Nishil Talati

Yuchen Xia, Jiho Kim, Yuhan Chen et al.

Due to the cost-prohibitive nature of training Large Language Models (LLMs), fine-tuning has emerged as an attractive alternative for specializing LLMs for specific tasks using limited compute resources in a cost-effective manner. In this paper, we characterize sparse Mixture of Experts (MoE) based LLM fine-tuning to understand their accuracy and runtime performance on a single GPU. Our evaluation provides unique insights into the training efficacy of sparse and dense versions of MoE models, as well as their runtime characteristics, including maximum batch size, execution time breakdown, end-to-end throughput, GPU hardware utilization, and load distribution. Our study identifies the optimization of the MoE layer as crucial for further improving the performance of LLM fine-tuning. Using our profiling results, we also develop and validate an analytical model to estimate the cost of LLM fine-tuning on the cloud. This model, based on parameters of the model and GPU architecture, estimates LLM throughput and the cost of training, aiding practitioners in industry and academia to budget the cost of fine-tuning a specific model.

Haojie Ye, Arjun Laxman, Yichao Yuan et al.

Money laundering detection faces challenges due to excessive false positives and inadequate adaptation to sophisticated multi-stage schemes that exploit modern financial networks. Graph analytics and AI are promising tools, but they struggle with the fuzziness of laundering patterns, which exhibit structural and temporal variations. Conventional data mining techniques require the detailed enumeration of pattern variants, which not only complicates the analyst's task to specify them, but also leads to large run-time overheads and difficulty training accurate AI models. The paper presents BlazingAML, a scalable AML system design that introduces: 1. A novel multi-stage framework for expressing fuzzy money laundering patterns 2. A domain-specific compiler that transforms high-level pattern descriptions into high-performance code for CPU and GPU back-ends The multi-stage abstraction decomposes complex laundering schemes into logical stages connected by graph operations, enabling diverse patterns to be expressed using unified primitives while capturing structural and temporal fuzziness. The compiler applies sophisticated optimizations, eliminating manual parallel programming requirements for financial analysts. Evaluation on IBM AML datasets shows BlazingAML achieves the same F1 score as state-of-the-art approaches while delivering 210x and 333x higher speedup on CPU and GPU respectively, with superior scalability.

Yichao Yuan, Ankita Nayak, Souvik Kundu et al.

Agentic AI shifts LLM serving from isolated prompt-generation requests to stateful, multi-turn executions that repeatedly invoke the model, call tools, and grow context over time. This paper characterizes ReAct-style agents from both the LLM-serving and tool-execution perspectives using an end-to-end tracing infrastructure across reasoning and non-reasoning Gemma and Qwen configurations on five agentic benchmarks. Our study shows that agentic workloads are not simply long-prompt workloads: with effective context caching, most input tokens are reused across turns, making execution decode-dominated while increasing dependence on long-lived KV-cache state. We also find that tool use has a clear temporal structure, with agents shifting from read/explore behavior early in execution to execute/write behavior later. These results show that efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior.

Yichao Yuan, Mosharaf Chowdhury, Nishil Talati

Power has become a central bottleneck for AI inference. This problem is becoming more urgent as agentic AI emerges as a major workload class, yet prior power-management techniques focus almost entirely on single-turn LLM serving. Our analysis shows that agentic serving behaves fundamentally differently: each request carries long-lived context that evolves across tool-interleaved turns, and lowering GPU frequency can push the system into a thrashing regime where memory pressure sharply worsens both performance and power efficiency. These observations show that power optimization for agentic serving requires rethinking. We present KAIROS, a context-aware power optimization system for agentic AI serving. KAIROS uses agent context as a first-class control signal to jointly manage GPU frequency, per-instance concurrency, and multi-instance request placement. This enables KAIROS to save power when memory headroom exists while avoiding thrashing and preserving performance targets. At a high level, KAIROS tracks requests at agent granularity, adapts local control to context growth and agent progress, and routes agents across instances to jointly improve power efficiency and memory stability. Evaluated across diverse software and data engineering agentic tasks, KAIROS achieves an average of 27% (up to 39.8%) power reduction while meeting the performance targets.

