Thomas Parnell

Paul Castro, Nick Mitchell, Nathan Ordonez et al.

Clients are evolving beyond chat completion, and now include a variety of innovative inference-time scaling and deep reasoning techniques. At the same time, inference servers remain heavily optimized for chat completion. Prior work has shown that large improvements to KV cache hit rate are possible if inference servers evolve towards these non-chat use cases. However, they offer solutions that are also optimized for a single use case, RAG. In this paper, we introduce the span query to generalize the interface to the inference server. We demonstrate that chat, RAG, inference-time scaling, and agentic workloads can all be expressed as span queries. We show how the critical distinction that had been assumed by prior work lies in whether the order of the inputs matter -- do they commute? In chat, they do not. In RAG, they often do. This paper introduces span queries, which are expression trees of inference calls, linked together with commutativity constraints. We describe span query syntax and semantics. We show how they can be automatically optimized to improve KV cache locality. We show how a small change to vLLM (affecting only 492 lines) can enable high-performance execution of span queries. Using this stack, we demonstrate that span queries can achieve 10-20x reductions in TTFT for two distinct non-chat use cases. Finally, we show that span queries can also be optimized to improve attention locality, so as to avoid the so-called lost-in-the-middle problem. We demonstrate that an attention-optimized span query on a 2b parameter model vastly outperforms the accuracy of a stock inference server using an 8b model.

Małgorzata Łazuka, Thomas Parnell, Andreea Anghel et al.

Multi-cloud computing has become increasingly popular with enterprises looking to avoid vendor lock-in. While most cloud providers offer similar functionality, they may differ significantly in terms of performance and/or cost. A customer looking to benefit from such differences will naturally want to solve the multi-cloud configuration problem: given a workload, which cloud provider should be chosen and how should its nodes be configured in order to minimize runtime or cost? In this work, we consider solutions to this optimization problem. We develop and evaluate possible adaptations of state-of-the-art cloud configuration solutions to the multi-cloud domain. Furthermore, we identify an analogy between multi-cloud configuration and the selection-configuration problems commonly studied in the automated machine learning (AutoML) field. Inspired by this connection, we utilize popular optimizers from AutoML to solve multi-cloud configuration. Finally, we propose a new algorithm for solving multi-cloud configuration, CloudBandit (CB). It treats the outer problem of cloud provider selection as a best-arm identification problem, in which each arm pull corresponds to running an arbitrary black-box optimizer on the inner problem of node configuration. Our experiments indicate that (a) many state-of-the-art cloud configuration solutions can be adapted to multi-cloud, with best results obtained for adaptations which utilize the hierarchical structure of the multi-cloud configuration domain, (b) hierarchical methods from AutoML can be used for the multi-cloud configuration task and can outperform state-of-the-art cloud configuration solutions and (c) CB achieves competitive or lower regret relative to other tested algorithms, whilst also identifying configurations that have 65% lower median cost and 20% lower median time in production, compared to choosing a random provider and configuration.

Burkhard Ringlein, Jan van Lunteren, Radu Stoica et al.

A long-standing goal in both industry and academia is to develop an LLM inference platform that is portable across hardware architectures, eliminates the need for low-level hand-tuning, and still delivers best-in-class efficiency. In this work, we demonstrate that portable, efficient cross-platform LLM inference is indeed possible and share our experience. We develop a state-of-the-art paged attention kernel, the core performance-critical component of many LLM deployments, that builds exclusively on the domain-specific just-in-time compiled language Triton to achieve state-of-the-art performance on both NVIDIA and AMD GPUs. We describe our high-level approach, the key algorithmic and system-level improvements, the parameter auto-tuning required to unlock efficiency, and the integrations into a popular inference server that are necessary to bring the performance of a generic Triton attention kernel from 19.7% of the state-of-the-art to 105.9%. Our results highlight how open-source domain-specific languages can be leveraged to unlock model portability across different GPU vendors.

Kristjan Greenewald, Luis Lastras, Thomas Parnell et al.

