Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve

Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt to produce one 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 inspired by the techniques we originally proposed for optimizing throughput in Sarathi. Sarathi-Serve leverages chunked-prefills from Sarathi to create stall-free schedules that can add 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. Our evaluation shows that Sarathi-Serve improves serving throughput within desired latency SLOs of Mistral-7B by up to 2.6x on a single A100 GPU and up to 6.9x for Falcon-180B on 8 A100 GPUs over Orca and vLLM.

SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills

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.

NGAME: Negative Mining-aware Mini-batching for Extreme Classification

Extreme Classification (XC) seeks to tag data points with the most relevant subset of labels from an extremely large label set. Performing deep XC with dense, learnt representations for data points and labels has attracted much attention due to its superiority over earlier XC methods that used sparse, hand-crafted features. Negative mining techniques have emerged as a critical component of all deep XC methods that allow them to scale to millions of labels. However, despite recent advances, training deep XC models with large encoder architectures such as transformers remains challenging. This paper identifies that memory overheads of popular negative mining techniques often force mini-batch sizes to remain small and slow training down. In response, this paper introduces NGAME, a light-weight mini-batch creation technique that offers provably accurate in-batch negative samples. This allows training with larger mini-batches offering significantly faster convergence and higher accuracies than existing negative sampling techniques. NGAME was found to be up to 16% more accurate than state-of-the-art methods on a wide array of benchmark datasets for extreme classification, as well as 3% more accurate at retrieving search engine queries in response to a user webpage visit to show personalized ads. In live A/B tests on a popular search engine, NGAME yielded up to 23% gains in click-through-rates.

Singularity: Planet-Scale, Preemptive and Elastic Scheduling of AI Workloads

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).

Singularity: Planet-Scale, Preemptible, Elastic Scheduling of AI Workloads

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).

LRTuner: A Learning Rate Tuner for Deep Neural Networks

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.

Wide-minima Density Hypothesis and the Explore-Exploit Learning Rate Schedule

While the generalization properties of neural networks are not yet well understood, 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 0.5\% higher absolute accuracy using the original training budget or up to 44\% reduced training time while achieving the original reported accuracy.

Privado: Practical and Secure DNN Inference

Recently, cloud providers have extended support for trusted hardware primitives such as Intel SGX. Simultaneously, the field of deep learning is seeing enormous innovation and increase in adoption. In this paper, we therefore ask the question: "Can third-party cloud services use SGX to provide practical, yet secure DNN Inference-as-a-service? " Our work addresses the three main challenges that SGX-based DNN inferencing faces, namely, security, ease-of-use, and performance. We first demonstrate that side-channel based attacks on DNN models are indeed possible. We show that, by observing access patterns, we can recover inputs to the DNN model. This motivates the need for Privado, a system we have designed for secure inference-as-a-service. Privado is input-oblivious: it transforms any deep learning framework written in C/C++ to be free of input-dependent access patterns. Privado is fully-automated and has a low TCB: with zero developer effort, given an ONNX description, it generates compact C code for the model which can run within SGX-enclaves. Privado has low performance overhead: we have used Privado with Torch, and have shown its overhead to be 20.77\% on average on 10 contemporary networks.