Recent advances in agentic large language models (LLMs) have substantially improved Text-to-SQL, enabling users without database expertise to query databases intuitively. However, deploying agentic LLM-based Text-to-SQL systems in production remains challenging due to multi-stage dependencies, strict latency requirements, and deployment complexity across heterogeneous GPUs in enterprise clusters. Existing LLM serving frameworks are designed mainly for independent inference tasks, leading to suboptimal performance and frequent service-level objective (SLO) violations for Text-to-SQL workloads. In this paper, we introduce \sys, a framework for scheduling and executing agentic multi-stage LLM-based Text-to-SQL workflows on heterogeneous GPU clusters serving multi-tenant requests. \sys adopts a hierarchical scheduler that combines global workload-balanced task dispatching with an adaptive local priority queue, guided by a systematic analysis of agentic Text-to-SQL workflows. We also propose a lightweight simulation-based method to tune key scheduling hyperparameters, improving robustness and adaptability. Evaluations on realistic Text-to-SQL benchmarks show that \sys significantly outperforms state-of-the-art LLM serving frameworks. Across all traces, \sys reduces P95 tail latency by $1.42{\sim}1.56\times$ and increases throughput by $1.49{\sim}1.81\times$, demonstrating consistent gains under diverse workloads.

Efficient data selection is crucial to accelerate the pretraining of language model (LMs). While various methods have been proposed to enhance data efficiency, limited research has addressed the inherent conflicts between these approaches to achieve optimal data selection for LM pretraining. To tackle this problem, we propose a multi-actor collaborative data selection mechanism: each data selection method independently prioritizes data based on its criterion and updates its prioritization rules using the current state of the model, functioning as an independent actor for data selection; and a console is designed to adjust the impacts of different actors at various stages and dynamically integrate information from all actors throughout the LM pretraining process. We conduct extensive empirical studies to evaluate our multi-actor framework. The experimental results demonstrate that our approach significantly improves data efficiency, accelerates convergence in LM pretraining, and achieves an average relative performance gain up to 10.5% across multiple language model benchmarks compared to the state-of-the-art methods. Code and checkpoints are publicly released at https://github.com/Relaxed-System-Lab/multi-actor-data-selection.

Many modern machine learning (ML) methods rely on embedding models to learn vector representations (embeddings) for a set of entities (embedding tables). As increasingly diverse ML applications utilize embedding models and embedding tables continue to grow in size and number, there has been a surge in the ad-hoc development of specialized frameworks targeted to train large embedding models for specific tasks. Although the scalability issues that arise in different embedding model training tasks are similar, each of these frameworks independently reinvents and customizes storage components for specific tasks, leading to substantial duplicated engineering efforts in both development and deployment. This paper presents MLKV, an efficient, extensible, and reusable data storage framework designed to address the scalability challenges in embedding model training, specifically data stall and staleness. MLKV augments disk-based key-value storage by democratizing optimizations that were previously exclusive to individual specialized frameworks and provides easy-to-use interfaces for embedding model training tasks. Extensive experiments on open-source workloads, as well as applications in eBay's payment transaction risk detection and seller payment risk detection, show that MLKV outperforms offloading strategies built on top of industrial-strength key-value stores by 1.6-12.6 ×. MLKV is open-source at https://github.com/llm-db/MLKV.

Nowadays, AI researchers become more and more interested in fine-tuning a pre-trained LLM, whose size has grown to up to over 100B parameters, for their downstream tasks. One approach to fine-tune such huge models is to aggregate device memory from many GPUs. However, this approach introduces prohibitive costs for most data scientists with a limited budget for high-end GPU servers. In this paper, we focus on LLM fine-tuning on a single consumer-grade GPU in a commodity server with limited main memory capacity, which is accessible to most AI researchers. In such a scenario, existing offloading-based methods fail to fine-tune an LLM efficiently due to a lack of holistic intra-server tensor movement management. To this end, we present Ratel, a low-cost, high-performance deep learning training framework that enables efficient 100B-scale model fine-tuning on a commodity server with a consumergrade GPU and limited main memory capacity. The key idea is to add holistic offloading traffic as an optimization dimension for 1) active gradient offloading, and 2) holistic traffic-aware activation swapping mechanism. The experimental results show that 1) Ratel is the first to fine-tune a 175B model on an RTX 4090 and 256 GB main memory, 2) Ratel achieves 2.32x throughput than the state-of-the-art baselines when fine-tuning a small 13B model, and 3) Ratel enables a cheap low-end consumer GPU to have higher cost-effectiveness than a DGX-A100 cluster when fine-tuning a 175B model.