Low-Rank Adaptation (LoRA) has emerged as a highly efficient framework for finetuning the weights of large foundation models, and has become the go-to method for data-driven customization of LLMs. Despite the promise of highly customized behaviors and capabilities, switching between relevant LoRAs in a multiturn setting is inefficient, as the key-value (KV) cache of the entire turn history must be recomputed with the LoRA weights before generation can begin. To address this problem, we propose Activated LoRA (aLoRA), an adapter architecture which modifies the LoRA framework to only adapt weights for the tokens in the sequence after the aLoRA is invoked. This change crucially allows aLoRA to accept the base model's KV cache of the input string, meaning that aLoRA can be instantly activated whenever needed in a chain without recomputing the prior keys and values. This enables building what we call intrinsics, i.e. specialized models invoked to perform well-defined operations on portions of an input chain or conversation that otherwise uses the base model by default. We train a set of aLoRA-based intrinsics models, demonstrating competitive accuracy with standard LoRA while significantly improving inference efficiency. We contributed our Activated LoRA implementation to the Huggingface PEFT library https://github.com/huggingface/peft.

Davis Wertheimer, Joshua Rosenkranz, Thomas Parnell et al.

This technical report describes the design and training of novel speculative decoding draft models, for accelerating the inference speeds of large language models in a production environment. By conditioning draft predictions on both context vectors and sampled tokens, we can train our speculators to efficiently predict high-quality n-grams, which the base model then accepts or rejects. This allows us to effectively predict multiple tokens per inference forward pass, accelerating wall-clock inference speeds of highly optimized base model implementations by a factor of 2-3x. We explore these initial results and describe next steps for further improvements.

Burkhard Ringlein, Thomas Parnell, Radu Stoica

As LLMs grow in complexity, achieving state-of-the-art performance requires tight co-design across algorithms, software, and hardware. Today's reliance on a single dominant platform limits portability, creates vendor lock-in, and raises barriers for new AI hardware. In this work, we make the case for combining just-in-time (JIT) compilation with comprehensive kernel parameter autotuning to enable portable LLM inference with state-of-the-art performance without code changes. Focusing on performance-critical LLM kernels, we demonstrate that this approach explores up to 15x more kernel parameter configurations, produces significantly more diverse code across multiple dimensions, and even outperforms vendor-optimized implementations by up to 230%, all while reducing kernel code size by 70x and eliminating manual code optimizations. Our results highlight autotuning as a promising path to unlocking model portability across GPU vendors.

Allison Li, Kristjan Greenewald, Thomas Parnell et al.

Modern large language model (LLM) systems increasingly rely on multi-turn pipelines that are composed of multiple task-specific adapters, yet existing serving frameworks remain inefficient, incurring substantial recomputation overhead when switching between adapters. We present the first LLM serving engine that supports cross-model prefix cache reuse between base and adapted models via Activated LoRA (aLoRA), enabling efficient and fine-grained adapter switching during inference. Our design extends the vLLM framework by introducing base-aligned block hashing and activation-aware masking within the model execution path, permitting cache reuse across models while preserving compatibility with existing serving engine optimizations. Integrated into a production-grade inference stack, this approach supports dynamic adapter activation without excessive key-value tensor recomputation. Evaluation across representative multi-turn, multi-adapter pipelines demonstrates up to 58x end-to-end latency reduction and over 100x time-to-first-token improvement relative to standard LoRA baselines, with benefits that scale with model size and sequence length and manifest across all stages of the request lifecycle. This work bridges parameter-efficient model adaptation with high-performance serving, providing the first complete realization of cross-model KV-cache reuse in modern LLM inference engines.

Thomas Schmied, Diego Didona, Andreas Döring et al.

Machine learning (ML) methods have recently emerged as an effective way to perform automated parameter tuning of databases. State-of-the-art approaches include Bayesian optimization (BO) and reinforcement learning (RL). In this work, we describe our experience when applying these methods to a database not yet studied in this context: FoundationDB. Firstly, we describe the challenges we faced, such as unknown valid ranges of configuration parameters and combinations of parameter values that result in invalid runs, and how we mitigated them. While these issues are typically overlooked, we argue that they are a crucial barrier to the adoption of ML self-tuning techniques in databases, and thus deserve more attention from the research community. Secondly, we present experimental results obtained when tuning FoundationDB using ML methods. Unlike prior work in this domain, we also compare with the simplest of baselines: random search. Our results show that, while BO and RL methods can improve the throughput of FoundationDB by up to 38%, random search is a highly competitive baseline, finding a configuration that is only 4% worse than the, vastly more complex, ML methods. We conclude that future work in this area may want to focus more on randomized, model-free optimization algorithms.

