arXiv subjects

Shivam Sahni

Publications and source records attributed to Shivam Sahni.

6 recordsLinked to original sources

arXiv2026

SonicSampler: Unified Tile-Aware Kernels for LLM Sampling and Speculative Verification

Sampling in LLM inference comprises a combinatorial set of logit processing, token selection, and verification operations for speculative decoding. However, existing implementations either accelerate only subsets of this pipeline, rely on multiple kernel launches, or assume homogeneous sampling behavior across a batch, limiting support for dynamic serving workloads and preventing efficient CUDA Graph execution. We present $\textbf{SonicSampler}$, a unified suite of tile-aware Triton kernels that vertically fuses the complete sampling pipeline into a fixed, workload-aware execution model. Our kernels support dynamic per-request sampling behaviors, including grammar-constrained decoding, repetition, frequency and presence penalties, logit bias, temperature scaling, top-$k$ / top-$p$ / min-$p$ filtering, and speculative verification - within a single batched kernel while remaining fully CUDA Graph-compatible. Central to our approach is a novel hierarchical two-stage top-$k$ algorithm that achieves up to $\textbf{10x speedup}$ over competitive baselines and exploits the low-entropy structure of LLM outputs to enable efficient selection over large vocabularies. Across heterogeneous speculative decoding workloads, SonicSampler achieves up to $\textbf{16x speedup}$ over state-of-the-art baselines while preserving flexible batched execution.

cs.AI↗

arXiv2025

Liger Kernel: Efficient Triton Kernels for LLM Training

Training Large Language Models (LLMs) efficiently at scale presents a formidable challenge, driven by their ever-increasing computational demands and the need for enhanced performance. In this work, we introduce Liger-Kernel, an open-sourced set of Triton kernels developed specifically for LLM training. With kernel optimization techniques like kernel operation fusing and input chunking, our kernels achieve on average a 20% increase in training throughput and a 60% reduction in GPU memory usage for popular LLMs compared to HuggingFace implementations. In addition, Liger-Kernel is designed with modularity, accessibility, and adaptability in mind, catering to both casual and expert users. Comprehensive benchmarks and integration tests are built in to ensure compatibility, performance, correctness, and convergence across diverse computing environments and model architectures. The source code is available under a permissive license at: github.com/linkedin/Liger-Kernel.

cs.LG↗

arXiv2023

Exploring Multilingual Text Data Distillation

With the rise of deep learning, large datasets and complex models have become common, requiring significant computing power. To address this, data distillation has emerged as a technique to quickly train models with lower memory and time requirements. However, data distillation on text-based datasets hasn't been explored much because of the challenges rising due to its discrete nature. Additionally, existing dataset distillation methods often struggle to generalize to new architectures. In the paper, we propose several data distillation techniques for multilingual text classification datasets using language-model-based learning methods. We conduct experiments to analyze their performance in terms of classification strength, and cross-architecture generalization. Furthermore, we investigate the language-specific fairness of the data summaries generated by these methods. Our approach builds upon existing techniques, enhancing cross-architecture generalization in the text data distillation domain.

cs.CL↗

arXiv2022

Exploring Explainability Methods for Graph Neural Networks

With the growing use of deep learning methods, particularly graph neural networks, which encode intricate interconnectedness information, for a variety of real tasks, there is a necessity for explainability in such settings. In this paper, we demonstrate the applicability of popular explainability approaches on Graph Attention Networks (GAT) for a graph-based super-pixel image classification task. We assess the qualitative and quantitative performance of these techniques on three different datasets and describe our findings. The results shed a fresh light on the notion of explainability in GNNs, particularly GATs.

cs.LG↗

arXiv2021

Trust as a Metric for Resiliency in Signed Social Networks

Recent technological advancements have resulted in a surge in online trading, raising severe concerns about theft and fraud, especially on platforms like Bitcoin OTC (over-the-counter), where users' identities remain anonymous. To mitigate the risk, it has become essential to capture the reputation of users based on their trade histories. The who-trusts-whom signed network of people has the capability to reflect the nature of such positive and negative relations between the users. It can be used to analyze linkage patterns, strength, and resiliency of such platforms. Due to the dynamic nature of trust between individuals, these trust networks are often vulnerable to link or node failures, making it critical to understand the stability of such systems. In this paper, we consider the problem of quantifying the resiliency of signed networks with the help of trustworthy community structures. We propose a metric for computing the Trustworthiness of a community structure. Using the trustworthiness scores of all communities structures, we generate a pipeline for assessing the resiliency of a signed network. We also show how these generated resiliency scores are concordant with the true nature of the network.

cs.SI↗

arXiv2021

Image Inpainting using Partial Convolution

Image Inpainting is one of the very popular tasks in the field of image processing with broad applications in computer vision. In various practical applications, images are often deteriorated by noise due to the presence of corrupted, lost, or undesirable information. There have been various restoration techniques used in the past with both classical and deep learning approaches for handling such issues. Some traditional methods include image restoration by filling gap pixels using the nearby known pixels or using the moving average over the same. The aim of this paper is to perform image inpainting using robust deep learning methods that use partial convolution layers.

cs.CV↗