Disaggregating the prefill and decoding phases represents an effective new paradigm for generative inference of large language models (LLM). This approach offers some significant system advantages, such as eliminating prefill-decoding interference and optimizing resource allocation. However, it is still an challenging open problem about how to deploy the disaggregated inference paradigm across a group of heterogeneous GPUs, which can be an economic alternative of the deployment over the homogeneous high performance GPUs. Towards this end, we introduce HexGen-2, a distributed system for high throughput and cost-efficient LLM serving on heterogeneous GPUs following the disaggragated paradigm. Built on top of HexGen, the core component of HexGen-2 is a sophisticated scheduling algorithm that formalizes the allocation of disaggregated LLM inference computations and communications over heterogeneous GPUs and network connections as a constraint optimization problem. We leverage the graph partitioning and max-flow algorithm to co-optimize resource allocation, parallel strategies for distinct inference phases, and the efficiency of inter-phase key-value (KV) cache communications. We conduct extensive experiments to evaluate HexGen-2, i.e., on OPT (30B) and Llama-2 (70B) models in various real-world settings, the results reveal that HexGen-2 delivers up to a 2.0$\times$ and on average a 1.3$\times$ improvement in serving throughput, reduces the average inference latency by 1.5$\times$ compared with state-of-the-art systems given the same price budget, and achieves comparable inference performance with a 30% lower price budget.

Large language models (LLMs) are increasingly employed for complex tasks that process multiple generation calls in a tree structure with shared prefixes of tokens, including few-shot prompting, multi-step reasoning, speculative decoding, etc. However, existing inference systems for tree-based applications are inefficient due to improper partitioning of queries and KV cache during attention calculation.This leads to two main issues: (1) a lack of memory access (IO) reuse for KV cache of shared prefixes, and (2) poor load balancing.As a result, there is redundant KV cache IO between GPU global memory and shared memory, along with low GPU utilization. To address these challenges, we propose DeFT(Decoding with Flash Tree-Attention), a hardware-efficient attention algorithm with prefix-aware and load-balanced KV cache partitions. DeFT reduces the number of read/write operations of KV cache during attention calculation through **KV-Guided Grouping**, a method that avoids repeatedly loading KV cache of shared prefixes in attention computation. Additionally, we propose **Flattened Tree KV Splitting**, a mechanism that ensures even distribution of the KV cache across partitions with little computation redundancy, enhancing GPU utilization during attention computations. By reducing 73-99% KV cache IO and nearly 100% IO for partial results during attention calculation, DeFT achieves up to 2.23/3.59$\times$ speedup in the end-to-end/attention latency across three practical tree-based workloads compared to state-of-the-art attention algorithms. Our code is available at https://github.com/LINs-lab/DeFT.

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.

Multimodal large language models (MLLMs) have achieved strong performance on vision-language tasks but still struggle with fine-grained visual differences, leading to hallucinations or missed semantic shifts. We attribute this to limitations in both training data and learning objectives. To address these issues, we propose a controlled data generation pipeline that produces minimally edited image pairs with semantically aligned captions. Using this pipeline, we construct the Micro Edit Dataset (MED), containing over 50K image-text pairs spanning 11 fine-grained edit categories, including attribute, count, position, and object presence changes. Building on MED, we introduce a supervised fine-tuning (SFT) framework with a feature-level consistency loss that promotes stable visual embeddings under small edits. We evaluate our approach on the Micro Edit Detection benchmark, which includes carefully balanced evaluation pairs designed to test sensitivity to subtle visual variations across the same edit categories. Our method improves difference detection accuracy and reduces hallucinations compared to strong baselines, including GPT-4o. Moreover, it yields consistent gains on standard vision-language tasks such as image captioning and visual question answering. These results demonstrate the effectiveness of combining targeted data and alignment objectives for enhancing fine-grained visual reasoning in MLLMs. Code and datasets are publicly released at https://github.com/Relaxed-System-Lab/hallu_med.