Thomas Parnell, Andreea Anghel, Malgorzata Lazuka et al.

Modern gradient boosting software frameworks, such as XGBoost and LightGBM, implement Newton descent in a functional space. At each boosting iteration, their goal is to find the base hypothesis, selected from some base hypothesis class, that is closest to the Newton descent direction in a Euclidean sense. Typically, the base hypothesis class is fixed to be all binary decision trees up to a given depth. In this work, we study a Heterogeneous Newton Boosting Machine (HNBM) in which the base hypothesis class may vary across boosting iterations. Specifically, at each boosting iteration, the base hypothesis class is chosen, from a fixed set of subclasses, by sampling from a probability distribution. We derive a global linear convergence rate for the HNBM under certain assumptions, and show that it agrees with existing rates for Newton's method when the Newton direction can be perfectly fitted by the base hypothesis at each boosting iteration. We then describe a particular realization of a HNBM, SnapBoost, that, at each boosting iteration, randomly selects between either a decision tree of variable depth or a linear regressor with random Fourier features. We describe how SnapBoost is implemented, with a focus on the training complexity. Finally, we present experimental results, using OpenML and Kaggle datasets, that show that SnapBoost is able to achieve better generalization loss than competing boosting frameworks, without taking significantly longer to tune.

Georgios Damaskinos, Celestine Mendler-Dünner, Rachid Guerraoui et al.

In this paper we tackle the challenge of making the stochastic coordinate descent algorithm differentially private. Compared to the classical gradient descent algorithm where updates operate on a single model vector and controlled noise addition to this vector suffices to hide critical information about individuals, stochastic coordinate descent crucially relies on keeping auxiliary information in memory during training. This auxiliary information provides an additional privacy leak and poses the major challenge addressed in this work. Driven by the insight that under independent noise addition, the consistency of the auxiliary information holds in expectation, we present DP-SCD, the first differentially private stochastic coordinate descent algorithm. We analyze our new method theoretically and argue that decoupling and parallelizing coordinate updates is essential for its utility. On the empirical side we demonstrate competitive performance against the popular stochastic gradient descent alternative (DP-SGD) while requiring significantly less tuning.

Nikolas Ioannou, Celestine Mendler-Dünner, Thomas Parnell

In this paper we propose a novel parallel stochastic coordinate descent (SCD) algorithm with convergence guarantees that exhibits strong scalability. We start by studying a state-of-the-art parallel implementation of SCD and identify scalability as well as system-level performance bottlenecks of the respective implementation. We then take a principled approach to develop a new SCD variant which is designed to avoid the identified system bottlenecks, such as limited scaling due to coherence traffic of model sharing across threads, and inefficient CPU cache accesses. Our proposed system-aware parallel coordinate descent algorithm (SySCD) scales to many cores and across numa nodes, and offers a consistent bottom line speedup in training time of up to x12 compared to an optimized asynchronous parallel SCD algorithm and up to x42, compared to state-of-the-art GLM solvers (scikit-learn, Vowpal Wabbit, and H2O) on a range of datasets and multi-core CPU architectures.

Andreea Anghel, Nikolas Ioannou, Thomas Parnell et al.

In this paper we analyze, evaluate, and improve the performance of training Random Forest (RF) models on modern CPU architectures. An exact, state-of-the-art binary decision tree building algorithm is used as the basis of this study. Firstly, we investigate the trade-offs between using different tree building algorithms, namely breadth-first-search (BFS) and depth-search-first (DFS). We design a novel, dynamic, hybrid BFS-DFS algorithm and demonstrate that it performs better than both BFS and DFS, and is more robust in the presence of workloads with different characteristics. Secondly, we identify CPU performance bottlenecks when generating trees using this approach, and propose optimizations to alleviate them. The proposed hybrid tree building algorithm for RF is implemented in the Snap Machine Learning framework, and speeds up the training of RFs by 7.8x on average when compared to state-of-the-art RF solvers (sklearn, H2O, and xgboost) on a range of datasets, RF configurations, and multi-core CPU architectures.

