arXiv subjects

Prashant Pandey

Publications and source records attributed to Prashant Pandey.

At least 19 recordsLinked to original sources

arXiv2026

InferScale: GPU-Native KV Injection for Personalized LLM Serving

Large language models are increasingly deployed with persistent personalized context, such as accumulated memory profiles or long conversation histories, that is shared across a user's many requests. Production memory systems (e.g., Mem0, MemGPT, and Zep) retrieve a relevant subset of this memory and inject it into the prompt, forcing the serving engine to repeatedly prefill the same content. As the retrieval budget grows, time-to-first-token (TTFT) increases even though the underlying memory is reused across requests. We present InferScale, a GPU-native LLM memory system that replaces repeated prompt prefilling with reusable KV state. InferScale precomputes each memory fact's KV representation, stores it alongside a semantic embedding on the GPU, retrieves relevant facts at serving time, and injects their KV directly into vLLM's paged cache. To support dynamically assembled memories under rotary position embeddings, we introduce Chunked RoPE, which stores keys before rotation and applies their serving-time positions during injection. However, encoding memory facts independently omits the cross-fact context available during joint prefilling. We mitigate this with Context-Window Encoding, which encodes each memory fact together with a small window of preceding conversation context while caching only the target fact's KV. InferScale is implemented through vLLM's KV-connector interface, requiring neither engine modifications nor model fine-tuning. Across three open-weight models on LoCoMo, InferScale keeps TTFT nearly constant as the retrieval budget increases: at k=50 it reduces TTFT by 72-79% (3.6-4.8x), achieves 60.3% accuracy versus 63.3% for Mem0 without serving-time recomputation, and delivers 3.7-4.5x the throughput under concurrent load. Reusable KV state thus decouples memory-conditioned serving latency from retrieved-context size while preserving application quality.

cs.DC↗

arXiv2026

Continual Segmentation under Joint Nonstationarity

Evolving data streams induce joint nonstationarity in continual semantic segmentation, where semantic classes, input distributions, and supervision availability change simultaneously over time. This setting reflects practical structured prediction systems, yet remains largely unexplored in prior continual learning work, which typically studies these factors in isolation. We formalize continual segmentation under coupled class, domain, and label shifts and investigate learning in heterogeneous dense prediction environments with limited annotations and abundant unlabeled data. To address instability and overfitting arising from few-shot supervision under distribution drift, we introduce gradient-adaptive stabilization, a parameter-wise regularization mechanism implemented via gradient-scaled stochastic perturbations that promotes a principled stability-plasticity tradeoff. We further leverage unlabeled data through semi-supervised learning and introduce prototype anchored supervision that validates pseudo-labels via joint confidence and prototype consistency. Together, these mechanisms enable learning under joint nonstationarity in continual segmentation. Extensive empirical evaluation across class-incremental, domain-incremental, and few-shot regimes demonstrates consistent improvements over prior methods in heterogeneous structured prediction settings. Our results expose fundamental failure modes of existing continual segmentation approaches and provide insight into learning robust dense predictors in dynamically evolving environments.

cs.CV↗

arXiv2026

How to Train Your Filter: Should You Learn, Stack or Adapt?

Filters are ubiquitous in computer science, enabling space-efficient approximate membership testing. Since Bloom filters were introduced in 1970, decades of work improved their space efficiency and performance. Recently, three new paradigms have emerged offering orders-of-magnitude improvements in false positive rates (FPRs) by using information beyond the input set: (1) learned filters train a model to distinguish (non)members, (2) stacked filters use negative workload samples to build cascading layers, and (3) adaptive filters update internal representation in response to false positive feedback. Yet each paradigm targets specific use cases, introduces complex configuration tuning, and has been evaluated in isolation. This results in unclear trade-offs and a gap in understanding how these approaches compare and when each is most appropriate. This paper presents the first comprehensive evaluation of learned, stacked, and adaptive filters across real-world datasets and query workloads. Our results reveal critical trade-offs: (1) Learned filters achieve up to 10^2 times lower FPRs but exhibit high variance and lack robustness under skewed or dynamic workloads. Critically, model inference overhead leads to up to 10^4 times slower query latencies than stacked or adaptive filters. (2) Stacked filters reliably achieve up to 10^3 times lower FPRs on skewed workloads but require workload knowledge. (3) Adaptive filters are robust across settings, achieving up to 10^3 times lower FPRs under adversarial queries without workload assumptions. Based on our analysis, learned filters suit stable workloads where input features enable effective model training and space constraints are paramount, stacked filters excel when reliable query distributions are known, and adaptive filters are most generalizable, providing robust theoretically bound guarantees even in dynamic or adversarial environments.

