Nipun Kwatra

Amey Agrawal, Anmol Agarwal, Nitin Kedia et al. · gatech

Serving large language models (LLMs) in production can incur substantial costs, which has prompted recent advances in inference system optimizations. Today, these systems are evaluated against conventional latency and throughput metrics (eg. TTFT, TBT, Normalised Latency and TPOT). However, these metrics fail to fully capture the nuances of LLM inference, leading to an incomplete assessment of user-facing performance crucial for real-time applications such as chat and translation. In this paper, we first identify the pitfalls of current performance metrics in evaluating LLM inference systems. We then propose Etalon, a comprehensive performance evaluation framework that includes fluidity-index -- a novel metric designed to reflect the intricacies of the LLM inference process and its impact on real-time user experience. Finally, we evaluate various existing open-source platforms and model-as-a-service offerings using Etalon, discussing their strengths and weaknesses. Etalon is available at https://github.com/project-etalon/etalon.

Raja Gond, Nipun Kwatra, Ramachandran Ramjee

Distributed inference of large language models (LLMs) using tensor parallelism can introduce communication overheads of $20$% even over GPUs connected via NVLink, a high-speed GPU interconnect. Several techniques have been proposed to mitigate these overheads by decomposing computations into smaller tasks and overlapping communication with these subtasks. However, none of these techniques are turned on by default during tensor-parallel serving in systems like vLLM, SGLang and TensorRT-LLM. This is because the number of tokens processed per iteration is typically kept small to support low-latency serving, and decomposing such smaller workloads to enable communication overlap results in worse performance. Further, the communication itself uses many streaming multiprocessors (SMs) that would otherwise be available for computation, increasing overhead. We present TokenWeave, the first system to enable efficient compute-communication overlap for tensor-parallel model inference for token lengths as small as 1024. TokenWeave identifies RMSNorm, a previously overlooked operation, as crucial and optimizes it along with communication by implementing a novel fused AllReduce--RMSNorm kernel. Further, this kernel leverages the NVSHARP/Multimem feature available on modern GPUs (e.g., Hopper, Blackwell) to jointly perform communication and RMSNorm efficiently using only $2-8$ streaming multiprocessors (SMs) on an $8\times$H100 DGX system. Our evaluations demonstrate up to $\boldsymbol{1.28\times}$ speedup in latency (baseline$÷$ours) and up to $\boldsymbol{1.19\times}$ higher throughput (ours$÷$baseline) across multiple models and workloads. In several settings, TokenWeave delivers better performance than an equivalent model with all communication removed. The source code is available at https://github.com/microsoft/tokenweave.

Amey Agrawal, Ashish Panwar, Jayashree Mohan et al. · gatech

Large Language Model (LLM) inference consists of two distinct phases - prefill phase which processes the input prompt and decode phase which generates output tokens autoregressively. While the prefill phase effectively saturates GPU compute at small batch sizes, the decode phase results in low compute utilization as it generates one token at a time per request. The varying prefill and decode times also lead to imbalance across micro-batches when using pipeline parallelism, resulting in further inefficiency due to bubbles. We present SARATHI to address these challenges. SARATHI employs chunked-prefills, which splits a prefill request into equal sized chunks, and decode-maximal batching, which constructs a batch using a single prefill chunk and populates the remaining slots with decodes. During inference, the prefill chunk saturates GPU compute, while the decode requests 'piggyback' and cost up to an order of magnitude less compared to a decode-only batch. Chunked-prefills allows constructing multiple decode-maximal batches from a single prefill request, maximizing coverage of decodes that can piggyback. Furthermore, the uniform compute design of these batches ameliorates the imbalance between micro-batches, significantly reducing pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware. For the LLaMA-13B model on A6000 GPU, SARATHI improves decode throughput by up to 10x, and accelerates end-to-end throughput by up to 1.33x. For LLaMa-33B on A100 GPU, we achieve 1.25x higher end-to-end-throughput and up to 4.25x higher decode throughput. When used with pipeline parallelism on GPT-3, SARATHI reduces bubbles by 6.29x, resulting in an end-to-end throughput improvement of 1.91x.

Abhinav Dutta, Sanjeev Krishnan, Nipun Kwatra et al.

