Wencong Xiao

Huaizheng Zhang, Yuanming Li, Wencong Xiao et al. · berkeley

New architecture GPUs like A100 are now equipped with multi-instance GPU (MIG) technology, which allows the GPU to be partitioned into multiple small, isolated instances. This technology provides more flexibility for users to support both deep learning training and inference workloads, but efficiently utilizing it can still be challenging. The vision of this paper is to provide a more comprehensive and practical benchmark study for MIG in order to eliminate the need for tedious manual benchmarking and tuning efforts. To achieve this vision, the paper presents MIGPerf, an open-source tool that streamlines the benchmark study for MIG. Using MIGPerf, the authors conduct a series of experiments, including deep learning training and inference characterization on MIG, GPU sharing characterization, and framework compatibility with MIG. The results of these experiments provide new insights and guidance for users to effectively employ MIG, and lay the foundation for further research on the orchestration of hybrid training and inference workloads on MIGs. The code and results are released on https://github.com/MLSysOps/MIGProfiler. This work is still in progress and more results will be published soon.

Tengju Ye, Wei Jing, Chunyong Hu et al.

Building a multi-modality multi-task neural network toward accurate and robust performance is a de-facto standard in perception task of autonomous driving. However, leveraging such data from multiple sensors to jointly optimize the prediction and planning tasks remains largely unexplored. In this paper, we present FusionAD, to the best of our knowledge, the first unified framework that fuse the information from two most critical sensors, camera and LiDAR, goes beyond perception task. Concretely, we first build a transformer based multi-modality fusion network to effectively produce fusion based features. In constrast to camera-based end-to-end method UniAD, we then establish a fusion aided modality-aware prediction and status-aware planning modules, dubbed FMSPnP that take advantages of multi-modality features. We conduct extensive experiments on commonly used benchmark nuScenes dataset, our FusionAD achieves state-of-the-art performance and surpassing baselines on average 15% on perception tasks like detection and tracking, 10% on occupancy prediction accuracy, reducing prediction error from 0.708 to 0.389 in ADE score and reduces the collision rate from 0.31% to only 0.12%.

Chenhao Ye, Huaizheng Zhang, Mingcong Han et al.

Modern LLM reinforcement learning (RL) workloads require a highly efficient weight transfer system to scale training across heterogeneous computational resources. However, existing weight transfer approaches either fail to provide flexibility for dynamically scaling clusters or incur fundamental data movement overhead, resulting in poor performance. We introduce Reference-Oriented Storage (ROS), a new storage abstraction for RL weight transfer that exploits the highly replicated model weights in place. ROS presents the illusion that certain versions of the model weights are stored and can be fetched on demand. Underneath, ROS does not physically store any copies of the weights; instead, it tracks the workers that hold these weights on GPUs for inference. Upon request, ROS directly uses them to serve reads. We build TensorHub, a production-quality system that extends the ROS idea with topology-optimized transfer, strong consistency, and fault tolerance. Evaluation shows that TensorHub fully saturates RDMA bandwidth and adapts to three distinct rollout workloads with minimal engineering effort. Specifically, TensorHub reduces total GPU stall time by up to 6.7x for standalone rollouts, accelerates weight update for elastic rollout by 4.8x, and cuts cross-datacenter rollout stall time by 19x. TensorHub has been deployed in production to support cutting-edge RL training.

Biao Sun, Ziming Huang, Hanyu Zhao et al.

Inference serving for large language models (LLMs) is the key to unleashing their potential in people's daily lives. However, efficient LLM serving remains challenging today because the requests are inherently heterogeneous and unpredictable in terms of resource and latency requirements, as a result of the diverse applications and the dynamic execution nature of LLMs. Existing systems are fundamentally limited in handling these characteristics and cause problems such as severe queuing delays, poor tail latencies, and SLO violations. We introduce Llumnix, an LLM serving system that reacts to such heterogeneous and unpredictable requests by runtime rescheduling across multiple model instances. Similar to context switching across CPU cores in modern operating systems, Llumnix reschedules requests to improve load balancing and isolation, mitigate resource fragmentation, and differentiate request priorities and SLOs. Llumnix implements the rescheduling with an efficient and scalable live migration mechanism for requests and their in-memory states, and exploits it in a dynamic scheduling policy that unifies the multiple rescheduling scenarios elegantly. Our evaluations show that Llumnix improves tail latencies by an order of magnitude, accelerates high-priority requests by up to 1.5x, and delivers up to 36% cost savings while achieving similar tail latencies, compared against state-of-the-art LLM serving systems. Llumnix is publicly available at https://github.com/AlibabaPAI/llumnix.