Multi-modal large language models (MLLMs) have achieved remarkable capabilities by integrating visual perception with language understanding, enabling applications such as image-grounded dialogue, visual question answering, and scientific analysis. However, most MLLMs adopt a static inference paradigm, encoding the entire image into fixed visual tokens upfront, which limits their ability to iteratively refine understanding or adapt to context during inference. This contrasts sharply with human perception, which is dynamic, selective, and feedback-driven. In this work, we introduce a novel framework for inference-time visual token scaling that enables MLLMs to perform iterative, verifier-guided reasoning over visual content. We formulate the problem as a Markov Decision Process, involving a reasoner that proposes visual actions and a verifier—trained via multi-step Direct Preference Optimization (DPO)—that evaluates these actions and determines when reasoning should terminate. To support this, we present a new dataset, VTS, comprising supervised reasoning trajectories (VTS-SFT) and preference-labeled reasoning comparisons (VTS-DPO). Our method significantly outperforms existing approaches across diverse visual reasoning benchmarks, offering not only improved accuracy but also more interpretable and grounded reasoning processes. These results demonstrate the promise of dynamic inference mechanisms for enabling fine-grained, context-aware visual reasoning in next-generation MLLMs. Code and datasets are publicly released at https://vts-v.github.io/.

Reinforcement learning (RL) has become a trending paradigm for training large language models (LLMs), particularly for reasoning tasks. Effective RL for LLMs requires massive parallelization and poses an urgent need for efficient training systems. Most existing large-scale RL systems for LLMs are synchronous by alternating generation and training in a batch setting, where the rollouts in each training batch are generated by the same (or latest) model. This stabilizes RL training but suffers from severe system-level inefficiency. Generation must wait until the longest output in the batch is completed before model update, resulting in GPU underutilization. We present AReaL, a fully asynchronous RL system that completely decouples generation from training. Rollout workers in AReaL continuously generate new outputs without waiting, while training workers update the model whenever a batch of data is collected. AReaL also incorporates a collection of system-level optimizations, leading to substantially higher GPU utilization. To stabilize RL training, AReaL balances the workload of rollout and training workers to control data staleness, and adopts a staleness-enhanced PPO variant to better handle outdated training samples. Extensive experiments on math and code reasoning benchmarks show that AReaL achieves up to 2.77x training speedup compared to synchronous systems with the same number of GPUs and matched or even improved final performance. The code of AReaL is available at https://github.com/inclusionAI/AReaL/.

The scaling law for large language models (LLMs) depicts that the path towards machine intelligence necessitates training at large scale. Thus, companies continuously build large-scale GPU clusters, and launch training jobs that span over thousands of computing nodes. However, LLM pre-training presents unique challenges due to its complex communication patterns, where GPUs exchange data in sparse yet high-volume bursts within specific groups. Inefficient resource scheduling exacerbates bandwidth contention, leading to suboptimal training performance. This paper presents Arnold, a scheduling system summarizing our experience to effectively align LLM communication patterns to data center topology at scale. In-depth characteristic study is performed to identify the impact of physical network topology to LLM pre-training jobs. Based on the insights, we develop a scheduling algorithm to effectively align communication patterns to physical network topology in data centers. Through simulation experiments, we show the effectiveness of our algorithm in reducing the maximum spread of communication groups by up to $1.67$x. In production training, our scheduling system improves the end-to-end performance by $10.6\%$ when training with more than $9600$ Hopper GPUs, a significant improvement for our training pipeline.

As distributed optimization scales to meet the demands of Large Language Model (LLM) training, hardware failures become increasingly non-negligible. Existing fault-tolerant training methods often introduce significant computational or memory overhead, demanding additional resources. To address this challenge, we propose **Me**mory- and **C**omputation- **e**fficient **F**ault-tolerant **O**ptimization (**MeCeFO**), a novel algorithm that ensures robust training with minimal overhead. When a computing node fails, MeCeFO seamlessly transfers its training task to a neighboring node while employing memory- and computation-efficient algorithmic optimizations to minimize the extra workload imposed on the neighboring node handling both tasks. MeCeFO leverages three key algorithmic designs: (i) Skip-connection, which drops the multi-head attention (MHA) module during backpropagation for memory- and computation-efficient approximation; (ii) Recomputation, which reduces activation memory in feedforward networks (FFNs); and (iii) Low-rank gradient approximation, enabling efficient estimation of FFN weight matrix gradients. Theoretically, MeCeFO matches the convergence rate of conventional distributed training, with a rate of $\mathcal{O}(1/\sqrt{nT})$, where $n$ is the data parallelism size and $T$ is the number of iterations. Empirically, MeCeFO maintains robust performance under high failure rates, incurring only a 4.18\% drop in throughput, demonstrating $5.0\times$ to $6.7\times$ greater resilience than previous SOTA approaches.

