Search for: All records

Modern model hubs, such as Hugging Face, store tens of petabytes of LLMs, with fine-tuned variants vastly outnumbering base models and dominating storage consumption. Existing storage reduction techniques---such as deduplication and compression---are either LLM-oblivious or not compatible with each other, limiting data reduction effectiveness. Our large-scale characterization study across all publicly available Hugging Face LLM repositories reveals several key insights: (1) fine-tuned models within the same family exhibit highly structured, sparse parameter differences suitable for delta compression; (2) bitwise similarity enables LLM family clustering; and (3) tensor-level deduplication is better aligned with model storage workloads, achieving high data reduction with low metadata overhead. Building on these insights, we design BitX, an effective, fast, lossless delta compression algorithm that compresses XORed difference between fine-tuned and base LLMs. We build ZipLLM, a model storage reduction pipeline that unifies tensor-level deduplication and lossless BitX compression. By synergizing deduplication and compression around LLM family clustering, ZipLLM reduces model storage consumption by 54%, over 20% higher than state-of-the-art deduplication and compression approaches. 

Interactive notebook programming is universal in modern ML and AI workflows, with interactive deep learning training (IDLT) emerging as a dominant use case. To ensure responsiveness, platforms like Jupyter and Colab reserve GPUs for long-running notebook sessions, despite their intermittent and sporadic GPU usage, leading to extremely low GPU utilization and prohibitively high costs. In this paper, we introduce NotebookOS, a GPU-efficient notebook platform tailored for the unique requirements of IDLT. NotebookOS employs replicated notebook kernels with Raft-synchronized replicas distributed across GPU servers. To optimize GPU utilization, NotebookOS oversubscribes server resources, leveraging high inter-arrival times in IDLT workloads, and allocates GPUs only during active cell execution. It also supports replica migration and automatic cluster scaling under high load. Altogether, this design enables interactive training with minimal delay. In evaluation on production workloads, NotebookOS saved over 1,187 GPU hours in 17.5 hours of real-world IDLT, while significantly improving interactivity. 

A series of perovskite oxides (Ln = La, Pr, Nd, Gd; A = Ba, Sr) was investigated to understand the effects of A-site cation size on oxygen vacancy formation. Quasirandom mixed structures were generated using Alloy Theoretic Automated Toolkit (ATAT), followed by density functional theory (DFT) calculations. While mixing the orthorhombic structures with the hexagonal AMnO3 structures leads to lattices and global symmetries closer to cubic, the average volume generally increases with the average ionic size, and the local bond and angles exhibit more variations due to A-site mixing. DFT calculations and a statistical model were combined to predict oxygen reduction abilities. Thermogravimetric analysis (TGA) provided experimental validation of these predictions by measuring changes in oxygen non-stoichiometry under controlled conditions. Both indicated that larger A-site ionic size differences lead to greater, consistent with the larger variation in local structures, and enhanced redox capabilities. This combined computational-experimental approach highlights the importance of local structure variation, instead of average properties, in A-site cation engineering to optimize perovskite oxides for different devices relying on oxygen vacancy redox activity. 

Understanding microbial interactions is fundamental for exploring population dynamics, particularly in microbial communities where interactions affect stability and host health. Generalized Lotka-Volterra (gLV) models have been widely used to investigate system dynamics but depend on absolute abundance data, which are often unavailable in microbiome studies. To address this limitation, we introduce an iterative Lotka-Volterra (iLV) model, a novel framework tailored for compositional data that leverages relative abundances and iterative refinements for parameter estimation. The iLV model features two key innovations: an adaptation of the gLV framework to compositional constraints and an iterative optimization strategy combining linear approximations with nonlinear refinements to enhance parameter estimation accuracy. Using simulations and real-world datasets, we demonstrate that iLV surpasses existing methodologies, such as the compositional LV (cLV) and the generalized LV (gLV) model, in recovering interaction coefficients and predicting species trajectories under varying noise levels and temporal resolutions. Applications to the lynx-hare predator-prey,Stylonychia pustula-P. caudatummixed culture, and cheese microbial systems revealed consistency between predicted and observed relative abundances showcasing its accuracy and robustness. In summary, the iLV model bridges theoretical gLV models and practical compositional data analysis, offering a robust framework to infer microbial interactions and predict community dynamics using relative abundance data, with significant potential for advancing microbial research.