Shaoyuan Chen, Wencong Xiao, Yutong Lin et al.

Transformer-based large language models (LLMs) exhibit impressive performance in generative tasks but also introduce significant challenges in real-world serving due to inefficient use of the expensive, computation-optimized accelerators. Although disaggregated serving architectures have been proposed to split different phases of LLM inference, the efficiency of decoding phase is still low. This is caused by the varying resource demands of different operators in the transformer-based LLMs. Specifically, the attention operator is memory-intensive, exhibiting a memory access pattern that clashes with the strengths of modern accelerators, especially for long context requests. To enhance the efficiency of LLM decoding, we introduce model-attention disaggregation. This approach leverages a collection of cheap, memory-optimized devices for the attention operator while still utilizing high-end accelerators for other parts of the model. This heterogeneous setup ensures that each component is tailored to its specific workload, maximizing overall performance and cost efficiency. Our comprehensive analysis and experiments confirm the viability of splitting the attention computation over multiple devices. Also, the communication bandwidth required between heterogeneous devices proves to be manageable with prevalent networking technologies. To further validate our theory, we develop and deploy Lamina, an LLM inference system that incorporates model-attention disaggregation in a distributed heterogeneous cluster. Experimental results indicate that Lamina can provide 16.1 ~ 90.1% higher estimated throughput than existing solutions with similar costs.

Borui Wan, Gaohong Liu, Zuquan Song et al.

The training scale of large language models (LLMs) has reached tens of thousands of GPUs and is still continuously expanding, enabling faster learning of larger models. Accompanying the expansion of the resource scale is the prevalence of failures (CUDA error, NaN values, job hang, etc.), which poses significant challenges to training stability. Any large-scale LLM training infrastructure should strive for minimal training interruption, efficient fault diagnosis, and effective failure tolerance to enable highly efficient continuous training. This paper presents ByteRobust, a large-scale GPU infrastructure management system tailored for robust and stable training of LLMs. It exploits the uniqueness of LLM training process and gives top priorities to detecting and recovering failures in a routine manner. Leveraging parallelisms and characteristics of LLM training, ByteRobust enables high-capacity fault tolerance, prompt fault demarcation, and localization with an effective data-driven approach, comprehensively ensuring continuous and efficient training of LLM tasks. ByteRobust is deployed on a production GPU platform and achieves 97% ETTR for a three-month training job on 9,600 GPUs.

Yangtao Deng, Lei Zhang, Qinlong Wang et al.

Reliability is essential for ensuring efficiency in LLM training. However, many real-world reliability issues remain difficult to resolve, resulting in wasted resources and degraded model performance. Unfortunately, today's collective communication libraries operate as black boxes, hiding critical information needed for effective root cause analysis. We propose Mycroft, a lightweight distributed tracing and root cause analysis system designed to address previously hidden reliability issues in collective communication. Mycroft's key idea is to trace collective communication states and leverage internal control and data dependencies to resolve reliability problems in LLM training. Mycroft has been deployed at ByteDance for over six months to debug collective communication related issues at runtime. It detected anomalies within 15 seconds in 90% of cases and identified the root cause within 20 seconds in 60% of cases. We also conducted extensive fault injection experiments to demonstrate Mycroft's capability and efficiency.

Rongxin Cheng, Kai Zhou, Xingda Wei et al.

Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. SpecActor achieves fast rollout with speculative decoding that deploys a fast path (e.g., a smaller model) to accelerate the unparallelizable generation, while the correctness is guaranteed by fast parallel verification of the outputs with the original model. SpecActor addresses two foundational challenges in speculative rollout by (1) a \emph{dynamic decoupled speculation} execution method that maximizes the GPU computational efficiency to realize speedup for large-batch execution -- a configuration common in training but unfriendly to speculative execution and (2) a \emph{dynamic Best-of-N speculation} method that selects and combines different drafting methods according to the rollout progress. It substantially improves the speculation accuracy even when the best drafting method is unknown a priori, meanwhile without requiring adding extra computation resources. {\sys} is {1.3--1.7}\,$\times$ faster than common post-training baselines, and is {1.3--1.5}\,$\times$ faster compared to naively adopting speculative decoding for rollout.

Rui Li, Xiaoyun Zhi, Jinxin Chi et al.

Large Language Models (LLMs) have become a cornerstone of modern AI, driving breakthroughs in natural language processing and expanding into multimodal jobs involving images, audio, and video. As with most computational software, it is important to distinguish between ordinary runtime performance and startup overhead. Prior research has focused on runtime performance: improving training efficiency and stability. This work focuses instead on the increasingly critical issue of startup overhead in training: the delay before training jobs begin execution. Startup overhead is particularly important in large, industrial-scale LLMs, where failures occur more frequently and multiple teams operate in iterative update-debug cycles. In one of our training clusters, more than 3.5% of GPU time is wasted due to startup overhead alone. In this work, we present the first in-depth characterization of LLM training startup overhead based on real production data. We analyze the components of startup cost, quantify its direct impact, and examine how it scales with job size. These insights motivate the design of Bootseer, a system-level optimization framework that addresses three primary startup bottlenecks: (a) container image loading, (b) runtime dependency installation, and (c) model checkpoint resumption. To mitigate these bottlenecks, Bootseer introduces three techniques: (a) hot block record-and-prefetch, (b) dependency snapshotting, and (c) striped HDFS-FUSE. Bootseer has been deployed in a production environment and evaluated on real LLM training workloads, demonstrating a 50% reduction in startup overhead.

Jianbo Dong, Bin Luo, Jun Zhang et al.

The emergence of Large Language Models (LLMs) has necessitated the adoption of distributed training techniques, involving the deployment of thousands of GPUs to train a single model. Unfortunately, the efficiency of large-scale distributed training systems is often suboptimal due to the increased likelihood of hardware errors in high-end GPU products and the heightened risk of network traffic collisions. Moreover, any local hardware failure can disrupt training tasks, and the inability to swiftly identify faulty components leads to a significant waste of GPU resources. And, prolonged communication due to traffic collisions can substantially increase GPU waiting times. To address these challenges, we propose a communication-driven solution, namely the C4. The key insights of C4 are twofold. First, the load in distributed training exhibits homogeneous characteristics and is divided into iterations through periodic synchronization, therefore hardware anomalies would incur certain syndrome in collective communication. By leveraging this feature, C4 can rapidly identify the faulty components, swiftly isolate the anomaly, and restart the task, thereby avoiding resource wastage caused by delays in anomaly detection. Second, the predictable communication model of collective communication, involving a limited number of long-lived flows, allows C4 to efficiently execute traffic planning, substantially reducing bandwidth competition among these flows. The C4 has been extensively deployed across real-world production systems in a hyperscale cloud provider, yielding a significant improvement in system efficiency, from 30% to 45%. This enhancement is attributed to a 30% reduction in error-induced overhead and a 15% reduction in communication costs.

Zhuliang Yao, Shijie Cao, Wencong Xiao et al.

In trained deep neural networks, unstructured pruning can reduce redundant weights to lower storage cost. However, it requires the customization of hardwares to speed up practical inference. Another trend accelerates sparse model inference on general-purpose hardwares by adopting coarse-grained sparsity to prune or regularize consecutive weights for efficient computation. But this method often sacrifices model accuracy. In this paper, we propose a novel fine-grained sparsity approach, balanced sparsity, to achieve high model accuracy with commercial hardwares efficiently. Our approach adapts to high parallelism property of GPU, showing incredible potential for sparsity in the widely deployment of deep learning services. Experiment results show that balanced sparsity achieves up to 3.1x practical speedup for model inference on GPU, while retains the same high model accuracy as fine-grained sparsity.