Ziming Mao

Ziming Mao, Yihan Zhang, Chihan Cui et al.

Mixture-of-Experts (MoE) workloads rely on expert parallelism (EP) to achieve high GPU efficiency. State-of-the-art EP communication systems such as DeepEP demonstrate strong performance but exhibit poor portability across heterogeneous GPU and NIC platforms. The poor portability is rooted in architecture: GPU-initiated token-level RDMA communication requires tight vertical integration between GPUs and NICs, e.g., GPU writes to NIC driver/MMIO interfaces. We present UCCL-EP, a portable EP communication system that delivers DeepEP-level performance across heterogeneous GPU and NIC hardware. UCCL-EP replaces GPU-initiated RDMA with a high-throughput GPU-CPU control channel: compact token-routing commands are transferred to multithreaded CPU proxies, which then issue GPUDirect RDMA operations on behalf of GPUs. UCCL-EP further emulates various ordering semantics required by specialized EP communication modes using RDMA immediate data, enabling correctness on NICs that lack such ordering, e.g., AWS EFA. We implement UCCL-EP on NVIDIA and AMD GPUs with EFA and Broadcom NICs. On EFA, it outperforms the best existing EP solution by up to $2.1\times$ for dispatch and combine throughput. On NVIDIA-only platform, UCCL-EP achieves comparable performance to the original DeepEP. UCCL-EP also improves token throughput on SGLang by up to 40% on the NVIDIA+EFA platform, and improves DeepSeek-V3 training throughput over the AMD Primus/Megatron-LM framework by up to 45% on a 16-node AMD+Broadcom platform.

Shu Liu, Alexander Krentsel, Shubham Agarwal et al.

Core systems like key-value stores have historically taken years to build, and are designed to be general so as to amortize cost across deployments, paying a significant performance cost. We argue that LLM-based coding agents now make a different approach tractable: Just-in-Time Systems, in which the entire system is synthesized from scratch, specialized to the environment, workload, and required system properties. We present a JIT system synthesis pipeline, Jitskit, and explore its effectiveness in synthesizing key-value stores from spec cards that span different YCSB workloads, deployment constraints (e.g., compute resources), and system properties (e.g., consistency and durability). Jitskit iteratively refines a system implementation to match the specification against an evolving evaluation test suite. The resulting synthesized systems are performant, beating comparable state-of-the-art systems on 18 of 18 specs tried, by up to 4.6x over the best off-the-shelf baseline on the most favorable spec. Naively running Claude Code either reward-hacks or underperforms Jitskit by up to 5.4x. We discuss the challenges we overcame in building Jitskit and our key takeaways.

Shiyi Cao, Ziming Mao, Joseph E. Gonzalez et al.

Optimizing GPU kernels is critical for efficient modern machine learning systems yet remains challenging due to the complex interplay of design factors and rapid hardware evolution. Existing automated approaches typically treat Large Language Models (LLMs) merely as stochastic code generators within heuristic-guided evolutionary loops. These methods often struggle with complex kernels requiring coordinated, multi-step structural transformations, as they lack explicit planning capabilities and frequently discard promising strategies due to inefficient or incorrect intermediate implementations. To address this, we propose Search via Co-Evolving World Model and build K-Search based on this method. By replacing static search heuristics with a co-evolving world model, our framework leverages LLMs' prior domain knowledge to guide the search, actively exploring the optimization space. This approach explicitly decouples high-level algorithmic planning from low-level program instantiation, enabling the system to navigate non-monotonic optimization paths while remaining resilient to temporary implementation defects. We evaluate K-Search on diverse, complex kernels from FlashInfer, including GQA, MLA, and MoE kernels. Our results show that K-Search significantly outperforms state-of-the-art evolutionary search methods, achieving an average 2.10x improvement and up to a 14.3x gain on complex MoE kernels. On the GPUMode TriMul task, K-Search achieves state-of-the-art performance on H100, reaching 1030us and surpassing both prior evolution and human-designed solutions.

Chon Lam Lao, Zhiying Xu, Zhuang Wang et al.

Collective communication incurs significant overhead in LLM workloads. Although overlapping communication with computation in application-level is a common strategy, it often requires substantial code modifications and is impractical for many workloads (e.g., tensor and expert parallelism). We present CCCL, a built-in compression-based collective communication library that supports operations such as allreduce, alltoall, and send/recv without requiring any user-side changes, thereby enabling seamless adoption in existing applications. CCCL tightly fuses compression kernels to minimize memory accesses and integrates with NCCL to eliminate the data coalescing stage, making it fast enough (up to 3x NVLink bandwidth) to sustain communication. Our evaluation shows that CCCL improves end-to-end throughput in vLLM PD disaggregation workloads by up to 10.1% and microbenchmark throughput by up to 30%.

