Haoran Qiu

Amey Agrawal, Haoran Qiu, Junda Chen et al. · gatech

Deploying million-token Large Language Models (LLMs) is challenging because production workloads are highly heterogeneous, mixing short queries and long documents. This heterogeneity, combined with the quadratic complexity of attention, creates severe convoy effects where long-running requests stall short, interactive ones, degrading system responsiveness. We present Medha, a serving system that eliminates these convoys by introducing fine-grained, preemptive scheduling to LLM inference. Medha makes preemption practical with a co-designed set of mechanisms -- including Adaptive Chunking and Stream Pipeline Parallel that overcome the perceived inefficiencies and scaling challenges of chunking. Additionally, we present a new parallelism strategy KV-Cache Parallelism to reduce the decode latency and afford interactivity despite very long context. These mechanisms are orchestrated by a Length-Aware Relative Slack (LARS) scheduler, a deadline and heterogeneity-aware scheduling policy that prevents both the convoy effect and the starvation that plagues simpler policies. Under a heterogeneous workload, Medha improves throughput by 5.7x while reducing median and 99th percentile latency by 30x and 174x, respectively, compared to state-of-the-art non-preemptive systems.

Haoran Qiu, Weichao Mao, Archit Patke et al.

Large language models (LLMs) have been driving a new wave of interactive AI applications across numerous domains. However, efficiently serving LLM inference requests is challenging due to their unpredictable execution times originating from the autoregressive nature of generative models. Existing LLM serving systems exploit first-come-first-serve (FCFS) scheduling, suffering from head-of-line blocking issues. To address the non-deterministic nature of LLMs and enable efficient interactive LLM serving, we present a speculative shortest-job-first (SSJF) scheduler that uses a light proxy model to predict LLM output sequence lengths. Our open-source SSJF implementation does not require changes to memory management or batching strategies. Evaluations on real-world datasets and production workload traces show that SSJF reduces average job completion times by 30.5-39.6% and increases throughput by 2.2-3.6x compared to FCFS schedulers, across no batching, dynamic batching, and continuous batching settings.

Haoran Qiu, Gohar Irfan Chaudhry, Chaojie Zhang et al.

Advances in multi-modal generative models are enabling new applications, from storytelling to automated media synthesis. Most current workloads generate simple outputs (e.g., image generation from a prompt) in batch mode, often requiring several seconds even for basic results. Serving real-time multi-modal workflows at scale is costly and complex, requiring efficient coordination of diverse models (each with unique resource needs) across language, audio, image, and video, all under strict latency and resource constraints. We tackle these challenges through the lens of real-time podcast video generation, integrating LLMs, text-to-speech, and video-audio generation. To meet tight SLOs, we design an adaptive, modular serving system, StreamWise, that dynamically manages quality (e.g., resolution, sharpness), model/content parallelism, and resource-aware scheduling. We leverage heterogeneous hardware to maximize responsiveness and efficiency. For example, the system can lower video resolution and allocate more resources to early scenes. We quantify the trade-offs between latency, cost, and quality. The cheapest setup generates a 10-minute podcast video on A100 GPUs in 1.4 hours (8.4x slower than the real-time) for less than \$25. StreamWise enables high-quality real-time streaming with a sub-second startup delay under $45.

Haoran Qiu, Anish Biswas, Zihan Zhao et al.

Large multimodal models (LMMs) demonstrate impressive capabilities in understanding images, videos, and audio beyond text. However, efficiently serving LMMs in production environments poses significant challenges due to their complex architectures and heterogeneous characteristics across their multi-stage inference pipelines. We present the first comprehensive systems analysis of two prominent LMM architectures, decoder-only and cross-attention, across six representative open-source models, revealing key systems design implications. We also present an in-depth analysis of production LMM inference traces, uncovering unique workload characteristics, including variable, heavy-tailed request distributions and bursty traffic patterns. Based on these insights, we propose ModServe, a modular LMM serving system that decouples stages for independent optimization and adaptive scaling. ModServe dynamically reconfigures stages and handles bursty traffic with modality-aware scheduling and autoscaling to meet tail latency SLOs while minimizing costs. ModServe achieves 3.3-5.5x higher throughput (leading to 25-41.3% cost saving) while meeting SLOs on a 128-GPU cluster with production traces.

Archit Patke, Dhemath Reddy, Saurabh Jha et al.

Large language model (LLM) serving is becoming an increasingly critical workload for cloud providers. Existing LLM serving systems focus on interactive requests, such as chatbots and coding assistants, with tight latency SLO requirements. However, when such systems execute batch requests that have relaxed SLOs along with interactive requests, it leads to poor multiplexing and inefficient resource utilization. To address these challenges, we propose QLM, a queue management system for LLM serving. QLM maintains batch and interactive requests across different models and SLOs in a request queue. Optimal ordering of the request queue is critical to maintain SLOs while ensuring high resource utilization. To generate this optimal ordering, QLM uses a Request Waiting Time (RWT) Estimator that estimates the waiting times for requests in the request queue. These estimates are used by a global scheduler to orchestrate LLM Serving Operations (LSOs) such as request pulling, request eviction, load balancing, and model swapping. Evaluation on heterogeneous GPU devices and models with real-world LLM serving dataset shows that QLM improves SLO attainment by 40-90% and throughput by 20-400% while maintaining or improving device utilization compared to other state-of-the-art LLM serving systems. QLM's evaluation is based on the production requirements of a cloud provider. QLM is publicly available at https://www.github.com/QLM-project/QLM.

Jovan Stojkovic, Chaojie Zhang, Íñigo Goiri et al.