cs.DS↗

arXiv2026

GPU-Accelerated ANNS: Quantized for Speed, Built for Change

Approximate nearest neighbor search (ANNS) is a core problem in machine learning and information retrieval applications. GPUs offer a promising path to high-performance ANNS: they provide massive parallelism for distance computations, are readily available, and can co-locate with downstream applications. Despite these advantages, current GPU-accelerated ANNS systems face three key limitations. First, real-world applications operate on evolving datasets that require fast batch updates, yet most GPU indices must be rebuilt from scratch when new data arrives. Second, high-dimensional vectors strain memory bandwidth, but current GPU systems lack efficient quantization techniques that reduce data movement without introducing costly random memory accesses. Third, the data-dependent memory accesses inherent to greedy search make overlapping compute and memory difficult, leading to reduced performance. We present Jasper, a GPU-native ANNS system with both high query throughput and updatability. Jasper builds on the Vamana graph index and overcomes existing bottlenecks via three contributions: (1) a CUDA batch-parallel construction algorithm that enables lock-free streaming insertions, (2) a GPU-efficient implementation of RaBitQ quantization that reduces memory footprint up to 8x without the random access penalties, and (3) an optimized greedy search kernel that increases compute utilization, resulting in better latency hiding and higher throughput. Our evaluation across five datasets shows that Jasper achieves up to 1.84x higher query throughput than CAGRA and achieves up to 80% peak utilization as measured by the roofline model. Jasper's construction scales efficiently and constructs indices an average of 7x faster than CAGRA while providing updatability that CAGRA lacks. Compared to BANG, the previous fastest GPU Vamana implementation, Jasper delivers 10-74x faster queries.

cs.DB↗

arXiv2025

KV Cache Recycling to Expand Usable Context Capacity in Low Parameter LLMs

Whether attention key value (KV) states computed for one prompt for a small LLM can be reused to accelerate inference on a new similar prompt, giving an increase to the space to its context memory using an approach called token recycling. Using a standard Hugging Face setup with DialoGPT-medium (a 345M parameter GPT-2 style decoder trained on 147M Reddit exchanges, 2005 to 2017) as the testbed, we build a cache of past activations and get entries by sentence embeddings, then reuse cached past key values when the cached prompt is an exact prefix of the new input. We compare recycled vs. baseline runs on latency and output fidelity, and log reuse depth in tokens. Reproducibility requires no model modifications, cached KVs are serialized to the CPU, reloaded, and supplied to the generate function to continue decoding from the cached prefix. In tests, we observe consistent speedups when prefix overlap exists, with no material degradation in output semantics, and when overlap is absent, behavior matches baseline.

cs.LG↗

arXiv2025

Time To Replace Your Filter: How Maplets Simplify System Design

Filters such as Bloom, quotient, and cuckoo filters are fundamental building blocks providing space-efficient approximate set membership testing. However, many applications need to associate small values with keys-functionality that filters do not provide. This mismatch forces complex workarounds that degrade performance. We argue that maplets-space-efficient data structures for approximate key-value mappings-are the right abstraction. A maplet provides the same space benefits as filters while natively supporting key-value associations with one-sided error guarantees. Through detailed case studies of SplinterDB (LSM-based key-value store), Squeakr (k-mer counter), and Mantis (genomic sequence search), we identify the common patterns and demonstrate how a unified maplet abstraction can lead to simpler designs and better performance. We conclude that applications benefit from defaulting to maplets rather than filters across domains including databases, computational biology, and networking.