The rapid advancements in large language models (LLMs), particularly in their reasoning capabilities, hold transformative potential for addressing complex challenges and boosting scientific discovery in atmospheric science. However, leveraging LLMs effectively in this domain requires a robust and comprehensive evaluation benchmark. Toward this end, we present AtmosSci-Bench, a novel benchmark designed to systematically assess LLM performance across five core categories of atmospheric science problems: hydrology, atmospheric dynamics, atmospheric physics, geophysics, and physical oceanography.AtmosSci-Bench features a dual-format design comprising both multiple-choice questions (MCQs) and open-ended questions (OEQs), enabling scalable automated evaluation alongside deeper analysis of conceptual understanding. We employ a template-based MCQ generation framework to create diverse, graduate-level problems with symbolic perturbation, while OEQs are used to probe open-ended reasoning.We conduct a comprehensive evaluation of representative LLMs, categorized into four groups: instruction-tuned models, advanced reasoning models, math-augmented models, and domain-specific climate models. Our analysis provides some interesting insights into the reasoning and problem-solving capabilities of LLMs in atmospheric science. We believe AtmosSci-Bench can serve as a critical step toward advancing LLM applications in climate services by offering a standard and rigorous evaluation framework. The source code of AtmosSci-Bench is available at https://github.com/Relaxed-System-Lab/AtmosSci-Bench.

Large language models (LLMs) have been widely applied for their remarkable capability of content generation. However, the practical use of open-source LLMs is hindered by high resource requirements, making deployment expensive and limiting widespread development. The collaborative inference is a promising solution for this problem, in which users collaborate by each hosting a subset of layers and transmitting intermediate activation. Many companies are building collaborative inference platforms to reduce LLM serving costs, leveraging users' underutilized GPUs. Despite widespread interest in collaborative inference within academia and industry, the privacy risks associated with LLM collaborative inference have not been well studied. This is largely because of the challenge posed by inverting LLM activation due to its strong non-linearity. In this paper, to validate the severity of privacy threats in LLM collaborative inference, we introduce the concept of prompt inversion attack (PIA), where a malicious participant intends to recover the input prompt through the activation transmitted by its previous participant. Specifically, we design a two-stage method to execute this attack. In the first stage, we optimize the input embedding with a constraint term derived from the LLM's embedding matrix to enforce the optimized embedding to be close to the ground truth. In the second stage, we accurately recover discrete tokens by incorporating activation calibration and semantic speculation. Extensive experiments show that our PIA method substantially outperforms existing baselines. For example, our method achieves an 88.4% token accuracy on the Skytrax dataset with the Llama-65B model when inverting the maximum number of transformer layers, while the best baseline method only achieves 22.8% accuracy. The results verify the effectiveness of our PIA attack and highlights its practical threat to LLM collaborative inference systems.

Serving generative inference of the large language model is a crucial component of contemporary AI applications. In this paper, our focus lies in deploying such services in a heterogeneous and cross-datacenter setting to mitigate the substantial inference costs typically associated with a single centralized datacenter. Towards this end, we propose HexGen, a flexible distributed inference engine that uniquely supports the asymmetric partition of generative inference computations over both tensor model parallelism and pipeline parallelism, which allows for effective deployment across diverse GPUs interconnected by a fully heterogeneous network. We further propose a sophisticated scheduling algorithm grounded in constrained optimization that can adaptively assign asymmetric inference computation across the GPUs to fulfill inference requests while maintaining acceptable latency levels. We conduct an extensive empirical study to evaluate the efficiency of HexGen by serving the state-of-the-art Llama-2 (70B) model. The experimental results suggest that HexGen can choose to achieve up to $2.3\times$ lower latency deadlines or tolerate up to $4\times$ more traffic request rates compared with the homogeneous baseline given the same budget.

Modern machine learning applications increasingly demand greater computational resources for training large models. Decentralized training has emerged as an effective means to democratize this technology. However, the potential threats associated with this approach remain inadequately discussed, posing a hurdle to the development of decentralized training infrastructures. This paper aims to initiate discussion towards this end by exploring the robustness of decentralized training from three primary perspectives. Firstly, we articulate our position on establishing robust decentralized training by outlining potential threats and the corresponding countermeasures. Secondly, we illustrate a nascent poisoning attack targeting decentralized training frameworks, easily executable by malicious stages. To mitigate this security threat and ensure efficient training, we propose a robust training framework, integrating a 100% detection strategy and efficient training mechanisms. Finally, we demonstrate the severity of the proposed attack and the effectiveness of our robust training framework. This position paper emphasizes the urgency of exploring the robustness of decentralized training and proposes a feasible solution. The code is available at https://github.com/dcx001016/pipeline_attack.

