arXiv subjects

Devesh Tiwari

Publications and source records attributed to Devesh Tiwari.

At least 19 recordsLinked to original sources

arXiv2026

Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition

Linear probes can decode safety-relevant concepts such as truthfulness from language-model activations, but probe accuracy may show only decodability, not that the features the probe weights causally drive model behavior. We demonstrate that this gap cannot be closed from the geometry of probe weights alone: the features geometrically aligned with probe direction need not be the ones the model uses, so causal relevance requires intervention. We introduce a feature-level diagnostic that decomposes a deployed True/False probe into sparse-autoencoder (SAE) features, ranks those features by both probe alignment and by gradient sensitivity of the model's behavior, and ablates the resulting shared, probe-only, and random feature sets under a coherence gate. On the truth probe of Buerger et al. (2024) (TTPD), applied in the instructed truth/deception setting of Long et al. (2025) for Gemma2-9B-Instruct, the two rankings overlap only weakly (about 12%, Spearman rho = 0.10), and ablation dissociates them sharply: features the probe shares with the model flip the output far more (up to 27%) than equally sized probe-only (6%) or random (1%) features at full coherence, while probe-only features instead perturb the probe's own readout. The dissociation holds across five seeds and a held-out split, and an activation-aware selection of features flips behavior nearly three times as often as the probe's geometric top features (17.6% vs. 6.1%). In this setting, therefore, the geometric projection of a probe's weight vector alone does not identify the features the model causally uses; however, combining probe information with feature activation statistics recovers substantially more behaviorally causal features, and coherence-gated SAE intervention is needed to separate them from probe readouts.

cs.AI↗

arXiv2026

The Diminishing Returns of Early-Exit Decoding in Modern LLMs

In Large Language Model (LLM) inference, early-exit refers to stopping computation at an intermediate layer once the prediction is sufficiently confident, thereby reducing latency and cost. However, recent LLMs adopt improved pretraining recipes and architectures that reduce layer redundancy, potentially limiting early-exit opportunities. We re-evaluate layer-wise early-exit in modern LLMs and analyze how intermediate representations evolve during training. We introduce a metric to quantify a model's intrinsic suitability for early-exit and propose a benchmark for researchers to explore the potential early-exit benefits on different models and workloads. Our results show a diminishing trend in early-exit effectiveness across newer model generations. We further find that dense transformers generally offer greater early-exit potential than Mixture-of-Experts and State Space Models. In addition, larger models, particularly those with more than 20 billion parameters, and base pretrained models without specialized tuning tend to exhibit higher early-exit potential.

cs.CL↗

arXiv2026

RLHFless: Serverless Computing for Efficient RLHF

Reinforcement Learning from Human Feedback (RLHF) has been widely applied to Large Language Model (LLM) post-training to align model outputs with human preferences. Recent models, such as DeepSeek-R1, have also shown RLHF's potential to improve LLM reasoning on complex tasks. In RL, inference and training co-exist, creating dynamic resource demands throughout the workflow. Compared to traditional RL, RLHF further challenges training efficiency due to expanding model sizes and resource consumption. Several RLHF frameworks aim to balance flexible abstraction and efficient execution. However, they rely on serverful infrastructures, which struggle with fine-grained resource variability. As a result, during synchronous RLHF training, idle time between or within RL components often causes overhead and resource wastage. To address these issues, we present RLHFless, the first scalable training framework for synchronous RLHF, built on serverless computing environments. RLHFless adapts to dynamic resource demands throughout the RLHF pipeline, pre-computes shared prefixes to avoid repeated computation, and uses a cost-aware actor scaling strategy that accounts for response length variation to find sweet spots with lower cost and higher speed. In addition, RLHFless assigns workloads efficiently to reduce intra-function imbalance and idle time. Experiments on both physical testbeds and a large-scale simulated cluster show that RLHFless achieves up to 1.35x speedup and 44.8% cost reduction compared to the state-of-the-art baseline.

cs.AI↗

arXiv2025

ThirstyFLOPS: Water Footprint Modeling and Analysis Toward Sustainable HPC Systems

High-performance computing (HPC) systems are becoming increasingly water-intensive due to their reliance on water-based cooling and the energy used in power generation. However, the water footprint of HPC remains relatively underexplored-especially in contrast to the growing focus on carbon emissions. In this paper, we present ThirstyFLOPS - a comprehensive water footprint analysis framework for HPC systems. Our approach incorporates region-specific metrics, including Water Usage Effectiveness, Power Usage Effectiveness, and Energy Water Factor, to quantify water consumption using real-world data. Using four representative HPC systems - Marconi, Fugaku, Polaris, and Frontier - as examples, we provide implications for HPC system planning and management. We explore the impact of regional water scarcity and nuclear-based energy strategies on HPC sustainability. Our findings aim to advance the development of water-aware, environmentally responsible computing infrastructures.