Johanna Sommer, Dimitrios Sarigiannis, Thomas Parnell

In this short paper we investigate whether meta-learning techniques can be used to more effectively tune the hyperparameters of machine learning models using successive halving (SH). We propose a novel variant of the SH algorithm (MeSH), that uses meta-regressors to determine which candidate configurations should be eliminated at each round. We apply MeSH to the problem of tuning the hyperparameters of a gradient-boosted decision tree model. By training and tuning our meta-regressors using existing tuning jobs from 95 datasets, we demonstrate that MeSH can often find a superior solution to both SH and random search.

Dimitrios Sarigiannis, Thomas Parnell, Haris Pozidis

The combined algorithm selection and hyperparameter tuning (CASH) problem is characterized by large hierarchical hyperparameter spaces. Model-free hyperparameter tuning methods can explore such large spaces efficiently since they are highly parallelizable across multiple machines. When no prior knowledge or meta-data exists to boost their performance, these methods commonly sample random configurations following a uniform distribution. In this work, we propose a novel sampling distribution as an alternative to uniform sampling and prove theoretically that it has a better chance of finding the best configuration in a worst-case setting. In order to compare competing methods rigorously in an experimental setting, one must perform statistical hypothesis testing. We show that there is little-to-no agreement in the automated machine learning literature regarding which methods should be used. We contrast this disparity with the methods recommended by the broader statistics literature, and identify a suitable approach. We then select three popular model-free solutions to CASH and evaluate their performance, with uniform sampling as well as the proposed sampling scheme, across 67 datasets from the OpenML platform. We investigate the trade-off between exploration and exploitation across the three algorithms, and verify empirically that the proposed sampling distribution improves performance in all cases.

Michael Kaufmann, Kornilios Kourtis, Celestine Mendler-Dünner et al.

Distributed machine learning training is one of the most common and important workloads running on data centers today, but it is rarely executed alone. Instead, to reduce costs, computing resources are consolidated and shared by different applications. In this scenario, elasticity and proper load balancing are vital to maximize efficiency, fairness, and utilization. Currently, most distributed training frameworks do not support the aforementioned properties. A few exceptions that do support elasticity, imitate generic distributed frameworks and use micro-tasks. In this paper we illustrate that micro-tasks are problematic for machine learning applications, because they require a high degree of parallelism which hinders the convergence of distributed training at a pure algorithmic level (i.e., ignoring overheads and scalability limitations). To address this, we propose Chicle, a new elastic distributed training framework which exploits the nature of machine learning algorithms to implement elasticity and load balancing without micro-tasks. We use Chicle to train deep neural network as well as generalized linear models, and show that Chicle achieves performance competitive with state of the art rigid frameworks, while efficiently enabling elastic execution and dynamic load balancing.

Martino Dazzi, Abu Sebastian, Pier Andrea Francese et al.

In-memory computing is an emerging computing paradigm that could enable deeplearning inference at significantly higher energy efficiency and reduced latency. The essential idea is to map the synaptic weights corresponding to each layer to one or more computational memory (CM) cores. During inference, these cores perform the associated matrix-vector multiply operations in place with O(1) time complexity, thus obviating the need to move the synaptic weights to an additional processing unit. Moreover, this architecture could enable the execution of these networks in a highly pipelined fashion. However, a key challenge is to design an efficient communication fabric for the CM cores. Here, we present one such communication fabric based on a graph topology that is well suited for the widely successful convolutional neural networks (CNNs). We show that this communication fabric facilitates the pipelined execution of all state of-the-art CNNs by proving the existence of a homomorphism between one graph representation of these networks and the proposed graph topology. We then present a quantitative comparison with established communication topologies and show that our proposed topology achieves the lowest bandwidth requirements per communication channel. Finally, we present a concrete example of mapping ResNet-32 onto an array of CM cores.

Alessandro De Palma, Celestine Mendler-Dünner, Thomas Parnell et al.

We present Acquisition Thompson Sampling (ATS), a novel technique for batch Bayesian Optimization (BO) based on the idea of sampling multiple acquisition functions from a stochastic process. We define this process through the dependency of the acquisition functions on a set of model hyper-parameters. ATS is conceptually simple, straightforward to implement and, unlike other batch BO methods, it can be employed to parallelize any sequential acquisition function or to make existing parallel methods scale further. We present experiments on a variety of benchmark functions and on the hyper-parameter optimization of a popular gradient boosting tree algorithm. These demonstrate the advantages of ATS with respect to classical parallel Thompson Sampling for BO, its competitiveness with two state-of-the-art batch BO methods, and its effectiveness if applied to existing parallel BO algorithms.