When Large Language Models (LLMs) are compressed using techniques such as quantization, the predominant way to demonstrate the validity of such techniques is by measuring the model's accuracy on various benchmarks.If the accuracies of the baseline model and the compressed model are close, it is assumed that there was negligible degradation in quality.However, even when the accuracy of baseline and compressed model are similar, we observe the phenomenon of flips, wherein answers change from correct to incorrect and vice versa in proportion.We conduct a detailed study of metrics across multiple compression techniques, models and datasets, demonstrating that the behavior of compressed models as visible to end-users is often significantly different from the baseline model, even when accuracy is similar.We further evaluate compressed models qualitatively and quantitatively using MT-Bench and show that compressed models are significantly worse than baseline models in this free-form generative task.Thus, we argue that compression techniques should also be evaluated using distance metrics.We propose two such metrics, KL-Divergence and flips, and show that they are well correlated.

Suril Mehta, Nipun Kwatra, Mohit Jain et al.

The use of observed wearable sensor data (e.g., photoplethysmograms [PPG]) to infer health measures (e.g., glucose level or blood pressure) is a very active area of research. Such technology can have a significant impact on health screening, chronic disease management and remote monitoring. A common approach is to collect sensor data and corresponding labels from a clinical grade device (e.g., blood pressure cuff), and train deep learning models to map one to the other. Although well intentioned, this approach often ignores a principled analysis of whether the input sensor data has enough information to predict the desired metric. We analyze the task of predicting blood pressure from PPG pulse wave analysis. Our review of the prior work reveals that many papers fall prey data leakage, and unrealistic constraints on the task and the preprocessing steps. We propose a set of tools to help determine if the input signal in question (e.g., PPG) is indeed a good predictor of the desired label (e.g., blood pressure). Using our proposed tools, we have found that blood pressure prediction using PPG has a high multi-valued mapping factor of 33.2% and low mutual information of 9.8%. In comparison, heart rate prediction using PPG, a well-established task, has a very low multi-valued mapping factor of 0.75% and high mutual information of 87.7%. We argue that these results provide a more realistic representation of the current progress towards to goal of wearable blood pressure measurement via PPG pulse wave analysis.

Dhruv Deshmukh, Saurabh Goyal, Nipun Kwatra et al.

Attention is the dominant source of latency during long-context LLM inference, an increasingly popular workload with reasoning models and RAG. We propose Kascade, a training-free sparse attention method that leverages known observations such as 1) post-softmax attention is intrinsically sparse, and 2) the identity of high-weight keys is stable across nearby layers. Kascade computes exact Top-k indices in a small set of anchor layers, then reuses those indices in intermediate reuse layers. The anchor layers are selected algorithmically, via a dynamic-programming objective that maximizes cross-layer similarity over a development set, allowing easy deployment across models. The method incorporates efficient implementation constraints (e.g. tile-level operations), across both prefill and decode attention. The Top-k selection and reuse in Kascade is head-aware and we show in our experiments that this is critical for high accuracy. Kascade achieves up to 4.1x speedup in decode attention and 2.2x speedup in prefill attention over FlashAttention-3 baseline on H100 GPUs while closely matching dense attention accuracy on long-context benchmarks such as LongBench and AIME-24.

Amey Agrawal, Nitin Kedia, Ashish Panwar et al. · gatech

Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt and produces the first output token and the second is decode which generates the rest of output tokens, one-at-a-time. Prefill iterations have high latency but saturate GPU compute due to parallel processing of the input prompt. In contrast, decode iterations have low latency but also low compute utilization because a decode iteration processes only a single token per request. This makes batching highly effective for decodes and consequently for overall throughput. However, batching multiple requests leads to an interleaving of prefill and decode iterations which makes it challenging to achieve both high throughput and low latency. We introduce an efficient LLM inference scheduler, Sarathi-Serve, to address this throughput-latency tradeoff. Sarathi-Serve introduces chunked-prefills which splits a prefill request into near equal sized chunks and creates stall-free schedules that adds new requests in a batch without pausing ongoing decodes. Stall-free scheduling unlocks the opportunity to improve throughput with large batch sizes while minimizing the effect of batching on latency. Furthermore, uniform batches in Sarathi-Serve ameliorate the imbalance between iterations resulting in minimal pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware under tail latency constraints. For Mistral-7B on single A100 GPUs, we achieve 2.6x higher serving capacity and up to 3.7x higher serving capacity for the Yi-34B model on two A100 GPUs as compared to vLLM. When used with pipeline parallelism on Falcon-180B, Sarathi-Serve provides up to 5.6x gain in the end-to-end serving capacity. The source code for Sarathi-Serve is available at https://github.com/microsoft/sarathi-serve.

Siddhartha Gairola, Pallavi Joshi, Anand Balasubramaniam et al.