cs.DC↗

arXiv2025

DarwinGame: Playing Tournaments for Tuning Applications in Noisy Cloud Environments

This work introduces a new subarea of performance tuning -- performance tuning in a shared interference-prone computing environment. We demonstrate that existing tuners are significantly suboptimal by design because of their inability to account for interference during tuning. Our solution, DarwinGame, employs a tournament-based design to systematically compare application executions with different tunable parameter configurations, enabling it to identify the relative performance of different tunable parameter configurations in a noisy environment. Compared to existing solutions, DarwinGame achieves more than 27% reduction in execution time, with less than 0.5% performance variability. DarwinGame is the first performance tuner that will help developers tune their applications in shared, interference-prone, cloud environments.

cs.PF↗

arXiv2025

HotSwap: Enabling Live Dependency Sharing in Serverless Computing

This work presents HotSwap, a novel provider-side cold-start optimization for serverless computing. This optimization reduces cold-start time when booting and loading dependencies at runtime inside a function container. Previous research has extensively focused on reducing cold-start latency for specific functions. However, little attention has been given to skewed production workloads. In such cases, cross-function optimization becomes essential. Without cross-function optimization, a cloud provider is left with two equally poor options: (i) Either the cloud provider gives up optimization for each function in the long tail (which is slow); or (ii) the cloud provider applies function-specific optimizations (e.g., cache function images) to every function in the long tail (which violates the vendor's cache constraints). HotSwap demonstrates cross-function optimization using a novel pre-warming strategy. In this strategy, a pre-initialized live dependency image is migrated to the new function instance. At the same time, HotSwap respects the provider's cache constraints, because a single pre-warmed dependency image in the cache can be shared among all serverless functions that require that image. HotSwap has been tested on seven representative functions from FunctionBench. In those tests, HotSwap accelerates dependency loading for those serverless functions with large dependency requirements by a factor ranging from 2.2 to 3.2. Simulation experiments using Azure traces indicate that HotSwap can save 88\% of space, compared with a previous function-specific method, PreBaking, when sharing a dependency image among ten different functions.

cs.DC↗

arXiv2025

ForgetMeNot: Understanding and Modeling the Impact of Forever Chemicals Toward Sustainable Large-Scale Computing

Fluorinated compounds, often referred to as forever chemicals, are critical in various steps of semiconductor fabrication like lithography, etching, chamber cleaning, and others. Forever chemical emissions can exhibit global warming potentials thousands of times greater than carbon dioxide and persist in the atmosphere for millennia. Despite their severe impact, most sustainability works in computer systems have focused on carbon emissions alone. We address this gap by introducing ForgetMeNot, a modeling tool that quantifies fluorinated compound emissions by integrating fabrication facility-specific practices and hardware specifications, and validate its accuracy using real-world emission data from fabrication facilities. We show how ForgetMeNot can enable fabrication facilities to optimize design and material usage decisions for emission reduction and provide researchers with a methodology to calibrate emission estimates for hardware designs. When ForgetMeNot is applied to analyze emissions for manufacturing CPUs, DRAM, and storage, it illustrates how hardware generations, lithography techniques, and capacities impact fluorinated compound emissions. Finally, we demonstrate how datacenter operators can assemble low-emission servers while balancing performance demands. By factoring in fluorinated emissions into manufacturing decisions, ForgetMeNot paves the way for building more sustainable systems.

cs.ET↗

arXiv2025

WaterWise: Co-optimizing Carbon- and Water-Footprint Toward Environmentally Sustainable Cloud Computing

The carbon and water footprint of large-scale computing systems poses serious environmental sustainability risks. In this study, we discover that, unfortunately, carbon and water sustainability are at odds with each other - and, optimizing one alone hurts the other. Toward that goal, we introduce, WaterWise, a novel job scheduler for parallel workloads that intelligently co-optimizes carbon and water footprint to improve the sustainability of geographically distributed data centers.

