinking dilemma, producing excessively long responses. Secondly, generating multiple reasoning branches for each request increases memory consumption, which is unsuitable for LLM serving since we can only batch a limited number of requests to process simultaneously. To address this, we present SART, a serving framework for efficient and accurate LLM reasoning. The essential idea is to manage the thinking to be short and right, rather than long. For one thing, we devise a redundant sampling with early stopping approach based on empirical observations and theoretic analysis, which increases the likelihood of obtaining short-thinking responses when sampling reasoning branches. For another, we propose to dynamically prune low-quality branches so that only right-thinking branches are maintained, reducing the memory consumption and allowing us to batch more requests. Experimental results demonstrate that SART not only improves the accuracy of LLM reasoning but also enhances the serving efficiency, outperforming existing methods by up to 28.2 times and on average 15.7 times in terms of efficiency when achieving the same level of accuracy.

Haoyang Li, Fangcheng Fu, Sheng Lin et al.

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 times.

Qiming Zeng, Xiao Yan, Hao Luo et al.

By retrieving contexts from knowledge graphs, graph-based retrieval-augmented generation (GraphRAG) enhances large language models (LLMs) to generate quality answers for user questions. Many GraphRAG methods have been proposed and reported inspiring performance in answer quality. However, we observe that the current answer evaluation framework for GraphRAG has two critical flaws, i.e., unrelated questions and evaluation biases, which may lead to biased or even wrong conclusions on performance. To tackle the two flaws, we propose an unbiased evaluation framework that uses graph-text-grounded question generation to produce questions that are more related to the underlying dataset and an unbiased evaluation procedure to eliminate the biases in LLM-based answer assessment. We apply our unbiased framework to evaluate 3 representative GraphRAG methods and find that their performance gains are much more moderate than reported previously. Although our evaluation framework may still have flaws, it calls for scientific evaluations to lay solid foundations for GraphRAG research.

Shiju Wang, Yujie Wang, Ao Sun et al.

Long context training is crucial for LLM's context extension. Existing schemes, such as sequence parallelism, incur substantial communication overhead. Pipeline parallelism (PP) reduces this cost, but its effectiveness hinges on partitioning granularity. Batch-level PP dividing input samples exhibits high memory consumption in long-context scenario, whereas token-level PP splitting sequences into slices alleviates memory overhead but may incur hardware under-utilization. This trade-off motivates adaptively selecting PP granularity to match resource and workload characteristics. Moreover, sequence length distribution of the real-world dataset exhibits skewness, posing a challenge on PP's workload balance and efficient scheduling. Current static PP scheduling methods overlook the variance of sequence length, leading to suboptimal performance. In this paper, we propose Elastic Pipeline Parallelism (EPP) that orchestrates token-level PP and batch-level PP to adapt to resource and workload heterogeneity. We build InfiniPipe, a distributed training system that unleashes the potential of EPP via (1) a resource-aware and workload-balanced sequence processor that splits long sequences and packs short ones; and (2) a co-optimization methodology that jointly optimizes pipeline schedule and gradient checkpointing via a mechanism named stage-aware chunk-level adaptive checkpointing. Comprehensive experiments demonstrate that InfiniPipe achieves a 1.69x speedup over state-of-the-art systems.

Xiaodong Ji, Hailin Zhang, Fangcheng Fu et al.

Many advanced Large Language Model (LLM) applications require long-context processing, but the self-attention module becomes a bottleneck during the prefilling stage of inference due to its quadratic time complexity with respect to sequence length. Existing sparse attention methods accelerate attention computation by skipping less significant regions of the attention map. However, these approaches typically perform coarse-grained inspection of the attention map, rendering considerable loss in model accuracy. In this paper, we propose SALE, a fine-grained sparse attention method that accelerates the long-context prefilling stage of LLM with negligible loss in model accuracy. SALE achieves fast and accurate fine-grained attention weight estimation through 4-bit quantized query-key products, followed by block-sparse attention to accelerate prefilling computations. For importance evaluation for query-key pairs, we adopt our Relative Attention Score metric, which offers significantly higher efficiency within our framework. We implement a custom CUDA kernel optimized for our approach for hardware efficiency, reducing the additional overhead to approximately 11% of the full attention latency. Notably, SALE requires no parameter training and can be seamlessly integrated into existing systems with trivial code modifications. Experiments on long-context benchmarks demonstrate that our method outperforms existing approaches in accuracy-efficiency trade-offs, achieving at least 3.36x speedups on Llama-3.1-8B for sequences longer than 64K while maintaining model quality.

Zihao Yu, Haoyang Li, Fangcheng Fu et al.

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 automatically identify the affected image regions and utilize the cached unchanged regions' feature map to accelerate the inference process. 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.

Shicheng Gao, Jie Xu, Xiaosen Li et al.

K-core decomposition is a commonly used metric to analyze graph structure or study the relative importance of nodes in complex graphs. Recent years have seen rapid growth in the scale of the graph, especially in industrial settings. For example, our industrial partner runs popular social applications with billions of users and is able to gather a rich set of user data. As a result, applying K-core decomposition on large graphs has attracted more and more attention from academics and the industry. A simple but effective method to deal with large graphs is to train them in the distributed settings, and some distributed K-core decomposition algorithms are also proposed. Despite their effectiveness, we experimentally and theoretically observe that these algorithms consume too many resources and become unstable on super-large-scale graphs, especially when the given resources are limited. In this paper, we deal with those super-large-scale graphs and propose a divide-and-conquer strategy on top of the distributed K-core decomposition algorithm. We evaluate our approach on three large graphs. The experimental results show that the consumption of resources can be significantly reduced, and the calculation on large-scale graphs becomes more stable than the existing methods. For example, the distributed K-core decomposition algorithm can scale to a large graph with 136 billion edges without losing correctness with our divide-and-conquer technique.

Fangcheng Fu, Jiawei Jiang, Yingxia Shao et al.

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.