cs.DS↗

arXiv2025

WarpSpeed: A High-Performance Library for Concurrent GPU Hash Tables

GPU hash tables are increasingly used to accelerate data processing, but their limited functionality restricts adoption in large-scale data processing applications. Current limitations include incomplete concurrency support and missing compound operations such as upserts. This paper presents WarpSpeed, a library of high-performance concurrent GPU hash tables with a unified benchmarking framework for performance analysis. WarpSpeed implements eight state-of-the-art Nvidia GPU hash table designs and provides a rich API designed for modern GPU applications. Our evaluation uses diverse benchmarks to assess both correctness and scalability, and we demonstrate real-world impact by integrating these hash tables into three downstream applications. We propose several optimization techniques to reduce concurrency overhead, including fingerprint-based metadata to minimize cache line probes and specialized Nvidia GPU instructions for lock-free queries. Our findings provide new insights into concurrent GPU hash table design and offer practical guidance for developing efficient, scalable data structures on modern GPUs.

cs.DC↗

arXiv2025

Bridging Cache-Friendliness and Concurrency: A Locality-Optimized In-Memory B-Skiplist

Skiplists are widely used for in-memory indexing in many key-value stores, such as RocksDB and LevelDB, due to their ease of implementation and simple concurrency control mechanisms. However, traditional skiplists suffer from poor cache locality, as they store only a single element per node, leaving performance on the table. Minimizing last-level cache misses is key to maximizing in-memory index performance, making high cache locality essential. In this paper, we present a practical concurrent B-skiplist that enhances cache locality and performance while preserving the simplicity of traditional skiplist structures and concurrency control schemes. Our key contributions include a top-down, single-pass insertion algorithm for B-skiplists and a corresponding simple and efficient top-down concurrency control scheme. On 128 threads, the proposed concurrent B-skiplist achieves between 2x-9x higher throughput compared to state-of-the-art concurrent skiplist implementations, including Facebook's concurrent skiplist from Folly and the Java ConcurrentSkipListMap. Furthermore, we find that the B-skiplist achieves competitive (0.9x-1.7x) throughput on point workloads compared to state-of-the-art cache-optimized tree-based indices (e.g., Masstree). For a more complete picture of the performance, we also measure the latency of skiplist and tree-based indices and find that the B-skiplist achieves between 3.5x-103x lower 99% latency compared to other concurrent skiplists and between 0.85x-64x lower 99% latency compared to tree-based indices on point workloads with inserts.

cs.DC↗

arXiv2024

Evaluating Learned Indexes for External-Memory Joins

Joins are among the most time-consuming and data-intensive operations in relational query processing. Much research effort has been applied to the optimization of join processing due to their frequent execution. Recent studies have shown that CDF-based learned models can create smaller and faster indexes, accelerating in-memory joins. However, their effectiveness for external-memory joins, which are crucial for large-scale databases, remains underexplored. This paper evaluates the impact of learned indexes on external-memory joins for both sorted and unsorted data. We compare learned index-based joins against traditional join methods such as hash joins, sort joins, and indexed nested-loop joins on real-world and simulated datasets. Additionally, we analyze learned index-based joins across multiple dimensions, including storage device types, data sorting, parallelism, constrained memory environments, and varying model error. The detailed evaluation enables us to determine the most appropriate learned index to employ for external-memory joins. Our experiments reveal that, unlike in-memory settings, learned indexes in external-memory joins can trade off accuracy for space without significantly degrading performance. While learned indexes provide smaller index sizes and faster lookups, they perform similarly to B-trees in external-memory joins since the total amount of I/O, which dominates runtime, remains unchanged. Additionally, the construction times of learned indexes are approximately $1000\times$ longer, and although they are $2-4\times$ smaller than the internal nodes of a B-tree, these nodes only represent $0.4%-1%$ of the data size and typically fit in main memory.

