Ana Klimovic

Andrew Audibert, Yang Chen, Dan Graur et al.

Machine learning (ML) computations commonly execute on expensive specialized hardware, such as GPUs and TPUs, which provide high FLOPs and performance-per-watt. For cost efficiency, it is essential to keep these accelerators highly utilized. This requires preprocessing input data at the rate at which the accelerators can ingest and perform ML computations on the data. To avoid data stalls, the host CPU and RAM required for input data processing per accelerator core used for ML computations varies across jobs. Hence, the traditional approach of processing input data on ML accelerator hosts with a fixed hardware ratio leads to either under-utilizing the accelerators or the host CPU and RAM. In this paper, we address these concerns by building a disaggregated ML data processing system. We present tf.data service, an open-source disaggregated input data processing service built on top of tf.data in TensorFlow. We show that disaggregating data preprocessing has three key advantages for large-scale ML training jobs. First, the service can horizontally scale-out to right-size CPU/RAM host resources for data processing in each job, saving 32x training time and 26x cost, on average. Second, the service can share ephemeral preprocessed data results across jobs, to optimize CPU usage and reduce redundant computations. Finally, the service supports coordinated reads, a technique that avoids stragglers due to different input sizes in distributed training, reducing training time by 2.2x, on average. Our design is inspired by lessons learned from deploying tf.data service in production, including relaxing data visitation guarantees without impacting model accuracy.

Alejandro Cuadron, Dacheng Li, Wenjie Ma et al.

Large Reasoning Models (LRMs) represent a breakthrough in AI problem-solving capabilities, but their effectiveness in interactive environments can be limited. This paper introduces and analyzes overthinking in LRMs. A phenomenon where models favor extended internal reasoning chains over environmental interaction. Through experiments on software engineering tasks using SWE Bench Verified, we observe three recurring patterns: Analysis Paralysis, Rogue Actions, and Premature Disengagement. We propose a framework to study these behaviors, which correlates with human expert assessments, and analyze 4018 trajectories. We observe that higher overthinking scores correlate with decreased performance, with reasoning models exhibiting stronger tendencies toward overthinking compared to non-reasoning models. Our analysis reveals that simple efforts to mitigate overthinking in agentic environments, such as selecting the solution with the lower overthinking score, can improve model performance by almost 30% while reducing computational costs by 43%. These results suggest that mitigating overthinking has strong practical implications. We suggest that by leveraging native function-calling capabilities and selective reinforcement learning overthinking tendencies could be mitigated. We also open-source our evaluation framework and dataset to facilitate research in this direction at https://github.com/AlexCuadron/Overthinking.

Cedric Renggli, Xiaozhe Yao, Luka Kolar et al.

Transfer learning can be seen as a data- and compute-efficient alternative to training models from scratch. The emergence of rich model repositories, such as TensorFlow Hub, enables practitioners and researchers to unleash the potential of these models across a wide range of downstream tasks. As these repositories keep growing exponentially, efficiently selecting a good model for the task at hand becomes paramount. By carefully comparing various selection and search strategies, we realize that no single method outperforms the others, and hybrid or mixed strategies can be beneficial. Therefore, we propose SHiFT, the first downstream task-aware, flexible, and efficient model search engine for transfer learning. These properties are enabled by a custom query language SHiFT-QL together with a cost-based decision maker, which we empirically validate. Motivated by the iterative nature of machine learning development, we further support efficient incremental executions of our queries, which requires a careful implementation when jointly used with our optimizations.

Christopher Zanoli, Andrea Giovannini, Tengjun Jin et al.

Constructing Extract-Load-Transform (ELT) pipelines is a labor-intensive data engineering task and a high-impact target for AI automation. On ELT-Bench, the first benchmark for end-to-end ELT pipeline construction, AI agents initially showed low success rates, suggesting they lacked practical utility. We revisit these results and identify two factors causing a substantial underestimation of agent capabilities. First, re-evaluating ELT-Bench with upgraded large language models reveals that the extraction and loading stage is largely solved, while transformation performance improves significantly. Second, we develop an Auditor-Corrector methodology that combines scalable LLM-driven root-cause analysis with rigorous human validation (inter-annotator agreement Fleiss' kappa = 0.85) to audit benchmark quality. Applying this to ELT-Bench uncovers that most failed transformation tasks contain benchmark-attributable errors -- including rigid evaluation scripts, ambiguous specifications, and incorrect ground truth -- that penalize correct agent outputs. Based on these findings, we construct ELT-Bench-Verified, a revised benchmark with refined evaluation logic and corrected ground truth. Re-evaluating on this version yields significant improvement attributable entirely to benchmark correction. Our results show that both rapid model improvement and benchmark quality issues contributed to underestimating agent capabilities. More broadly, our findings echo observations of pervasive annotation errors in text-to-SQL benchmarks, suggesting quality issues are systemic in data engineering evaluation. Systematic quality auditing should be standard practice for complex agentic tasks. We release ELT-Bench-Verified to provide a more reliable foundation for progress in AI-driven data engineering automation.