cs.DC↗

arXiv2025

A Micromechanical Model for Light-interactive Molecular Crystals

Molecular crystals respond to a light stimulus by bending, twisting, rolling, jumping, or other kinematic behaviors. These behaviors are known to be affected by, among others, the intensity of the incident light, the aspect ratios of crystal geometries, and the volume changes accompanying phase transformation. While these factors, individually, explain the increase in internal energy of the system and its subsequent minimization through macroscopic deformation, they do not fully explain the diversity of deformations observed in molecular crystals. Here, we propose a micromechanical model based on the Cauchy-Born rule and photoreaction theory to predict the macroscopic response in molecular crystals. By accounting for lattice geometry changes and microstructural patterns that emerge during phase transformation, we predict a range of deformations in a representative molecular crystal (salicylideneamine). Doing so, we find that the interplay between photoexcited states and the energy minimization pathways, across a multi-well energy landscape, is crucial to the bending and twisting deformations. We use our model to analyze the role of particle geometries and the intensity of incident light on macroscopic deformation, and identify geometric regimes for shearing and twisting deformations in salicylideneamine crystals. Our micromechanical model is general and can be adapted to predict photomechanical deformation in other molecular crystals undergoing a solid-to-solid phase change and has potential as a computational design tool to engineer reversible and controllable actuation in molecular crystals.

cond-mat.mtrl-sci↗

arXiv2024

EcoLife: Carbon-Aware Serverless Function Scheduling for Sustainable Computing

This work introduces ECOLIFE, the first carbon-aware serverless function scheduler to co-optimize carbon footprint and performance. ECOLIFE builds on the key insight of intelligently exploiting multi-generation hardware to achieve high performance and lower carbon footprint. ECOLIFE designs multiple novel extensions to Particle Swarm Optimization (PSO) in the context of serverless execution environment to achieve high performance while effectively reducing the carbon footprint.

cs.DC↗

arXiv2024

Qompose: A Technique to Select Optimal Algorithm- Specific Layout for Neutral Atom Quantum Architectures

As quantum computing architecture matures, it is important to investigate new technologies that lend unique advantages. In this work, we propose, Qompose, a neutral atom quantum computing framework for efficiently composing quantum circuits on 2-D topologies of neutral atoms. Qompose selects an efficient topology for any given circuit in order to optimize for length of execution through efficient parallelism and for overall fidelity. our extensive evaluation demonstrates the Qompose is effective for a large collection of randomly-generated quantum circuits and a range of real-world benchmarks including VQE, ISING, and QAOA.

quant-ph↗

arXiv2024

OrganiQ: Mitigating Classical Resource Bottlenecks of Quantum Generative Adversarial Networks on NISQ-Era Machines

Driven by swift progress in hardware capabilities, quantum machine learning has emerged as a research area of interest. Recently, quantum image generation has produced promising results. However, prior quantum image generation techniques rely on classical neural networks, limiting their quantum potential and image quality. To overcome this, we introduce OrganiQ, the first quantum GAN capable of producing high-quality images without using classical neural networks.

quant-ph↗

arXiv2024

LLM Inference Serving: Survey of Recent Advances and Opportunities

This survey offers a comprehensive overview of recent advancements in Large Language Model (LLM) serving systems, focusing on research since the year 2023. We specifically examine system-level enhancements that improve performance and efficiency without altering the core LLM decoding mechanisms. By selecting and reviewing high-quality papers from prestigious ML and system venues, we highlight key innovations and practical considerations for deploying and scaling LLMs in real-world production environments. This survey serves as a valuable resource for LLM practitioners seeking to stay abreast of the latest developments in this rapidly evolving field.

cs.DC↗

arXiv2024

Toward Sustainable GenAI using Generation Directives for Carbon-Friendly Large Language Model Inference

The rapid advancement of Generative Artificial Intelligence (GenAI) across diverse sectors raises significant environmental concerns, notably the carbon emissions from their cloud and high performance computing (HPC) infrastructure. This paper presents Sprout, an innovative framework designed to address these concerns by reducing the carbon footprint of generative Large Language Model (LLM) inference services. Sprout leverages the innovative concept of "generation directives" to guide the autoregressive generation process, thereby enhancing carbon efficiency. Our proposed method meticulously balances the need for ecological sustainability with the demand for high-quality generation outcomes. Employing a directive optimizer for the strategic assignment of generation directives to user prompts and an original offline quality evaluator, Sprout demonstrates a significant reduction in carbon emissions by over 40% in real-world evaluations using the Llama2 LLM and global electricity grid data. This research marks a critical step toward aligning AI technology with sustainable practices, highlighting the potential for mitigating environmental impacts in the rapidly expanding domain of generative artificial intelligence.