Abstract Modern machine learning (ML) systems commonly use stochastic gradient descent (SGD) to train ML models. However, SGD relies on random data order to converge, which usually requires a full data shuffle. For in-DB ML systems and deep learning systems with large datasets stored onblock-addressable secondary storagesuch as HDD and SSD, this full data shuffle leads to low I/O performance—the data shuffling time can be even longer than the training itself, due to massive random data accesses. To balance the convergence rate of SGD (which favors data randomness) and its I/O performance (which favors sequential access), previous work has proposed several data shuffling strategies. In this paper, we first perform an empirical study on existing data shuffling strategies, showing that these strategies suffer from either low performance or low convergence rate. To solve this problem, we propose a simple but noveltwo-leveldata shuffling strategy named , which canavoida full data shuffle while maintainingcomparableconvergence rate of SGD as if a full shuffle were performed. We further theoretically analyze the convergence behavior of and empirically evaluate its efficacy in both in-DB ML and deep learning systems. For in-DB ML systems, we integrate into PostgreSQL by introducing three newphysicaloperators with optimizations. For deep learning systems, we extend single-process to multi-process for the parallel/distributed environment and integrate it into PyTorch. Our evaluation shows that can achieve comparable convergence rate with the full-shuffle-based SGD for both linear models and deep learning models. For in-DB ML with linear models, is 1.6 $$\times $$ × $$-$$ - 12.8 $$\times $$ × faster than two state-of-the-art systems, Apache MADlib and Bismarck, on both HDD and SSD. For deep learning models on ImageNet, is 1.5 $$\times $$ × faster than PyTorch with full data shuffle.

The high computational and memory requirements of large language model (LLM) inference make it feasible only with multiple high-end accelerators. Motivated by the emerging demand for latency-insensitive tasks with batched processing, this paper initiates the study of high-throughput LLM inference using limited resources, such as a single commodity GPU. We present FlexGen, a high-throughput generation engine for running LLMs with limited GPU memory. FlexGen can be flexibly configured under various hardware resource constraints by aggregating memory and computation from the GPU, CPU, and disk. By solving a linear programming problem, it searches for efficient patterns to store and access tensors. FlexGen further compresses the weights and the attention cache to 4 bits with negligible accuracy loss. These techniques enable FlexGen to have a larger space of batch size choices and thus significantly increase maximum throughput. As a result, when running OPT-175B on a single 16GB GPU, FlexGen achieves significantly higher throughput compared to state-of-the-art offloading systems, reaching a generation throughput of 1 token/s for the first time with an effective batch size of 144. On the HELM benchmark, FlexGen can benchmark a 30B model with a 16GB GPU on 7 representative sub-scenarios in 21 hours. The code is available at https://github.com/FMInference/FlexGen.

Large language models (LLMs) with hundreds of billions of parameters have sparked a new wave of exciting AI applications. However, they are computationally expensive at inference time. Sparsity is a natural approach to reduce this cost, but existing methods either require costly retraining, have to forgo LLM's in-context learning ability, or do not yield wall-clock time speedup on modern hardware. We hypothesize that contextual sparsity, which are small, input-dependent sets of attention heads and MLP parameters that yield approximately the same output as the dense model for a given input, can address these issues. We show that contextual sparsity exists, that it can be accurately predicted, and that we can exploit it to speed up LLM inference in wall-clock time without compromising LLM's quality or in-context learning ability. Based on these insights, we propose DejaVu, a system that uses a low-cost algorithm to predict contextual sparsity on the fly given inputs to each layer, along with an asynchronous and hardware-aware implementation that speeds up LLM inference. We validate that DejaVu can reduce the inference latency of OPT-175B by over 2$\times$ compared to the state-of-the-art FasterTransformer, and over 6$\times$ compared to the widely used Hugging Face implementation, without compromising model quality. The code is available at https://github.com/FMInference/DejaVu.

The relational data model was designed to facilitate large-scale data management and analytics. We consider the problem of how to differentiate computations expressed relationally. We show experimentally that a relational engine running an auto-differentiated relational algorithm can easily scale to very large datasets, and is competitive with state-of-the-art, special-purpose systems for large-scale distributed machine learning.

Distributed training of foundation models, especially large language models (LLMs), is communication-intensive and so has heavily relied on centralized data centers with fast interconnects. Can we train on slow networks and unlock the potential of decentralized infrastructure for foundation models? In this paper, we propose CocktailSGD, a novel communication-efficient training framework that combines three distinct compression techniques -- random sparsification, top-K sparsification, and quantization -- to achieve much greater compression than each individual technique alone. We justify the benefit of such a hybrid approach through a theoretical analysis of convergence. Empirically, we show that CocktailSGD achieves up to 117$\times$ compression in fine-tuning LLMs up to 20 billion parameters without hurting convergence. On a 500Mbps network, CocktailSGD only incurs $\sim$1.2$\times$ slowdown compared with data center networks.

