Fine-tuning large language models (LLMs) is increasingly costly as models scale to hundreds of billions of parameters, and even parameter-efficient fine-tuning (PEFT) methods like LoRA remain resource-intensive. We introduce LowRA, the first framework to enable LoRA fine-tuning below 2 bits per parameter with minimal performance loss. LowRA optimizes fine-grained quantization - mapping, threshold selection, and precision assignment - while leveraging efficient CUDA kernels for scalable deployment. Extensive evaluations across 4 LLMs and 4 datasets show that LowRA achieves a superior performance-precision trade-off above 2 bits and remains accurate down to 1.15 bits, reducing memory usage by up to 50%. Our results highlight the potential of ultra-low-bit LoRA fine-tuning for resource-constrained environments.

ML libraries, often written in architecture-specific programming languages (ASPLs) that target domain-specific architectures, are key to efficient ML systems. However, writing these high-performance ML libraries is challenging because it requires expert knowledge of ML algorithms and the ASPL. Large language models (LLMs), on the other hand, have shown general coding capabilities. However, challenges remain when using LLMs for generating ML libraries using ASPLs because 1) this task is complicated even for experienced human programmers and 2) there are limited code examples because of the esoteric and evolving nature of ASPLs. Therefore, LLMs need complex reasoning with limited data in order to complete this task. To address these challenges, we introduce an adaptive self-improvement agentic system. In order to evaluate the effectiveness of our system, we construct a benchmark of a typical ML library and generate ASPL code with both open and closed-source LLMs on this benchmark. Our results show improvements of up to $3.9\times$ over a baseline single LLM.

Monolithic large language models (LLMs) like GPT-4 have paved the way for modern generative AI applications. Training, serving, and maintaining monolithic LLMs at scale, however, remains prohibitively expensive and challenging. The disproportionate increase in compute-to-memory ratio of modern AI accelerators have created a memory wall, necessitating new methods to deploy AI. Composition of Experts (CoE) is an alternative modular approach that lowers the cost and complexity of training and serving. However, this approach presents two key challenges when using conventional hardware: (1) without fused operations, smaller models have lower operational intensity, which makes high utilization more challenging to achieve; and (2) hosting a large number of models can be either prohibitively expensive or slow when dynamically switching between them. In this paper, we describe how combining CoE, streaming dataflow, and a three-tier memory system scales the AI memory wall. We describe Samba-CoE, a CoE system with 150 experts and a trillion total parameters. We deploy Samba-CoE on the SambaNova SN40L Reconfigurable Dataflow Unit (RDU) - a commercial dataflow accelerator architecture that has been co-designed for enterprise inference and training applications. The chip introduces a new three-tier memory system with on-chip distributed SRAM, on-package HBM, and off-package DDR DRAM. A dedicated inter-RDU network enables scaling up and out over multiple sockets. We demonstrate speedups ranging from 2x to 13x on various benchmarks running on eight RDU sockets compared with an unfused baseline. We show that for CoE inference deployments, the 8-socket RDU Node reduces machine footprint by up to 19x, speeds up model switching time by 15x to 31x, and achieves an overall speedup of 3.7x over a DGX H100 and 6.6x over a DGX A100.

We introduce the Bayesian Compiler Optimization framework (BaCO), a general purpose autotuner for modern compilers targeting CPUs, GPUs, and FPGAs. BaCO provides the flexibility needed to handle the requirements of modern autotuning tasks. Particularly, it deals with permutation, ordered, and continuous parameter types along with both known and unknown parameter constraints. To reason about these parameter types and efficiently deliver high-quality code, BaCO uses Bayesian optimiza tion algorithms specialized towards the autotuning domain. We demonstrate BaCO's effectiveness on three modern compiler systems: TACO, RISE & ELEVATE, and HPVM2FPGA for CPUs, GPUs, and FPGAs respectively. For these domains, BaCO outperforms current state-of-the-art autotuners by delivering on average 1.36x-1.56x faster code with a tiny search budget, and BaCO is able to reach expert-level performance 2.9x-3.9x faster.

While Bayesian Optimization (BO) is a very popular method for optimizing expensive black-box functions, it fails to leverage the experience of domain experts. This causes BO to waste function evaluations on bad design choices (e.g., machine learning hyperparameters) that the expert already knows to work poorly. To address this issue, we introduce Prior-guided Bayesian Optimization (PrBO). PrBO allows users to inject their knowledge into the optimization process in the form of priors about which parts of the input space will yield the best performance, rather than BO's standard priors over functions (which are much less intuitive for users). PrBO then combines these priors with BO's standard probabilistic model to form a pseudo-posterior used to select which points to evaluate next. We show that PrBO is around 12x faster than state-of-the-art methods without user priors and 10,000x faster than random search on a common suite of benchmarks, and achieves a new state-of-the-art performance on a real-world hardware design application. We also show that PrBO converges faster even if the user priors are not entirely accurate and that it robustly recovers from misleading priors.