Michael Kaufmann, Thomas Parnell, Kornilios Kourtis

In this paper we experimentally analyze the convergence behavior of CoCoA and show, that the number of workers required to achieve the highest convergence rate at any point in time, changes over the course of the training. Based on this observation, we build Chicle, an elastic framework that dynamically adjusts the number of workers based on feedback from the training algorithm, in order to select the number of workers that results in the highest convergence rate. In our evaluation of 6 datasets, we show that Chicle is able to accelerate the time-to-accuracy by a factor of up to 5.96x compared to the best static setting, while being robust enough to find an optimal or near-optimal setting automatically in most cases.

Nikolas Ioannou, Celestine Dünner, Kornilios Kourtis et al.

In this paper we analyze, evaluate, and improve the performance of training generalized linear models on modern CPUs. We start with a state-of-the-art asynchronous parallel training algorithm, identify system-level performance bottlenecks, and apply optimizations that improve data parallelism, cache line locality, and cache line prefetching of the algorithm. These modifications reduce the per-epoch run-time significantly, but take a toll on algorithm convergence in terms of the required number of epochs. To alleviate these shortcomings of our systems-optimized version, we propose a novel, dynamic data partitioning scheme across threads which allows us to approach the convergence of the sequential version. The combined set of optimizations result in a consistent bottom line speedup in convergence of up to 12x compared to the initial asynchronous parallel training algorithm and up to 42x, compared to state of the art implementations (scikit-learn and h2o) on a range of multi-core CPU architectures.

Andreea Anghel, Nikolaos Papandreou, Thomas Parnell et al.

Gradient boosting decision trees (GBDTs) have seen widespread adoption in academia, industry and competitive data science due to their state-of-the-art performance in many machine learning tasks. One relative downside to these models is the large number of hyper-parameters that they expose to the end-user. To maximize the predictive power of GBDT models, one must either manually tune the hyper-parameters, or utilize automated techniques such as those based on Bayesian optimization. Both of these approaches are time-consuming since they involve repeatably training the model for different sets of hyper-parameters. A number of software GBDT packages have started to offer GPU acceleration which can help to alleviate this problem. In this paper, we consider three such packages: XGBoost, LightGBM and Catboost. Firstly, we evaluate the performance of the GPU acceleration provided by these packages using large-scale datasets with varying shapes, sparsities and learning tasks. Then, we compare the packages in the context of hyper-parameter optimization, both in terms of how quickly each package converges to a good validation score, and in terms of generalization performance.

Celestine Dünner, Thomas Parnell, Dimitrios Sarigiannis et al.

We describe a new software framework for fast training of generalized linear models. The framework, named Snap Machine Learning (Snap ML), combines recent advances in machine learning systems and algorithms in a nested manner to reflect the hierarchical architecture of modern computing systems. We prove theoretically that such a hierarchical system can accelerate training in distributed environments where intra-node communication is cheaper than inter-node communication. Additionally, we provide a review of the implementation of Snap ML in terms of GPU acceleration, pipelining, communication patterns and software architecture, highlighting aspects that were critical for achieving high performance. We evaluate the performance of Snap ML in both single-node and multi-node environments, quantifying the benefit of the hierarchical scheme and the data streaming functionality, and comparing with other widely-used machine learning software frameworks. Finally, we present a logistic regression benchmark on the Criteo Terabyte Click Logs dataset and show that Snap ML achieves the same test loss an order of magnitude faster than any of the previously reported results, including those obtained using TensorFlow and scikit-learn.

Kubilay Atasu, Thomas Parnell, Celestine Dünner et al.