This paper discusses why flow scheduling does not apply to distributed deep learning training and presents EchelonFlow, the first network abstraction to bridge the gap. EchelonFlow deviates from the common belief that semantically related flows should finish at the same time. We reached the key observation, after extensive workflow analysis of diverse training paradigms, that distributed training jobs observe strict computation patterns, which may consume data at different times. We devise a generic method to model the drastically different computation patterns across training paradigms, and formulate EchelonFlow to regulate flow finish times accordingly. Case studies of mainstream training paradigms under EchelonFlow demonstrate the expressiveness of the abstraction, and our system sketch suggests the feasibility of an EchelonFlow scheduling system.

Recent years have witnessed an exponential growth of model scale in deep learning-based recommender systems---from Google's 2016 model with 1 billion parameters to the latest Facebook's model with 12 trillion parameters. Significant quality boost has come with each jump of the model capacity, which makes us believe the era of 100 trillion parameters is around the corner. However, the training of such models is challenging even within industrial scale data centers. We resolve this challenge by careful co-design of both optimization algorithm and distributed system architecture. Specifically, to ensure both the training efficiency and the training accuracy, we design a novel hybrid training algorithm, where the embedding layer and the dense neural network are handled by different synchronization mechanisms; then we build a system called Persia (short for parallel recommendation training system with hybrid acceleration) to support this hybrid training algorithm. Both theoretical demonstrations and empirical studies with up to 100 trillion parameters have been conducted to justify the system design and implementation of Persia. We make Persia publicly available (at github.com/PersiaML/Persia) so that anyone can easily train a recommender model at the scale of 100 trillion parameters.

Communication compression is a crucial technique for modern distributed learning systems to alleviate their communication bottlenecks over slower networks. Despite recent intensive studies of gradient compression for data parallel-style training, compressing the activations for models trained with pipeline parallelism is still an open problem. In this paper, we propose AQ-SGD, a novel activation compression algorithm for communication-efficient pipeline parallelism training over slow networks. Different from previous efforts in activation compression, instead of compressing activation values directly, AQ-SGD compresses the changes of the activations. This allows us to show, to the best of our knowledge for the first time, that one can still achieve $O(1/\sqrt{T})$ convergence rate for non-convex objectives under activation compression, without making assumptions on gradient unbiasedness that do not hold for deep learning models with non-linear activation functions. We then show that AQ-SGD can be optimized and implemented efficiently, without additional end-to-end runtime overhead. We evaluated AQ-SGD to fine-tune language models with up to 1.5 billion parameters, compressing activation to 2-4 bits. AQ-SGD provides up to $4.3\times$ end-to-end speed-up in slower networks, without sacrificing model quality. Moreover, we also show that AQ-SGD can be combined with state-of-the-art gradient compression algorithms to enable end-to-end communication compression: All communications between machines, including model gradients, forward activations, and backward gradients are compressed into lower precision. This provides up to $4.9\times$ end-to-end speed-up, without sacrificing model quality.

Training foundation models, such as GPT-3 and PaLM, can be extremely expensive, often involving tens of thousands of GPUs running continuously for months. These models are typically trained in specialized clusters featuring fast, homogeneous interconnects and using carefully designed software systems that support both data parallelism and model/pipeline parallelism. Such dedicated clusters can be costly and difficult to obtain. Can we instead leverage the much greater amount of decentralized, heterogeneous, and lower-bandwidth interconnected compute? Previous works examining the heterogeneous, decentralized setting focus on relatively small models that can be trained in a purely data parallel manner. State-of-the-art schemes for model parallel foundation model training, such as Megatron and Deepspeed, only consider the homogeneous data center setting. In this paper, we present the first study of training large foundation models with model parallelism in a decentralized regime over a heterogeneous network. Our key technical contribution is a scheduling algorithm that allocates different computational “tasklets” in the training of foundation models to a group of decentralized GPU devices connected by a slow heterogeneous network. We provide a formal cost model and further propose an efficient evolutionary algorithm to find the optimal allocation strategy. We conduct extensive experiments that represent different scenarios for learning over geo-distributed devices simulated using real-world network measurements. In the most extreme case, across 8 different cities spanning 3 continents, our approach is 4.8× faster than prior state-of-the-art training systems.