Jae-Won Chung, Nishil Talati, Mosharaf Chowdhury

The "AI for Science, Energy, and Security" report from DOE outlines a significant focus on developing and optimizing artificial intelligence workflows for a foundational impact on a broad range of DOE missions. With the pervasive usage of artificial intelligence (AI) and machine learning (ML) tools and techniques, their energy efficiency is likely to become the gating factor toward adoption. This is because generative AI (GenAI) models are massive energy hogs: for instance, training a 200-billion parameter large language model (LLM) at Amazon is estimated to have taken 11.9 GWh, which is enough to power more than a thousand average U.S. households for a year. Inference consumes even more energy, because a model trained once serve millions. Given this scale, high energy efficiency is key to addressing the power delivery problem of constructing and operating new supercomputers and datacenters specialized for AI workloads. In that regard, we outline software- and architecture-level research challenges and opportunities, setting the stage for creating cross-layer energy optimizations in AI systems.

Yichao Yuan, Lin Ma, Nishil Talati

Mixture of Experts (MoE) LLMs, characterized by their sparse activation patterns, offer a promising approach to scaling language models while avoiding proportionally increasing the inference cost. However, their large parameter sizes present deployment challenges in resource-constrained environments with limited GPU memory capacity, as GPU memory is often insufficient to accommodate the full set of model weights. Consequently, typical deployments rely on CPU-GPU hybrid execution: the GPU handles compute-intensive GEMM operations, while the CPU processes the relatively lightweight attention mechanism. This setup introduces a key challenge: how to effectively optimize resource utilization across CPU and GPU? Prior work has designed system optimizations based on performance models with limited scope. Specifically, such models do not capture the complex interactions between hardware properties and system execution mechanisms. Therefore, previous approaches neither identify nor achieve the hardware limit. This paper presents MoE-Lens, a high-throughput MoE LLM inference system designed through holistic performance modeling for resource-constrained environments. Our performance model thoroughly analyzes various fundamental system components, including CPU memory capacity, GPU compute power, and workload characteristics, to understand the theoretical performance upper bound of MoE inference. Furthermore, it captures the system execution mechanisms to identify the key hardware bottlenecks and accurately predict the achievable throughput. Informed by our performance model, MoE-Lens introduces an inference system approaching hardware limits. Evaluated on diverse MoE models and datasets, MoE-Lens outperforms the state-of-the-art solution by 4.6x on average (up to 25.5x), with our theoretical model predicting performance with an average 94% accuracy.

Yuchen Xia, Souvik Kundu, Mosharaf Chowdhury et al.

Novel View Synthesis (NVS) is the task of generating new images of a scene from viewpoints that were not part of the original input. Diffusion-based NVS can generate high-quality, temporally consistent images, however, remains computationally prohibitive. Conversely, regression-based NVS offers suboptimal generation quality despite requiring significantly lower compute; leaving the design objective of a high-quality, inference-efficient NVS framework an open challenge. To close this critical gap, we present Sphinx, a training-free hybrid inference framework that achieves diffusion-level fidelity at a significantly lower compute. Sphinx proposes to use regression-based fast initialization to guide and reduce the denoising workload for the diffusion model. Additionally, it integrates selective refinement with adaptive noise scheduling, allowing more compute to uncertain regions and frames. This enables Sphinx to provide flexible navigation of the performance-quality trade-off, allowing adaptation to latency and fidelity requirements for dynamically changing inference scenarios. Our evaluation shows that Sphinx achieves an average 1.8x speedup over diffusion model inference with negligible perceptual degradation of less than 5%, establishing a new Pareto frontier between quality and latency in NVS serving.

Savini Kashmira, Jayanaka Dantanarayana, Thamirawaran Sathiyalogeswaran et al.

This paper presents GraphMend, a high-level compiler that eliminates FX graph breaks in PyTorch 2 programs. Although PyTorch 2 introduced TorchDynamo and TorchInductor to enable just-in-time graph compilation, unresolved dynamic control flow and unsupported Python constructs often fragment models into multiple FX graphs. These fragments force frequent fallbacks to eager mode, incur costly CPU-to-GPU synchronizations, and reduce optimization opportunities. GraphMend addresses this limitation by analyzing and transforming source code before execution. Built on the Jac compilation framework, GraphMend introduces two code transformations that remove graph breaks due to dynamic control flow and Python I/O functions. This design allows PyTorch's compilation pipeline to capture larger, uninterrupted FX graphs without requiring manual refactoring by developers. Evaluation across eight Hugging Face models shows that GraphMend removes all fixable graph breaks due to dynamic control flow and Python I/O functions, driving the break count to 0 in 6 models and reducing it from 5 to 2 in another model. On NVIDIA RTX 3090 and A40 GPUs, GraphMend achieves up to 75% latency reductions and up to 8% higher end-to-end throughput. These results demonstrate that high-level code transformation is an effective complement to PyTorch's dynamic JIT compilation pipeline, substantially improving both usability and performance.