The amount of unstructured text-based data is growing every day. Querying, clustering, and classifying this big data requires similarity computations across large sets of documents. Whereas low-complexity similarity metrics are available, attention has been shifting towards more complex methods that achieve a higher accuracy. In particular, the Word Mover's Distance (WMD) method proposed by Kusner et al. is a promising new approach, but its time complexity grows cubically with the number of unique words in the documents. The Relaxed Word Mover's Distance (RWMD) method, again proposed by Kusner et al., reduces the time complexity from qubic to quadratic and results in a limited loss in accuracy compared with WMD. Our work contributes a low-complexity implementation of the RWMD that reduces the average time complexity to linear when operating on large sets of documents. Our linear-complexity RWMD implementation, henceforth referred to as LC-RWMD, maps well onto GPUs and can be efficiently distributed across a cluster of GPUs. Our experiments on real-life datasets demonstrate 1) a performance improvement of two orders of magnitude with respect to our GPU-based distributed implementation of the quadratic RWMD, and 2) a performance improvement of three to four orders of magnitude with respect to our distributed WMD implementation that uses GPU-based RWMD for pruning.

Celestine Dünner, Thomas Parnell, Martin Jaggi

We propose a generic algorithmic building block to accelerate training of machine learning models on heterogeneous compute systems. Our scheme allows to efficiently employ compute accelerators such as GPUs and FPGAs for the training of large-scale machine learning models, when the training data exceeds their memory capacity. Also, it provides adaptivity to any system's memory hierarchy in terms of size and processing speed. Our technique is built upon novel theoretical insights regarding primal-dual coordinate methods, and uses duality gap information to dynamically decide which part of the data should be made available for fast processing. To illustrate the power of our approach we demonstrate its performance for training of generalized linear models on a large-scale dataset exceeding the memory size of a modern GPU, showing an order-of-magnitude speedup over existing approaches.

Thomas Parnell, Celestine Dünner, Kubilay Atasu et al.

In this work we propose an accelerated stochastic learning system for very large-scale applications. Acceleration is achieved by mapping the training algorithm onto massively parallel processors: we demonstrate a parallel, asynchronous GPU implementation of the widely used stochastic coordinate descent/ascent algorithm that can provide up to 35x speed-up over a sequential CPU implementation. In order to train on very large datasets that do not fit inside the memory of a single GPU, we then consider techniques for distributed stochastic learning. We propose a novel method for optimally aggregating model updates from worker nodes when the training data is distributed either by example or by feature. Using this technique, we demonstrate that one can scale out stochastic learning across up to 8 worker nodes without any significant loss of training time. Finally, we combine GPU acceleration with the optimized distributed method to train on a dataset consisting of 200 million training examples and 75 million features. We show by scaling out across 4 GPUs, one can attain a high degree of training accuracy in around 4 seconds: a 20x speed-up in training time compared to a multi-threaded, distributed implementation across 4 CPUs.

Celestine Dünner, Thomas Parnell, Kubilay Atasu et al.

In this paper we explore the performance limits of Apache Spark for machine learning applications. We begin by analyzing the characteristics of a state-of-the-art distributed machine learning algorithm implemented in Spark and compare it to an equivalent reference implementation using the high performance computing framework MPI. We identify critical bottlenecks of the Spark framework and carefully study their implications on the performance of the algorithm. In order to improve Spark performance we then propose a number of practical techniques to alleviate some of its overheads. However, optimizing computational efficiency and framework related overheads is not the only key to performance -- we demonstrate that in order to get the best performance out of any implementation it is necessary to carefully tune the algorithm to the respective trade-off between computation time and communication latency. The optimal trade-off depends on both the properties of the distributed algorithm as well as infrastructure and framework-related characteristics. Finally, we apply these technical and algorithmic optimizations to three different distributed linear machine learning algorithms that have been implemented in Spark. We present results using five large datasets and demonstrate that by using the proposed optimizations, we can achieve a reduction in the performance difference between Spark and MPI from 20x to 2x.

Reinhard Heckel, Michail Vlachos, Thomas Parnell et al.

We consider the problem of generating interpretable recommendations by identifying overlapping co-clusters of clients and products, based only on positive or implicit feedback. Our approach is applicable on very large datasets because it exhibits almost linear complexity in the input examples and the number of co-clusters. We show, both on real industrial data and on publicly available datasets, that the recommendation accuracy of our algorithm is competitive to that of state-of-art matrix factorization techniques. In addition, our technique has the advantage of offering recommendations that are textually and visually interpretable. Finally, we examine how to implement our technique efficiently on Graphical Processing Units (GPUs).