arXiv subjects

Rhea Sanjay Sukthanker

Publications and source records attributed to Rhea Sanjay Sukthanker.

10 recordsLinked to original sources

arXiv2026

Sliding-window beats linear attention

Due to the nature of quadratic attention, Large Language Models (LLMs) consume a lot of memory and energy: every new token costs more than the previous one, and its keys and values must be stored in memory indefinitely, which is unsustainable. Two main lines of work address this: compressing the KV cache, e.g., by evicting or quantizing keys and values, and retrofitting LLMs to use Linear Attention, which replaces the KV cache with a fixed-size state. Retrofitting has attracted a lot of attention, given its promise to solve the quadratic scaling problem with state-of-the-art performance at low cost. However, it has not been properly compared to the simplest form of KV-cache eviction: Sliding Window Attention (SWA) with attention sinks. In this work, we show that SWA with sinks performs as well or better than most retrofitted Linear Attention models across multiple LLMs and downstream tasks, with the largest gains on long-context and generative tasks. On long-context reasoning tasks (Needle-in-a-Haystack and BABILong), SWA achieves massively higher performance (2 to 10 times higher than linear attention). SWA requires no additional training, is extremely fast, and requires little memory, making it an extremely cheap and reliable solution. When the training budget is limited, switching to SWA is a much more effective way to reduce inference memory cost than retrofitting linear attention. Linear attention models have shown promise, but they require training from scratch or extensive retrofitting to reap their architectural benefits and come close to SWA.

cs.CL↗

arXiv2026

Where to Begin: Efficient Pretraining via Subnetwork Selection and Distillation

Small Language models (SLMs) offer an efficient and accessible alternative to Large Language Models (LLMs), delivering strong performance while using far fewer resources. We introduce a simple and effective framework for pretraining SLMs that brings together three complementary ideas. First, we identify structurally sparse sub-network initializations that consistently outperform randomly initialized models of similar size under the same compute budget. Second, we use evolutionary search to automatically discover high-quality sub-network initializations, providing better starting points for pretraining. Third, we apply knowledge distillation from larger teacher models to speed up training and improve generalization. Together, these components make SLM pretraining substantially more efficient: our best model, discovered using evolutionary search and initialized with LLM weights, matches the validation perplexity of a comparable Pythia SLM while requiring 5.16x and 1.26x fewer floating point operations for token budgets of 10B and 100B, respectively. We release all code publicly, offering a practical and reproducible path toward cost-efficient small language model development at scale.

cs.CL↗

arXiv2025

Weight-Entanglement Meets Gradient-Based Neural Architecture Search

Weight sharing is a fundamental concept in neural architecture search (NAS), enabling gradient-based methods to explore cell-based architectural spaces significantly faster than traditional black-box approaches. In parallel, weight-entanglement has emerged as a technique for more intricate parameter sharing amongst macro-architectural spaces. Since weight-entanglement is not directly compatible with gradient-based NAS methods, these two paradigms have largely developed independently in parallel sub-communities. This paper aims to bridge the gap between these sub-communities by proposing a novel scheme to adapt gradient-based methods for weight-entangled spaces. This enables us to conduct an in-depth comparative assessment and analysis of the performance of gradient-based NAS in weight-entangled search spaces. Our findings reveal that this integration of weight-entanglement and gradient-based NAS brings forth the various benefits of gradient-based methods, while preserving the memory efficiency of weight-entangled spaces. The code for our work is openly accessible https://github.com/automl/TangleNAS.

cs.LG↗

arXiv2025

Compressing Large Language Models with Automated Sub-Network Search

Large Language Models (LLMs) demonstrate exceptional reasoning abilities, enabling strong generalization across diverse tasks such as commonsense reasoning and instruction following. However, as LLMs scale, inference costs become increasingly prohibitive, accumulating significantly over their life cycle. In this paper we consider model compression for LLMs to reduce model size while improving downstream task performance. We phrase this as a neural architecture search problem that automatically prunes structural components, such as attention heads, neurons, and layers by searching for the Pareto-optimal set of sub-networks balancing between performance and on-device latency. Compared to state-of-the-art structural pruning approaches and fine-tuned smaller sub-networks extracted from the pre-trained model, our method achieves upto 9.85% improvement on average on 11 diverse downstream tasks, while achieving up to 22% improvement of on-device latency.

cs.CL↗

arXiv2025

Multi-objective Differentiable Neural Architecture Search

Pareto front profiling in multi-objective optimization (MOO), i.e., finding a diverse set of Pareto optimal solutions, is challenging, especially with expensive objectives that require training a neural network. Typically, in MOO for neural architecture search (NAS), we aim to balance performance and hardware metrics across devices. Prior NAS approaches simplify this task by incorporating hardware constraints into the objective function, but profiling the Pareto front necessitates a computationally expensive search for each constraint. In this work, we propose a novel NAS algorithm that encodes user preferences to trade-off performance and hardware metrics, yielding representative and diverse architectures across multiple devices in just a single search run. To this end, we parameterize the joint architectural distribution across devices and multiple objectives via a hypernetwork that can be conditioned on hardware features and preference vectors, enabling zero-shot transferability to new devices. Extensive experiments involving up to 19 hardware devices and 3 different objectives demonstrate the effectiveness and scalability of our method. Finally, we show that, without any additional costs, our method outperforms existing MOO NAS methods across a broad range of qualitatively different search spaces and datasets, including MobileNetV3 on ImageNet-1k, an encoder-decoder transformer space for machine translation and a decoder-only space for language modelling.