cs.DB↗

arXiv2024

BYO: A Unified Framework for Benchmarking Large-Scale Graph Containers

A fundamental building block in any graph algorithm is a graph container - a data structure used to represent the graph. Ideally, a graph container enables efficient access to the underlying graph, has low space usage, and supports updating the graph efficiently. In this paper, we conduct an extensive empirical evaluation of graph containers designed to support running algorithms on large graphs. To our knowledge, this is the first apples-to-apples comparison of graph containers rather than overall systems, which include confounding factors such as differences in algorithm implementations and infrastructure. We measure the running time of 10 highly-optimized algorithms across over 20 different containers and 10 graphs. Somewhat surprisingly, we find that the average algorithm running time does not differ much across containers, especially those that support dynamic updates. Specifically, a simple container based on an off-the-shelf B-tree is only 1.22x slower on average than a highly optimized static one. Moreover, we observe that simplifying a graph-container Application Programming Interface (API) to only a few simple functions incurs a mere 1.16x slowdown compared to a complete API. Finally, we also measure batch-insert throughput in dynamic-graph containers for a full picture of their performance. To perform the benchmarks, we introduce BYO, a unified framework that standardizes evaluations of graph-algorithm performance across different graph containers. BYO extends the Graph Based Benchmark Suite (Dhulipala et al. 18), a state-of-the-art graph algorithm benchmark, to easily plug into different dynamic graph containers and enable fair comparisons between them on a large suite of graph algorithms. While several graph algorithm benchmarks have been developed to date, to the best of our knowledge, BYO is the first system designed to benchmark graph containers

cs.DS↗

arXiv2024

Adaptive Quotient Filters

Adaptive filters, such as telescoping and adaptive cuckoo filters, update their representation upon detecting a false positive to avoid repeating the same error in the future. Adaptive filters require an auxiliary structure, typically much larger than the main filter and often residing on slow storage, to facilitate adaptation. However, existing adaptive filters are not practical and have seen no adoption in real-world systems due to two main reasons. Firstly, they offer weak adaptivity guarantees, meaning that fixing a new false positive can cause a previously fixed false positive to come back. Secondly, the sub-optimal design of the auxiliary structure results in adaptivity overheads so substantial that they can actually diminish the overall system performance compared to a traditional filter. In this paper, we design and implement AdaptiveQF, the first practical adaptive filter with minimal adaptivity overhead and strong adaptivity guarantees, which means that the performance and false-positive guarantees continue to hold even for adversarial workloads. The AdaptiveQF is based on the state-of-the-art quotient filter design and preserves all the critical features of the quotient filter such as cache efficiency and mergeability. Furthermore, we employ a new auxiliary structure design which results in considerably low adaptivity overhead and makes the AdaptiveQF practical in real systems.

cs.DS↗

arXiv2023

A Language-Guided Benchmark for Weakly Supervised Open Vocabulary Semantic Segmentation

Increasing attention is being diverted to data-efficient problem settings like Open Vocabulary Semantic Segmentation (OVSS) which deals with segmenting an arbitrary object that may or may not be seen during training. The closest standard problems related to OVSS are Zero-Shot and Few-Shot Segmentation (ZSS, FSS) and their Cross-dataset variants where zero to few annotations are needed to segment novel classes. The existing FSS and ZSS methods utilize fully supervised pixel-labelled seen classes to segment unseen classes. Pixel-level labels are hard to obtain, and using weak supervision in the form of inexpensive image-level labels is often more practical. To this end, we propose a novel unified weakly supervised OVSS pipeline that can perform ZSS, FSS and Cross-dataset segmentation on novel classes without using pixel-level labels for either the base (seen) or the novel (unseen) classes in an inductive setting. We propose Weakly-Supervised Language-Guided Segmentation Network (WLSegNet), a novel language-guided segmentation pipeline that i) learns generalizable context vectors with batch aggregates (mean) to map class prompts to image features using frozen CLIP (a vision-language model) and ii) decouples weak ZSS/FSS into weak semantic segmentation and Zero-Shot segmentation. The learned context vectors avoid overfitting on seen classes during training and transfer better to novel classes during testing. WLSegNet avoids fine-tuning and the use of external datasets during training. The proposed pipeline beats existing methods for weak generalized Zero-Shot and weak Few-Shot semantic segmentation by 39 and 3 mIOU points respectively on PASCAL VOC and weak Few-Shot semantic segmentation by 5 mIOU points on MS COCO. On a harder setting of 2-way 1-shot weak FSS, WLSegNet beats the baselines by 13 and 22 mIOU points on PASCAL VOC and MS COCO, respectively.