Stochastic gradient descent (SGD) is the cornerstone of modern ML systems. Despite its computational efficiency, SGD requires random data access that is inherently inefficient when implemented in systems that rely on block-addressable secondary storage such as HDD and SSD, e.g., in-DB ML systems and TensorFlow/PyTorch over large files. To address this impedance mismatch, various data shuffling strategies have been proposed to balance the convergence rate of SGD (which favors randomness) and its I/O performance (which favors sequential access).

Distributed machine learning (ML) can bring more computational resources to bear than single-machine learning, thus enabling reductions in training time. Distributed learning partitions models and data over many machines, allowing model and dataset sizes beyond the available compute power and memory of a single machine. In practice though, distributed ML is challenging when distribution is mandatory, rather than chosen by the practitioner. In such scenarios, data could unavoidably be separated among workers due to limited memory capacity per worker or even because of data privacy issues. There, existing distributed methods will utterly fail due to dominant transfer costs across workers, or do not even apply. We propose a new approach to distributed fully connected neural network learning, called independent subnet training (IST), to handle these cases. In IST, the original network is decomposed into a set of narrow subnetworks with the same depth. These subnetworks are then trained locally before parameters are exchanged to produce new subnets and the training cycle repeats. Such a naturally "model parallel" approach limits memory usage by storing only a portion of network parameters on each device. Additionally, no requirements exist for sharing data between workers (i.e., subnet training is local and independent) and communication volume and frequency are reduced by decomposing the original network into independent subnets. These properties of IST can cope with issues due to distributed data, slow interconnects, or limited device memory, making IST a suitable approach for cases of mandatory distribution. We show experimentally that IST results in training times that are much lower than common distributed learning approaches.

Machine learning (ML) computations are often expressed using vectors, matrices, or higher-dimensional tensors. Such data structures can have many different implementations, especially in a distributed environment: a matrix could be stored as row or column vectors, tiles of different sizes, or relationally, as a set of (rowIndex, colIndex, value) triples. Many other storage formats are possible. The choice of format can have a profound impact on the performance of a ML computation. In this paper, we propose a framework for automatic optimization of the physical implementation of a complex ML or linear algebra (LA) computation in a distributed environment, develop algorithms for solving this problem, and show, through a prototype on top of a distributed relational database system, that our ideas can radically speed up common ML and LA computations.

Recent years have witnessed a growing list of systems for distributed data-parallel training. Existing systems largely fit into two paradigms, i.e., parameter server and MPI-style collective operations. On the algorithmic side, researchers have proposed a wide range of techniques to lower the communication via "system relaxations": quantization, decentralization, and communication delay. However, most, if not all, existing systems only rely on standard synchronous and asynchronous stochastic gradient (SG) based optimization, therefore, cannot take advantage of all possible optimizations that the machine learning community has been developing recently. Given this emerging gap between the current landscapes of systems and theory, we build Bagua, a MPI-style communication library, providing a collection of primitives, that is both flexible and modular to support state-of-the-art system relaxation techniques of distributed training. Powered by this design, Bagua has a great ability to implement and extend various state-of-the-art distributed learning algorithms. In a production cluster with up to 16 machines (128 GPUs), Bagua can outperform PyTorch-DDP, Horovod and BytePS in the end-to-end training time by a significant margin (up to 2X) across a diverse range of tasks. Moreover, we conduct a rigorous tradeoff exploration showing that different algorithms and system relaxations achieve the best performance over different network conditions.

When numerical and machine learning (ML) computations are expressed relationally, classical query execution strategies (hash-based joins and aggregations) can do a poor job distributing the computation. In this paper, we propose a two-phase execution strategy for numerical computations that are expressed relationally, as aggregated join trees (that is, expressed as a series of relational joins followed by an aggregation). In a pilot run, lineage information is collected; this lineage is used to optimally plan the computation at the level of individual records. Then, the computation is actually executed. We show experimentally that a relational system making use of this two-phase strategy can be an excellent platform for distributed ML computations.

We consider the question: what is the abstraction that should be implemented by the computational engine of a machine learning system? Current machine learning systems typically push whole tensors through a series of compute kernels such as matrix multiplications or activation functions, where each kernel runs on an AI accelerator (ASIC) such as a GPU. This implementation abstraction provides little built-in support for ML systems to scale past a single machine, or for handling large models with matrices or tensors that do not easily fit into the RAM of an ASIC. In this paper, we present an alternative implementation abstraction called the tensor relational algebra (TRA). The TRA is a set-based algebra based on the relational algebra. Expressions in the TRA operate over binary tensor relations, where keys are multi-dimensional arrays and values are tensors. The TRA is easily executed with high efficiency in a parallel or distributed environment, and amenable to automatic optimization. Our empirical study shows that the optimized TRA-based back-end can significantly outperform alternatives for running ML workflows in distributed clusters.