Machine learning (ML) techniques are enjoying rapidly increasing adoption. However, designing and implementing the systems that support ML models in real-world deployments remains a significant obstacle, in large part due to the radically different development and deployment profile of modern ML methods, and the range of practical concerns that come with broader adoption. We propose to foster a new systems machine learning research community at the intersection of the traditional systems and ML communities, focused on topics such as hardware systems for ML, software systems for ML, and ML optimized for metrics beyond predictive accuracy. To do this, we describe a new conference, SysML, that explicitly targets research at the intersection of systems and machine learning with a program committee split evenly between experts in systems and ML, and an explicit focus on topics at the intersection of the two.

Serial crystallography is the field of science that studies the structure and properties of crystals via diffraction patterns. In this paper, we introduce a new serial crystallography dataset comprised of real and synthetic images; the synthetic images are generated through the use of a simulator that is both scalable and accurate. The resulting dataset is called DiffraNet, and it is composed of 25,457 512x512 grayscale labeled images. We explore several computer vision approaches for classification on DiffraNet such as standard feature extraction algorithms associated with Random Forests and Support Vector Machines but also an end-to-end CNN topology dubbed DeepFreak tailored to work on this new dataset. All implementations are publicly available and have been fine-tuned using off-the-shelf AutoML optimization tools for a fair comparison. Our best model achieves 98.5% accuracy on synthetic images and 94.51% accuracy on real images. We believe that the DiffraNet dataset and its classification methods will have in the long term a positive impact in accelerating discoveries in many disciplines, including chemistry, geology, biology, materials science, metallurgy, and physics.

Multi-objective optimization is a crucial matter in computer systems design space exploration because real-world applications often rely on a trade-off between several objectives. Derivatives are usually not available or impractical to compute and the feasibility of an experiment can not always be determined in advance. These problems are particularly difficult when the feasible region is relatively small, and it may be prohibitive to even find a feasible experiment, let alone an optimal one. We introduce a new methodology and corresponding software framework, HyperMapper 2.0, which handles multi-objective optimization, unknown feasibility constraints, and categorical/ordinal variables. This new methodology also supports injection of user prior knowledge in the search when available. All of these features are common requirements in computer systems but rarely exposed in existing design space exploration systems. The proposed methodology follows a white-box model which is simple to understand and interpret (unlike, for example, neural networks) and can be used by the user to better understand the results of the automatic search. We apply and evaluate the new methodology to automatic static tuning of hardware accelerators within the recently introduced Spatial programming language, with minimization of design runtime and compute logic under the constraint of the design fitting in a target field programmable gate array chip. Our results show that HyperMapper 2.0 provides better Pareto fronts compared to state-of-the-art baselines, with better or competitive hypervolume indicator and with 8x improvement in sampling budget for most of the benchmarks explored.

The deep learning community has proposed optimizations spanning hardware, software, and learning theory to improve the computational performance of deep learning workloads. While some of these optimizations perform the same operations faster (e.g., switching from a NVIDIA K80 to P100), many modify the semantics of the training procedure (e.g., large minibatch training, reduced precision), which can impact a model's generalization ability. Due to a lack of standard evaluation criteria that considers these trade-offs, it has become increasingly difficult to compare these different advances. To address this shortcoming, DAWNBENCH and the upcoming MLPERF benchmarks use time-to-accuracy as the primary metric for evaluation, with the accuracy threshold set close to state-of-the-art and measured on a held-out dataset not used in training; the goal is to train to this accuracy threshold as fast as possible. In DAWNBENCH , the winning entries improved time-to-accuracy on ImageNet by two orders of magnitude over the seed entries. Despite this progress, it is unclear how sensitive time-to-accuracy is to the chosen threshold as well as the variance between independent training runs, and how well models optimized for time-to-accuracy generalize. In this paper, we provide evidence to suggest that time-to-accuracy has a low coefficient of variance and that the models tuned for it generalize nearly as well as pre-trained models. We additionally analyze the winning entries to understand the source of these speedups, and give recommendations for future benchmarking efforts.

Low-precision computation is often used to lower the time and energy cost of machine learning, and recently hardware accelerators have been developed to support it. Still, it has been used primarily for inference - not training. Previous low-precision training algorithms suffered from a fundamental tradeoff: as the number of bits of precision is lowered, quantization noise is added to the model, which limits statistical accuracy. To address this issue, we describe a simple low-precision stochastic gradient descent variant called HALP. HALP converges at the same theoretical rate as full-precision algorithms despite the noise introduced by using low precision throughout execution. The key idea is to use SVRG to reduce gradient variance, and to combine this with a novel technique called bit centering to reduce quantization error. We show that on the CPU, HALP can run up to $4 \times$ faster than full-precision SVRG and can match its convergence trajectory. We implemented HALP in TensorQuant, and show that it exceeds the validation performance of plain low-precision SGD on two deep learning tasks.