cs.CV↗

arXiv2022

High-Performance Filters For GPUs

Filters approximately store a set of items while trading off accuracy for space-efficiency and can address the limited memory on accelerators, such as GPUs. However, there is a lack of high-performance and feature-rich GPU filters as most advancements in filter research has focused on CPUs. In this paper, we explore the design space of filters with a goal to develop massively parallel, high performance, and feature rich filters for GPUs. We evaluate various filter designs in terms of performance, usability, and supported features and identify two filter designs that offer the right trade off in terms of performance, features, and usability. We present two new GPU-based filters, the TCF and GQF, that can be employed in various high performance data analytics applications. The TCF is a set membership filter and supports faster inserts and queries, whereas the GQF supports counting which comes at an additional performance cost. Both the GQF and TCF provide point and bulk insertion API and are designed to exploit the massive parallelism in the GPU without sacrificing usability and necessary features. The TCF and GQF are up to $4.4\times$ and $1.4\times$ faster than the previous GPU filters in our benchmarks and at the same time overcome the fundamental constraints in performance and usability in current GPU filters.

cs.DC↗

arXiv2022

Injectable Bubbles for Physiological Pressure Measurement

Microbubbles - used as contrast agents in ultrasound imaging - are important tools in biomedical research, having been used together with ultrasound to develop significant diagnostic and therapeutic techniques. It has been suggested that the dynamic behaviour of microbubbles is dependent on the surrounding fluid's ambient (hydrostatic) pressure, and the potential to non-invasively determine blood pressure has numerous medical applications. To study this dependence, a computational mathematical model was created based on Marmottant's dynamic model of a microbubble. A pulse inversion (PI) protocol was incorporated into the model to emphasize the nonlinear behaviour of the microbubble's response. The mathematical model was also used to assess the sensitivity of the microbubbles to undesirable changes in parameters other than the ambient pressure. It found that a variation in the microbubble's initial radius and surface tension would cause the most significant changes in signal energy and hence pose a risk to ambient pressure measurements. To test the practicality of detecting a change in the dynamic behaviour of microbubbles, in vitro experiments were designed and carried out using clinically available contrast agent with two different ultrasound systems. The experiments, while possessing certain limitations, confirmed that there is a change in microbubbles' dynamic behaviour when the ambient pressure is varied, in cases by as little as 7.36 mmHg (981 Pa - a 0.98% change in atmospheric pressure). The experimental results establish a proof-of-principle that future experimental work can build upon to verify the mathematical model, and hence aid in developing a non-invasive blood pressure measurement procedure.

physics.med-ph↗

arXiv2022

IcebergHT: High Performance PMEM Hash Tables Through Stability and Low Associativity

Modern hash table designs strive to minimize space while maximizing speed. The most important factor in speed is the number of cache lines accessed during updates and queries. This is especially important on PMEM, which is slower than DRAM and in which writes are more expensive than reads. This paper proposes two stronger design objectives: stability and low-associativity. A stable hash table doesn't move items around, and a hash table has low associativity if there are only a few locations where an item can be stored. Low associativity ensures that queries need to examine only a few memory locations, and stability ensures that insertions write to very few cache lines. Stability also simplifies scaling and crash safety. We present IcebergHT, a fast, crash-safe, concurrent, and space-efficient hash table for PMEM based on the design principles of stability and low associativity. IcebergHT combines in-memory metadata with a new hashing technique, iceberg hashing, that is (1) space efficient, (2) stable, and (3) supports low associativity. In contrast, existing hash-tables either modify numerous cache lines during insertions (e.g. cuckoo hashing), access numerous cache lines during queries (e.g. linear probing), or waste space (e.g. chaining). Moreover, the combination of (1)-(3) yields several emergent benefits: IcebergHT scales better than other hash tables, supports crash-safety, and has excellent performance on PMEM (where writes are particularly expensive).