Keratoconus is a severe eye disease that leads to deformation of the cornea. It impacts people aged 10-25 years and is the leading cause of blindness in that demography. Corneal topography is the gold standard for keratoconus diagnosis. It is a non-invasive process performed using expensive and bulky medical devices called corneal topographers. This makes it inaccessible to large populations, especially in the Global South. Low-cost smartphone-based corneal topographers, such as SmartKC, have been proposed to make keratoconus diagnosis accessible. Similar to medical-grade topographers, SmartKC outputs curvature heatmaps and quantitative metrics that need to be evaluated by doctors for keratoconus diagnosis. An automatic scheme for evaluation of these heatmaps and quantitative values can play a crucial role in screening keratoconus in areas where doctors are not available. In this work, we propose a dual-head convolutional neural network (CNN) for classifying keratoconus on the heatmaps generated by SmartKC. Since SmartKC is a new device and only had a small dataset (114 samples), we developed a 2-stage transfer learning strategy -- using historical data collected from a medical-grade topographer and a subset of SmartKC data -- to satisfactorily train our network. This, combined with our domain-specific data augmentations, achieved a sensitivity of 91.3% and a specificity of 94.2%.

Aditya Aggarwal, Siddhartha Gairola, Uddeshya Upadhyay et al.

Refractive error is the most common eye disorder and is the key cause behind correctable visual impairment, responsible for nearly 80% of the visual impairment in the US. Refractive error can be diagnosed using multiple methods, including subjective refraction, retinoscopy, and autorefractors. Although subjective refraction is the gold standard, it requires cooperation from the patient and hence is not suitable for infants, young children, and developmentally delayed adults. Retinoscopy is an objective refraction method that does not require any input from the patient. However, retinoscopy requires a lens kit and a trained examiner, which limits its use for mass screening. In this work, we automate retinoscopy by attaching a smartphone to a retinoscope and recording retinoscopic videos with the patient wearing a custom pair of paper frames. We develop a video processing pipeline that takes retinoscopic videos as input and estimates the net refractive error based on our proposed extension of the retinoscopy mathematical model. Our system alleviates the need for a lens kit and can be performed by an untrained examiner. In a clinical trial with 185 eyes, we achieved a sensitivity of 91.0% and specificity of 74.0% on refractive error diagnosis. Moreover, the mean absolute error of our approach was 0.75$\pm$0.67D on net refractive error estimation compared to subjective refraction measurements. Our results indicate that our approach has the potential to be used as a retinoscopy-based refractive error screening tool in real-world medical settings.

Amey Agrawal, Nitin Kedia, Jayashree Mohan et al. · gatech

Optimizing the deployment of Large language models (LLMs) is expensive today since it requires experimentally running an application workload against an LLM implementation while exploring large configuration space formed by system knobs such as parallelization strategies, batching techniques, and scheduling policies. To address this challenge, we present Vidur - a large-scale, high-fidelity, easily-extensible simulation framework for LLM inference performance. Vidur models the performance of LLM operators using a combination of experimental profiling and predictive modeling, and evaluates the end-to-end inference performance for different workloads by estimating several metrics of interest such as latency and throughput. We validate the fidelity of Vidur on several LLMs and show that it estimates inference latency with less than 9% error across the range. Further, we present Vidur-Search, a configuration search tool that helps optimize LLM deployment. Vidur-Search uses Vidur to automatically identify the most cost-effective deployment configuration that meets application performance constraints. For example, Vidur-Search finds the best deployment configuration for LLaMA2-70B in one hour on a CPU machine, in contrast to a deployment-based exploration which would require 42K GPU hours - costing ~218K dollars. Source code for Vidur is available at https://github.com/microsoft/vidur.

Aditya Chetan, Nipun Kwatra

The manifold hypothesis (real world data concentrates near low-dimensional manifolds) is suggested as the principle behind the effectiveness of machine learning algorithms in very high dimensional problems that are common in domains such as vision and speech. Multiple methods have been proposed to explicitly incorporate the manifold hypothesis as a prior in modern Deep Neural Networks (DNNs), with varying success. In this paper, we propose a new method, Distance Learner, to incorporate this prior for DNN-based classifiers. Distance Learner is trained to predict the distance of a point from the underlying manifold of each class, rather than the class label. For classification, Distance Learner then chooses the class corresponding to the closest predicted class manifold. Distance Learner can also identify points as being out of distribution (belonging to neither class), if the distance to the closest manifold is higher than a threshold. We evaluate our method on multiple synthetic datasets and show that Distance Learner learns much more meaningful classification boundaries compared to a standard classifier. We also evaluate our method on the task of adversarial robustness, and find that it not only outperforms standard classifier by a large margin, but also performs at par with classifiers trained via state-of-the-art adversarial training.