Maximilian Böther, Ties Robroek, Viktor Gsteiger et al.

In real-world machine learning (ML) pipelines, datasets are continuously growing. Models must incorporate this new training data to improve generalization and adapt to potential distribution shifts. The cost of model retraining is proportional to how frequently the model is retrained and how much data it is trained on, which makes the naive approach of retraining from scratch each time impractical. We present Modyn, a data-centric end-to-end machine learning platform. Modyn's ML pipeline abstraction enables users to declaratively describe policies for continuously training a model on a growing dataset. Modyn pipelines allow users to apply data selection policies (to reduce the number of data points) and triggering policies (to reduce the number of trainings). Modyn executes and orchestrates these continuous ML training pipelines. The system is open-source and comes with an ecosystem of benchmark datasets, models, and tooling. We formally discuss how to measure the performance of ML pipelines by introducing the concept of composite models, enabling fair comparison of pipelines with different data selection and triggering policies. We empirically analyze how various data selection and triggering policies impact model accuracy, and also show that Modyn enables high throughput training with sample-level data selection.

Qinghao Hu, Shang Yang, Junxian Guo et al.

The emergence of Large Language Models (LLMs) with strong reasoning capabilities marks a significant milestone, unlocking new frontiers in complex problem-solving. However, training these reasoning models, typically using Reinforcement Learning (RL), encounters critical efficiency bottlenecks: response generation during RL training exhibits a persistent long-tail distribution, where a few very long responses dominate execution time, wasting resources and inflating costs. To address this, we propose TLT, a system that accelerates reasoning RL training losslessly by integrating adaptive speculative decoding. Applying speculative decoding in RL is challenging due to the dynamic workloads, evolving target model, and draft model training overhead. TLT overcomes these obstacles with two synergistic components: (1) Adaptive Drafter, a lightweight draft model trained continuously on idle GPUs during long-tail generation to maintain alignment with the target model at no extra cost; and (2) Adaptive Rollout Engine, which maintains a memory-efficient pool of pre-captured CUDAGraphs and adaptively select suitable SD strategies for each input batch. Evaluations demonstrate that TLT achieves over 1.7x end-to-end RL training speedup over state-of-the-art systems, preserves the model accuracy, and yields a high-quality draft model as a free byproduct suitable for efficient deployment. Code is released at https://github.com/mit-han-lab/fastrl.

Xiaozhe Yao, Qinghao Hu, Ana Klimovic

Fine-tuning large language models (LLMs) greatly improves model quality for downstream tasks. However, serving many fine-tuned LLMs concurrently is challenging due to the sporadic, bursty, and varying request patterns of different LLMs. To bridge this gap, we present DeltaZip, an LLM serving system that efficiently serves multiple full-parameter fine-tuned models concurrently by aggressively compressing model deltas by up to 10x while maintaining high model quality. The key insight behind this design is that fine-tuning results in small-magnitude changes to the pre-trained model. By co-designing the serving system with the compression algorithm, DeltaZip achieves 2x to 12x improvement in throughput compared to the state-of-the-art systems.

Maximilian Böther, Abraham Sebastian, Pranjal Awasthi et al.

Modern datasets span billions of samples, making training on all available data infeasible. Selecting a high quality subset helps in reducing training costs and enhancing model quality. Submodularity, a discrete analogue of convexity, is commonly used for solving such subset selection problems. However, existing algorithms for optimizing submodular functions are sequential, and the prior distributed methods require at least one central machine to fit the target subset in DRAM. At billion datapoint scale, even the subset may not fit a single machine, and the sequential algorithms are prohibitively slow. In this paper, we relax the requirement of having a central machine for the target subset by proposing a novel distributed bounding algorithm with provable approximation guarantees. The algorithm iteratively bounds the minimum and maximum utility values to select high quality points and discard the unimportant ones. When bounding does not find the complete subset, we use a multi-round, partition-based distributed greedy algorithm to identify the remaining subset. We discuss how to implement these algorithms in a distributed data processing framework and empirically analyze different configurations. We find high quality subsets on CIFAR-100 and ImageNet with marginal or no loss in quality compared to centralized methods, and scale to a dataset with 13 billion points.

Zerui Wang, Qinghao Hu, Ana Klimovic et al.

Deep learning (DL) schedulers are pivotal in optimizing resource allocation in GPU clusters, but operate with a critical limitation: they are largely blind to the semantic context of the jobs they manage. This forces them to rely on limited metadata, leading to high profiling overhead, unreliable duration estimation, inadequate failure handling, and poor observability. To this end, we propose SchedMate, a framework that bridges this semantic gap by systematically extracting deep insights from overlooked, unstructured data sources: source code, runtime logs, and historical jobs. SchedMate enhances existing schedulers non-intrusively through three LLM-based components. Our implementation integrates seamlessly with existing deep learning schedulers. Evaluations on a 128-GPU physical cluster and extensive simulations on production traces show SchedMate reduces average job completion times by up to 1.91x, substantially enhancing the scheduling performance, demonstrating the critical role of semantic-awareness in modern DL scheduling.

Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang et al. · eth-zurich

We present Apertus, a fully open suite of large language models (LLMs) designed to address two systemic shortcomings in today's open model ecosystem: data compliance and multilingual representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of memorization, we adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. The Apertus models also expand multilingual coverage, training on 15T tokens from over 1800 languages, with ~40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivalling or surpassing open-weight counterparts. Beyond model weights, we release all scientific artifacts from our development cycle with a permissive license, including data preparation scripts, checkpoints, evaluation suites, and training code, enabling transparent audit and extension.

Maximilian Böther, Xiaozhe Yao, Tolga Kerimoglu et al.

State-of-the-art large language and vision models are trained over trillions of tokens that are aggregated from a large variety of sources. As training data collections grow, manually managing the samples becomes time-consuming, tedious, and prone to errors. Yet recent research shows that the data mixture and the order in which samples are visited during training can significantly influence model accuracy. We build and present Mixtera, a data plane for foundation model training that enables users to declaratively express which data samples should be used in which proportion and in which order during training. Mixtera is a centralized, read-only layer that is deployed on top of existing training data collections and can be declaratively queried. It operates independently of the filesystem structure and supports mixtures across arbitrary properties (e.g., language, source dataset) as well as dynamic adjustment of the mixture based on model feedback. We experimentally evaluate Mixtera and show that our implementation does not bottleneck training and scales to 256 GH200 superchips. We demonstrate how Mixtera supports recent advancements in mixing strategies by implementing the proposed Adaptive Data Optimization (ADO) algorithm in the system and evaluating its performance impact. We also explore the role of mixtures for vision-language models.

Michael Kuchnik, Ana Klimovic, Jiri Simsa et al.

Input pipelines, which ingest and transform input data, are an essential part of training Machine Learning (ML) models. However, it is challenging to implement efficient input pipelines, as it requires reasoning about parallelism, asynchrony, and variability in fine-grained profiling information. Our analysis of over two million ML jobs in Google datacenters reveals that a significant fraction of model training jobs could benefit from faster input data pipelines. At the same time, our analysis indicates that most jobs do not saturate host hardware, pointing in the direction of software-based bottlenecks. Motivated by these findings, we propose Plumber, a tool for finding bottlenecks in ML input pipelines. Plumber uses an extensible and interpretable operational analysis analytical model to automatically tune parallelism, prefetching, and caching under host resource constraints. Across five representative ML pipelines, Plumber obtains speedups of up to 47x for misconfigured pipelines. By automating caching, Plumber obtains end-to-end speedups of over 50% compared to state-of-the-art tuners.

Jiawei Jiang, Shaoduo Gan, Yue Liu et al.

The appeal of serverless (FaaS) has triggered a growing interest on how to use it in data-intensive applications such as ETL, query processing, or machine learning (ML). Several systems exist for training large-scale ML models on top of serverless infrastructures (e.g., AWS Lambda) but with inconclusive results in terms of their performance and relative advantage over "serverful" infrastructures (IaaS). In this paper we present a systematic, comparative study of distributed ML training over FaaS and IaaS. We present a design space covering design choices such as optimization algorithms and synchronization protocols, and implement a platform, LambdaML, that enables a fair comparison between FaaS and IaaS. We present experimental results using LambdaML, and further develop an analytic model to capture cost/performance tradeoffs that must be considered when opting for a serverless infrastructure. Our results indicate that ML training pays off in serverless only for models with efficient (i.e., reduced) communication and that quickly converge. In general, FaaS can be much faster but it is never significantly cheaper than IaaS.

Derek G. Murray, Jiri Simsa, Ana Klimovic et al.

Training machine learning models requires feeding input data for models to ingest. Input pipelines for machine learning jobs are often challenging to implement efficiently as they require reading large volumes of data, applying complex transformations, and transferring data to hardware accelerators while overlapping computation and communication to achieve optimal performance. We present tf.data, a framework for building and executing efficient input pipelines for machine learning jobs. The tf.data API provides operators which can be parameterized with user-defined computation, composed, and reused across different machine learning domains. These abstractions allow users to focus on the application logic of data processing, while tf.data's runtime ensures that pipelines run efficiently. We demonstrate that input pipeline performance is critical to the end-to-end training time of state-of-the-art machine learning models. tf.data delivers the high performance required, while avoiding the need for manual tuning of performance knobs. We show that tf.data features, such as parallelism, caching, static optimizations, and non-deterministic execution are essential for high performance. Finally, we characterize machine learning input pipelines for millions of jobs that ran in Google's fleet, showing that input data processing is highly diverse and consumes a significant fraction of job resources. Our analysis motivates future research directions, such as sharing computation across jobs and pushing data projection to the storage layer.