cs.DS↗

arXiv2022

Adversarially Robust Prototypical Few-shot Segmentation with Neural-ODEs

Few-shot Learning (FSL) methods are being adopted in settings where data is not abundantly available. This is especially seen in medical domains where the annotations are expensive to obtain. Deep Neural Networks have been shown to be vulnerable to adversarial attacks. This is even more severe in the case of FSL due to the lack of a large number of training examples. In this paper, we provide a framework to make few-shot segmentation models adversarially robust in the medical domain where such attacks can severely impact the decisions made by clinicians who use them. We propose a novel robust few-shot segmentation framework, Prototypical Neural Ordinary Differential Equation (PNODE), that provides defense against gradient-based adversarial attacks. We show that our framework is more robust compared to traditional adversarial defense mechanisms such as adversarial training. Adversarial training involves increased training time and shows robustness to limited types of attacks depending on the type of adversarial examples seen during training. Our proposed framework generalises well to common adversarial attacks like FGSM, PGD and SMIA while having the model parameters comparable to the existing few-shot segmentation models. We show the effectiveness of our proposed approach on three publicly available multi-organ segmentation datasets in both in-domain and cross-domain settings by attacking the support and query sets without the need for ad-hoc adversarial training.

cs.CV↗

arXiv2022

Robust Prototypical Few-Shot Organ Segmentation with Regularized Neural-ODEs

Despite the tremendous progress made by deep learning models in image semantic segmentation, they typically require large annotated examples, and increasing attention is being diverted to problem settings like Few-Shot Learning (FSL) where only a small amount of annotation is needed for generalisation to novel classes. This is especially seen in medical domains where dense pixel-level annotations are expensive to obtain. In this paper, we propose Regularized Prototypical Neural Ordinary Differential Equation (R-PNODE), a method that leverages intrinsic properties of Neural-ODEs, assisted and enhanced by additional cluster and consistency losses to perform Few-Shot Segmentation (FSS) of organs. R-PNODE constrains support and query features from the same classes to lie closer in the representation space thereby improving the performance over the existing Convolutional Neural Network (CNN) based FSS methods. We further demonstrate that while many existing Deep CNN based methods tend to be extremely vulnerable to adversarial attacks, R-PNODE exhibits increased adversarial robustness for a wide array of these attacks. We experiment with three publicly available multi-organ segmentation datasets in both in-domain and cross-domain FSS settings to demonstrate the efficacy of our method. In addition, we perform experiments with seven commonly used adversarial attacks in various settings to demonstrate R-PNODE's robustness. R-PNODE outperforms the baselines for FSS by significant margins and also shows superior performance for a wide array of attacks varying in intensity and design.

cs.CV↗

arXiv2021

Contrastive Semi-Supervised Learning for 2D Medical Image Segmentation

Contrastive Learning (CL) is a recent representation learning approach, which encourages inter-class separability and intra-class compactness in learned image representations. Since medical images often contain multiple semantic classes in an image, using CL to learn representations of local features (as opposed to global) is important. In this work, we present a novel semi-supervised 2D medical segmentation solution that applies CL on image patches, instead of full images. These patches are meaningfully constructed using the semantic information of different classes obtained via pseudo labeling. We also propose a novel consistency regularization (CR) scheme, which works in synergy with CL. It addresses the problem of confirmation bias, and encourages better clustering in the feature space. We evaluate our method on four public medical segmentation datasets and a novel histopathology dataset that we introduce. Our method obtains consistent improvements over state-of-the-art semi-supervised segmentation approaches for all datasets.

cs.CV↗