cs.LG↗

arXiv2024

HW-GPT-Bench: Hardware-Aware Architecture Benchmark for Language Models

The increasing size of language models necessitates a thorough analysis across multiple dimensions to assess trade-offs among crucial hardware metrics such as latency, energy consumption, GPU memory usage, and performance. Identifying optimal model configurations under specific hardware constraints is becoming essential but remains challenging due to the computational load of exhaustive training and evaluation on multiple devices. To address this, we introduce HW-GPT-Bench, a hardware-aware benchmark that utilizes surrogate predictions to approximate various hardware metrics across 13 devices of architectures in the GPT-2 family, with architectures containing up to 1.55B parameters. Our surrogates, via calibrated predictions and reliable uncertainty estimates, faithfully model the heteroscedastic noise inherent in the energy and latency measurements. To estimate perplexity, we employ weight-sharing techniques from Neural Architecture Search (NAS), inheriting pretrained weights from the largest GPT-2 model. Finally, we demonstrate the utility of HW-GPT-Bench by simulating optimization trajectories of various multi-objective optimization algorithms in just a few seconds.

cs.LG↗

arXiv2023

Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face Recognition

Face recognition systems are widely deployed in safety-critical applications, including law enforcement, yet they exhibit bias across a range of socio-demographic dimensions, such as gender and race. Conventional wisdom dictates that model biases arise from biased training data. As a consequence, previous works on bias mitigation largely focused on pre-processing the training data, adding penalties to prevent bias from effecting the model during training, or post-processing predictions to debias them, yet these approaches have shown limited success on hard problems such as face recognition. In our work, we discover that biases are actually inherent to neural network architectures themselves. Following this reframing, we conduct the first neural architecture search for fairness, jointly with a search for hyperparameters. Our search outputs a suite of models which Pareto-dominate all other high-performance architectures and existing bias mitigation methods in terms of accuracy and fairness, often by large margins, on the two most widely used datasets for face identification, CelebA and VGGFace2. Furthermore, these models generalize to other datasets and sensitive attributes. We release our code, models and raw data files at https://github.com/dooleys/FR-NAS.

cs.CV↗

arXiv2022

Generative Flows with Invertible Attentions

Flow-based generative models have shown an excellent ability to explicitly learn the probability density function of data via a sequence of invertible transformations. Yet, learning attentions in generative flows remains understudied, while it has made breakthroughs in other domains. To fill the gap, this paper introduces two types of invertible attention mechanisms, i.e., map-based and transformer-based attentions, for both unconditional and conditional generative flows. The key idea is to exploit a masked scheme of these two attentions to learn long-range data dependencies in the context of generative flows. The masked scheme allows for invertible attention modules with tractable Jacobian determinants, enabling its seamless integration at any positions of the flow-based models. The proposed attention mechanisms lead to more efficient generative flows, due to their capability of modeling the long-term data dependencies. Evaluation on multiple image synthesis tasks shows that the proposed attention flows result in efficient models and compare favorably against the state-of-the-art unconditional and conditional generative flows.

cs.LG↗

arXiv2021

Neural Architecture Search of SPD Manifold Networks

In this paper, we propose a new neural architecture search (NAS) problem of Symmetric Positive Definite (SPD) manifold networks, aiming to automate the design of SPD neural architectures. To address this problem, we first introduce a geometrically rich and diverse SPD neural architecture search space for an efficient SPD cell design. Further, we model our new NAS problem with a one-shot training process of a single supernet. Based on the supernet modeling, we exploit a differentiable NAS algorithm on our relaxed continuous search space for SPD neural architecture search. Statistical evaluation of our method on drone, action, and emotion recognition tasks mostly provides better results than the state-of-the-art SPD networks and traditional NAS algorithms. Empirical results show that our algorithm excels in discovering better performing SPD network design and provides models that are more than three times lighter than searched by the state-of-the-art NAS algorithms.

cs.LG↗

arXiv2021

Trilevel Neural Architecture Search for Efficient Single Image Super-Resolution

Modern solutions to the single image super-resolution (SISR) problem using deep neural networks aim not only at better performance accuracy but also at a lighter and computationally efficient model. To that end, recently, neural architecture search (NAS) approaches have shown some tremendous potential. Following the same underlying, in this paper, we suggest a novel trilevel NAS method that provides a better balance between different efficiency metrics and performance to solve SISR. Unlike available NAS, our search is more complete, and therefore it leads to an efficient, optimized, and compressed architecture. We innovatively introduce a trilevel search space modeling, i.e., hierarchical modeling on network-, cell-, and kernel-level structures. To make the search on trilevel spaces differentiable and efficient, we exploit a new sparsestmax technique that is excellent at generating sparse distributions of individual neural architecture candidates so that they can be better disentangled for the final selection from the enlarged search space. We further introduce the sorting technique to the sparsestmax relaxation for better network-level compression. The proposed NAS optimization additionally facilitates simultaneous search and training in a single phase, reducing search time and train time. Comprehensive evaluations on the benchmark datasets show our method's clear superiority over the state-of-the-art NAS in terms of a good trade-off between model size, performance, and efficiency.

cs.CV↗