The rising demand for generative large language models (LLMs) poses challenges for thermal and power management in cloud datacenters. Traditional techniques often are inadequate for LLM inference due to the fine-grained, millisecond-scale execution phases, each with distinct performance, thermal, and power profiles. Additionally, LLM inference workloads are sensitive to various configuration parameters (e.g., model parallelism, size, and quantization) that involve trade-offs between performance, temperature, power, and output quality. Moreover, clouds often co-locate SaaS and IaaS workloads, each with different levels of visibility and flexibility. We propose TAPAS, a thermal- and power-aware framework designed for LLM inference clusters in the cloud. TAPAS enhances cooling and power oversubscription capabilities, reducing the total cost of ownership (TCO) while effectively handling emergencies (e.g., cooling and power failures). The system leverages historical temperature and power data, along with the adaptability of SaaS workloads, to: (1) efficiently place new GPU workload VMs within cooling and power constraints, (2) route LLM inference requests across SaaS VMs, and (3) reconfigure SaaS VMs to manage load spikes and emergency situations. Our evaluation on a large GPU cluster demonstrates significant reductions in thermal and power throttling events, boosting system efficiency.

Xingqi Cui, Chieh-Jan Mike Liang, Jiarong Xing et al.

Serving large generative models such as LLMs and multi- modal transformers requires balancing user-facing SLOs (e.g., time-to-first-token, time-between-tokens) with provider goals of efficiency and cost reduction. Existing solutions rely on static provisioning or model-level autoscaling, both of which treat the model as a monolith. This coarse-grained resource management leads to degraded performance or significant resource underutilization due to poor adaptability to dynamic inference traffic that is common online. The root cause of this inefficiency lies in the internal structure of generative models: they are executed as graphs of interconnected operators. Through detailed characterization and systematic analysis, we find that operators are heterogeneous in their compute and memory footprints and exhibit diverse sensitivity to workload and resource factors such as batch size, sequence length, and traffic rate. This heterogeneity suggests that the operator, rather than the entire model, is the right granularity for scaling decisions. We propose an operator-level autoscaling framework, which allocates resources at finer (operator)-granularity, optimizing the scaling, batching, and placement based on individual operator profiles. Evaluated on production-scale traces, our approach preserves SLOs with up to 40% fewer GPUs and 35% less energy, or under fixed resources achieves 1.6x higher throughput with 5% less energy. These results show that the operator, rather than the model, is fundamentally a more effective unit for scaling large generative workloads.

Xiangyuan Zhang, Weichao Mao, Haoran Qiu et al.

Closed-loop control of nonlinear dynamical systems with partial-state observability demands expert knowledge of a diverse, less standardized set of theoretical tools. Moreover, it requires a delicate integration of controller and estimator designs to achieve the desired system behavior. To establish a general controller synthesis framework, we explore the Decision Transformer (DT) architecture. Specifically, we first frame the control task as predicting the current optimal action based on past observations, actions, and rewards, eliminating the need for a separate estimator design. Then, we leverage the pre-trained language models, i.e., the Generative Pre-trained Transformer (GPT) series, to initialize DT and subsequently train it for control tasks using low-rank adaptation (LoRA). Our comprehensive experiments across five distinct control tasks, ranging from maneuvering aerospace systems to controlling partial differential equations (PDEs), demonstrate DT's capability to capture the parameter-agnostic structures intrinsic to control tasks. DT exhibits remarkable zero-shot generalization abilities for completely new tasks and rapidly surpasses expert performance levels with a minimal amount of demonstration data. These findings highlight the potential of DT as a foundational controller for general control applications.

Gohar Irfan Chaudhry, Esha Choukse, Haoran Qiu et al.

Agentic workflows commonly coordinate multiple models and tools with complex control logic. They are quickly becoming the dominant paradigm for AI applications. However, serving them remains inefficient with today's frameworks. The key problem is that they expose workflows as opaque sequences of model and tool calls that tightly couple agent logic with model and hardware choices. Often, these workflow components are fragmented across different entities, preventing systems from reasoning about trade-offs across accuracy, latency, energy, and cost. This leads to resource waste and degraded service-level objectives (SLOs). We present Murakkab, a resource-efficient serving system for agentic workflows. Murakkab introduces a declarative abstraction that decouples workflow specification from execution configuration. A profile-guided optimizer and adaptive runtime jointly manage the full stack: orchestrating workflow components, mapping them to models and hardware, and dynamically reconfiguring execution to satisfy user-defined SLOs. By exposing the internal structure of agentic workflows, Murakkab enables cross-layer optimization that existing frameworks and cloud schedulers cannot achieve. Our evaluation on diverse workflows shows that Murakkab reduces GPU usage by up to 2.8$\times$, energy consumption by 3.7$\times$, and cost by 4.3$\times$ while maintaining SLOs.

Weichao Mao, Haoran Qiu, Chen Wang et al.

No-regret learning has a long history of being closely connected to game theory. Recent works have devised uncoupled no-regret learning dynamics that, when adopted by all the players in normal-form games, converge to various equilibrium solutions at a near-optimal rate of $\widetilde{O}(T^{-1})$, a significant improvement over the $O(1/\sqrt{T})$ rate of classic no-regret learners. However, analogous convergence results are scarce in Markov games, a more generic setting that lays the foundation for multi-agent reinforcement learning. In this work, we close this gap by showing that the optimistic-follow-the-regularized-leader (OFTRL) algorithm, together with appropriate value update procedures, can find $\widetilde{O}(T^{-1})$-approximate (coarse) correlated equilibria in full-information general-sum Markov games within $T$ iterations. Numerical results are also included to corroborate our theoretical findings.