Kanishk Goel, Jayashree Mohan, Nipun Kwatra et al.

The widespread adoption of Large Language Models (LLMs) has enabled diverse applications with very different latency requirements. Existing LLM serving frameworks rely on siloed infrastructure with coarse-grained workload segregation -- interactive and batch -- leading to inefficient resource utilization and limited support for fine-grained Quality-of-Service (QoS) differentiation. This results in operational inefficiencies, over-provisioning and poor load management during traffic surges. We present Niyama, a novel QoS-driven inference serving system that enables efficient co-scheduling of diverse workloads on shared infrastructure. Niyama introduces fine-grained QoS classification allowing applications to specify precise latency requirements, and dynamically adapts scheduling decisions based on real-time system state. Leveraging the predictable execution characteristics of LLM inference, Niyama implements a dynamic chunking mechanism to improve overall throughput while maintaining strict QoS guarantees. Additionally, Niyama employs a hybrid prioritization policy that balances fairness and efficiency, and employs selective request relegation that enables graceful service degradation during overload conditions. Our evaluation demonstrates that Niyama increases serving capacity by 32% compared to current siloed deployments, while maintaining QoS guarantees. Notably, under extreme load, our system reduces SLO violations by an order of magnitude compared to current strategies.

Amey Agrawal, Nitin Kedia, Anmol Agarwal et al. · gatech

The rapid evolution of Large Language Model (LLM) inference systems has yielded significant efficiency improvements. However, our systematic analysis reveals that current evaluation methodologies frequently exhibit fundamental flaws, often manifesting as common evaluation anti-patterns that obscure true performance characteristics and impede scientific progress. Through a comprehensive examination of recent systems, we identify recurring anti-patterns across three key dimensions: Baseline Fairness, Evaluation Setup, and Metric Design. These anti-patterns are uniquely problematic for LLM inference due to its dual-phase nature combining distinct prefill and decode operations, its handling of highly heterogeneous workloads, and its strict temporal requirements for interactive use. We demonstrate how common anti-patterns -- such as inadequate baseline comparisons that conflate engineering effort with algorithmic novelty, workload selections that fail to represent production scenarios, and metric normalizations that hide substantial performance variability like generation stalls-lead to misleading conclusions. To address these challenges, we provide a comprehensive checklist derived from our analysis, establishing a framework for recognizing and avoiding these anti-patterns in favor of robust LLM inference evaluation. To demonstrate the practical application of our framework, we present a case study analyzing speculative decoding, a technique whose bursty, non-uniform token generation is easily misinterpreted when evaluated using approaches characteristic of these anti-patterns. Our work establishes a rigorous foundation for evaluation methodology, enabling meaningful comparisons, ensuring reproducible results, and ultimately accelerating genuine progress in LLM inference systems by moving beyond common anti-patterns to align evaluation with real-world requirements.

Dharma Shukla, Muthian Sivathanu, Srinidhi Viswanatha et al.

Lowering costs by driving high utilization across deep learning workloads is a crucial lever for cloud providers. We present Singularity, Microsoft's globally distributed scheduling service for highly-efficient and reliable execution of deep learning training and inference workloads. At the heart of Singularity is a novel, workload-aware scheduler that can transparently preempt and elastically scale deep learning workloads to drive high utilization without impacting their correctness or performance, across a global fleet of AI accelerators (e.g., GPUs, FPGAs). All jobs in Singularity are preemptable, migratable, and dynamically resizable (elastic) by default: a live job can be dynamically and transparently (a) preempted and migrated to a different set of nodes, cluster, data center or a region and resumed exactly from the point where the execution was preempted, and (b) resized (i.e., elastically scaled-up/down) on a varying set of accelerators of a given type. Our mechanisms are transparent in that they do not require the user to make any changes to their code or require using any custom libraries that may limit flexibility. Additionally, our approach significantly improves the reliability of deep learning workloads. We show that the resulting efficiency and reliability gains with Singularity are achieved with negligible impact on the steady-state performance. Finally, our design approach is agnostic of DNN architectures and handles a variety of parallelism strategies (e.g., data/pipeline/model parallelism).

Siddhartha Gairola, Murtuza Bohra, Nadeem Shaheer et al.