Partitioning is effective in avoiding expensive shuffling operations. However, it remains a significant challenge to automate this process for Big Data analytics workloads that extensively use user defined functions (UDFs), where sub-computations are hard to be reused for partitionings compared to relational applications. In addition, functional dependency that is widely utilized for partitioning selection is often unavailable in the unstructured data that is ubiquitous in UDF-centric analytics. We propose the Lachesis system, which represents UDF-centric workloads as workflows of analyzable and reusable sub-computations. Lachesis further adopts a deep reinforcement learning model to infer which sub-computations should be used to partition the underlying data. This analysis is then applied to automatically optimize the storage of the data across applications to improve the performance and users' productivity.

Automatic analysis of biomedical signals plays an important role in the auxiliary diagnosis of diseases. Traditional methods extract hand-crafted features by imitating doctors' experience, while recent methods focus on extracting deep features automatically by designing the architectures of deep neural networks (DNNs). Combining these two kinds of features can not only take advantage of doctors' experience but also mine the hidden information in the raw data. But directly combining these features by fully connected layers may cause complex optimization hyper-planes. To better integrate doctors' experience and deep features that doctors can hardly describe, we propose a feature fusion framework named hybrid plus framework (HPF) and apply this framework to seizure detection. HPF mainly consists of two parts: (1) the FET module, where hand-crafted features are extracted and transformed to sparse categorical features; (2) the enhanced DNN, which contains a carefully designed neural network structure with the input being original signals and sparse categorical features. Experiments on the dataset of CHB-MIT show that HPF outperforms the state-of-the-art methods. Further experiments indicate that HPF is very flexible as many of its modules can be replaced.

A number of popular systems, most notably Google's TensorFlow, have been implemented from the ground up to support machine learning tasks. We consider how to make a very small set of changes to a modern relational database management system (RDBMS) to make it suitable for distributed learning computations. Changes include adding better support for recursion, and optimization and execution of very large compute plans. We also show that there are key advantages to using an RDBMS as a machine learning platform. In particular, learning based on a database management system allows for trivial scaling to large data sets and especially large models, where different computational units operate on different parts of a model that may be too large to fit into RAM.

This paper describes PlinyCompute, a system for development of high-performance, data-intensive, distributed computing tools and libraries. \emphIn the large, PlinyCompute presents the programmer with a very high-level, declarative interface, relying on automatic, relational-database style optimization to figure out how to stage distributed computations. However, in the small, PlinyCompute presents the capable systems programmer with a persistent object data model and API (the "PC object model'') and associated memory management system that has been designed from the ground-up for high performance, distributed, data-intensive computing. This contrasts with most other Big Data systems, which are constructed on top of the Java Virtual Machine (JVM), and hence must at least partially cede performance-critical concerns such as memory management (including layout and de/allocation) and virtual method/function dispatch to the JVM. This hybrid approach---declarative in the large, trusting the programmer's ability to utilize PC object model efficiently in the small---results in a system that is ideal for the development of reusable, data-intensive tools and libraries.

In this paper, we consider the problem of source code abridgment, where the goal is to remove statements from a source code in order to display the source code in a small space, while at the same time leaving the ``important'' parts of the source code intact, so that an engineer can read the code and quickly understand purpose of the code. To this end, we develop an algorithm that looks at a number of examples, human-created source code abridgments, and learns how to remove lines from the code in order to mimic the human abridger. The learning algorithm takes into account syntactic features of the code, as well as semantic features such as control flow and data dependencies. Through a comprehensive user study, we show that the abridgments that our system produces can decrease the time that a user must look at code in order to understand its functionality, as well as increase the accuracy of the assessment, while displaying the code in a greatly reduced area.

This paper presents an effective video retargeting method with the assessment of jittery artifact. Our method firstly constructs a new energy function by including both spatial and temporal constrains. Then the retargeting processing can be performed effectively by minimizing our proposed energy function for each frame. We also propose a new objective measurement to assess temporal coherence after video retargeting, the accuracy of which is verified by psycho-visual tests. Experimental results show that our video retargeting method is able to provide comfortable resized videos in terms of subjective and objective measurements.