cs.DC↗

arXiv2024

Sustainable Supercomputing for AI: GPU Power Capping at HPC Scale

As research and deployment of AI grows, the computational burden to support and sustain its progress inevitably does too. To train or fine-tune state-of-the-art models in NLP, computer vision, etc., some form of AI hardware acceleration is virtually a requirement. Recent large language models require considerable resources to train and deploy, resulting in significant energy usage, potential carbon emissions, and massive demand for GPUs and other hardware accelerators. However, this surge carries large implications for energy sustainability at the HPC/datacenter level. In this paper, we study the aggregate effect of power-capping GPUs on GPU temperature and power draw at a research supercomputing center. With the right amount of power-capping, we show significant decreases in both temperature and power draw, reducing power consumption and potentially improving hardware life-span with minimal impact on job performance. While power-capping reduces power draw by design, the aggregate system-wide effect on overall energy consumption is less clear; for instance, if users notice job performance degradation from GPU power-caps, they may request additional GPU-jobs to compensate, negating any energy savings or even worsening energy consumption. To our knowledge, our work is the first to conduct and make available a detailed analysis of the effects of GPU power-capping at the supercomputing scale. We hope our work will inspire HPCs/datacenters to further explore, evaluate, and communicate the impact of power-capping AI hardware accelerators for more sustainable AI.

cs.AR↗

arXiv2024

Autonomy Loops for Monitoring, Operational Data Analytics, Feedback, and Response in HPC Operations

Many High Performance Computing (HPC) facilities have developed and deployed frameworks in support of continuous monitoring and operational data analytics (MODA) to help improve efficiency and throughput. Because of the complexity and scale of systems and workflows and the need for low-latency response to address dynamic circumstances, automated feedback and response have the potential to be more effective than current human-in-the-loop approaches which are laborious and error prone. Progress has been limited, however, by factors such as the lack of infrastructure and feedback hooks, and successful deployment is often site- and case-specific. In this position paper we report on the outcomes and plans from a recent Dagstuhl Seminar, seeking to carve a path for community progress in the development of autonomous feedback loops for MODA, based on the established formalism of similar (MAPE-K) loops in autonomous computing and self-adaptive systems. By defining and developing such loops for significant cases experienced across HPC sites, we seek to extract commonalities and develop conventions that will facilitate interoperability and interchangeability with system hardware, software, and applications across different sites, and will motivate vendors and others to provide telemetry interfaces and feedback hooks to enable community development and pervasive deployment of MODA autonomy loops.

cs.DC↗

arXiv2023

Toward Sustainable HPC: Carbon Footprint Estimation and Environmental Implications of HPC Systems

The rapid growth in demand for HPC systems has led to a rise in carbon footprint, which requires urgent intervention. In this work, we present a comprehensive analysis of the carbon footprint of high-performance computing (HPC) systems, considering the carbon footprint during both the hardware manufacturing and system operational stages. Our work employs HPC hardware component carbon footprint modeling, regional carbon intensity analysis, and experimental characterization of the system life cycle to highlight the importance of quantifying the carbon footprint of HPC systems.

cs.DC↗

arXiv2023

From Words to Watts: Benchmarking the Energy Costs of Large Language Model Inference

Large language models (LLMs) have exploded in popularity due to their new generative capabilities that go far beyond prior state-of-the-art. These technologies are increasingly being leveraged in various domains such as law, finance, and medicine. However, these models carry significant computational challenges, especially the compute and energy costs required for inference. Inference energy costs already receive less attention than the energy costs of training LLMs -- despite how often these large models are called on to conduct inference in reality (e.g., ChatGPT). As these state-of-the-art LLMs see increasing usage and deployment in various domains, a better understanding of their resource utilization is crucial for cost-savings, scaling performance, efficient hardware usage, and optimal inference strategies. In this paper, we describe experiments conducted to study the computational and energy utilization of inference with LLMs. We benchmark and conduct a preliminary analysis of the inference performance and inference energy costs of different sizes of LLaMA -- a recent state-of-the-art LLM -- developed by Meta AI on two generations of popular GPUs (NVIDIA V100 \& A100) and two datasets (Alpaca and GSM8K) to reflect the diverse set of tasks/benchmarks for LLMs in research and practice. We present the results of multi-node, multi-GPU inference using model sharding across up to 32 GPUs. To our knowledge, our work is the one of the first to study LLM inference performance from the perspective of computational and energy resources at this scale.

cs.CL↗