Keratoconus is a severe eye disease affecting the cornea (the clear, dome-shaped outer surface of the eye), causing it to become thin and develop a conical bulge. The diagnosis of keratoconus requires sophisticated ophthalmic devices which are non-portable and very expensive. This makes early detection of keratoconus inaccessible to large populations in low- and middle-income countries, making it a leading cause for partial/complete blindness among such populations. We propose SmartKC, a low-cost, smartphone-based keratoconus diagnosis system comprising of a 3D-printed placido's disc attachment, an LED light strip, and an intelligent smartphone app to capture the reflection of the placido rings on the cornea. An image processing pipeline analyzes the corneal image and uses the smartphone's camera parameters, the placido rings' 3D location, the pixel location of the reflected placido rings and the setup's working distance to construct the corneal surface, via the Arc-Step method and Zernike polynomials based surface fitting. In a clinical study with 101 distinct eyes, we found that SmartKC achieves a sensitivity of 94.1% and a specificity of 100.0%. Moreover, the quantitative curvature estimates (sim-K) strongly correlate with a gold-standard medical device (Pearson correlation coefficient =0.78). Our results indicate that SmartKC has the potential to be used as a keratoconus screening tool under real-world medical settings.

Nikhil Iyer, V Thejas, Nipun Kwatra et al.

One very important hyperparameter for training deep neural networks is the learning rate schedule of the optimizer. The choice of learning rate schedule determines the computational cost of getting close to a minima, how close you actually get to the minima, and most importantly the kind of local minima (wide/narrow) attained. The kind of minima attained has a significant impact on the generalization accuracy of the network. Current systems employ hand tuned learning rate schedules, which are painstakingly tuned for each network and dataset. Given that the state space of schedules is huge, finding a satisfactory learning rate schedule can be very time consuming. In this paper, we present LRTuner, a method for tuning the learning rate as training proceeds. Our method works with any optimizer, and we demonstrate results on SGD with Momentum, and Adam optimizers. We extensively evaluate LRTuner on multiple datasets, models, and across optimizers. We compare favorably against standard learning rate schedules for the given dataset and models, including ImageNet on Resnet-50, Cifar-10 on Resnet-18, and SQuAD fine-tuning on BERT. For example on ImageNet with Resnet-50, LRTuner shows up to 0.2% absolute gains in test accuracy compared to the hand-tuned baseline schedule. Moreover, LRTuner can achieve the same accuracy as the baseline schedule in 29% less optimization steps.

Siddhartha Gairola, Francis Tom, Nipun Kwatra et al.

Auscultation of respiratory sounds is the primary tool for screening and diagnosing lung diseases. Automated analysis, coupled with digital stethoscopes, can play a crucial role in enabling tele-screening of fatal lung diseases. Deep neural networks (DNNs) have shown a lot of promise for such problems, and are an obvious choice. However, DNNs are extremely data hungry, and the largest respiratory dataset ICBHI has only 6898 breathing cycles, which is still small for training a satisfactory DNN model. In this work, RespireNet, we propose a simple CNN-based model, along with a suite of novel techniques -- device specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding -- enabling us to efficiently use the small-sized dataset. We perform extensive evaluation on the ICBHI dataset, and improve upon the state-of-the-art results for 4-class classification by 2.2%

Arunava Chakraborty, Rahul Ragesh, Mahir Shah et al.

Generative adversarial networks (GANs) have been remarkably successful in learning complex high dimensional real word distributions and generating realistic samples. However, they provide limited control over the generation process. Conditional GANs (cGANs) provide a mechanism to control the generation process by conditioning the output on a user defined input. Although training GANs requires only unsupervised data, training cGANs requires labelled data which can be very expensive to obtain. We propose a framework for semi-supervised training of cGANs which utilizes sparse labels to learn the conditional mapping, and at the same time leverages a large amount of unsupervised data to learn the unconditional distribution. We demonstrate effectiveness of our method on multiple datasets and different conditional tasks.

Nikhil Iyer, V Thejas, Nipun Kwatra et al.

Several papers argue that wide minima generalize better than narrow minima. In this paper, through detailed experiments that not only corroborate the generalization properties of wide minima, we also provide empirical evidence for a new hypothesis that the density of wide minima is likely lower than the density of narrow minima. Further, motivated by this hypothesis, we design a novel explore-exploit learning rate schedule. On a variety of image and natural language datasets, compared to their original hand-tuned learning rate baselines, we show that our explore-exploit schedule can result in either up to 0.84% higher absolute accuracy using the original training budget or up to 57% reduced training time while achieving the original reported accuracy. For example, we achieve state-of-the-art (SOTA) accuracy for IWSLT'14 (DE-EN) dataset by just modifying the learning rate schedule of a high performing model.