Ziming Mao, Hao wu, Yongxi Tan et al.

Depression is a very common but serious mood disorder.In this paper, We built a generative detection network(GDN) in accordance with three physiological laws. Our aim is that we expect the neural network to learn the relevant brain activity based on the EEG signal and, at the same time, to regenerate the target electrode signal based on the brain activity. We trained two generators, the first one learns the characteristics of depressed brain activity, and the second one learns the characteristics of control group's brain activity. In the test, a segment of EEG signal was put into the two generators separately, if the relationship between the EEG signal and brain activity conforms to the characteristics of a certain category, then the signal generated by the generator of the corresponding category is more consistent with the original signal. Thus it is possible to determine the category corresponding to a certain segment of EEG signal. We obtained an accuracy of 92.30\% on the MODMA dataset and 86.73\% on the HUSM dataset. Moreover, this model is able to output explainable information, which can be used to help the user to discover possible misjudgments of the network.Our code will be released.

Yusen Zhang, Ansong Ni, Ziming Mao et al.

Text summarization helps readers capture salient information from documents, news, interviews, and meetings. However, most state-of-the-art pretrained language models (LM) are unable to efficiently process long text for many summarization tasks. In this paper, we propose Summ$^N$, a simple, flexible, and effective multi-stage framework for input texts that are longer than the maximum context length of typical pretrained LMs. Summ$^N$ first splits the data samples and generates a coarse summary in multiple stages and then produces the final fine-grained summary based on it. Our framework can process input text of arbitrary length by adjusting the number of stages while keeping the LM input size fixed. Moreover, it can deal with both single-source documents and dialogues, and it can be used on top of different backbone abstractive summarization models. To the best of our knowledge, Summ$^N$ is the first multi-stage split-then-summarize framework for long input summarization. Our experiments demonstrate that Summ$^N$ outperforms previous state-of-the-art methods by improving ROUGE scores on three long meeting summarization datasets AMI, ICSI, and QMSum, two long TV series datasets from SummScreen, and a long document summarization dataset GovReport. Our data and code are available at https://github.com/psunlpgroup/Summ-N.

Ziming Mao, Tian Xia, Zhanghao Wu et al.

Recent years have witnessed an explosive growth of AI models. The high cost of hosting AI services on GPUs and their demanding service requirements, make it timely and challenging to lower service costs and guarantee service quality. While spot instances have long been offered with a large discount, spot preemptions have discouraged users from using them to host model replicas when serving AI models. To address this, we propose a simple yet efficient policy, SpotHedge, that leverages spot replicas across different failure domains (e.g., regions and clouds) to ensure availability, lower costs, and high service quality. SpotHedge intelligently spreads spot replicas across different regions and clouds to improve availability and reduce correlated preemptions, overprovisions cheap spot replicas than required as a safeguard against possible preemptions, and dynamically falls back to on-demand replicas when spot replicas become unavailable. We built SkyServe, a system leveraging SpotHedge to efficiently serve AI models over a mixture of spot and on-demand replicas across regions and clouds. We compared SkyServe with both research and production systems on real AI workloads: SkyServe reduces cost by 43% on average while achieving high resource availability compared to using on-demand replicas. Additionally, SkyServe improves P50, P90, and P99 latency by 2.3$\times$, 2.1$\times$, 2.1$\times$ on average compared to other research and production systems.

Yi Xu, Ziming Mao, Xiangxi Mo et al.

The rapid growth of LLMs has revolutionized natural language processing and AI analysis, but their increasing size and memory demands present significant challenges. A common solution is to spill over to CPU memory; however, traditional GPU-CPU memory swapping often results in higher latency and lower throughput. This paper introduces Pie, an LLM inference framework that addresses these challenges with performance-transparent swapping and adaptive expansion. By leveraging predictable memory access patterns and the high bandwidth of modern hardware like the NVIDIA GH200 Grace Hopper Superchip, Pie enables concurrent data swapping without affecting foreground computation, expanding effective memory without added latency. Adaptive expansion dynamically adjusts CPU memory allocation based on real-time information, optimizing memory usage and performance under varying conditions. Pie maintains low computation latency, high throughput, and high elasticity. Our experimental evaluation demonstrates that Pie achieves optimal swapping policy during cache warmup and effectively balances increased memory capacity with negligible impact on computation. With its extended capacity, Pie outperforms vLLM by up to 1.9X in throughput and 2X in latency. Additionally, Pie can reduce GPU memory usage by up to 1.67X while maintaining the same performance. Compared to FlexGen, an offline profiling-based swapping solution, Pie achieves magnitudes lower latency and 9.4X higher throughput.

Shiyi Cao, Yichuan Wang, Ziming Mao et al.

Large language model (LLM) inference workload dominates a wide variety of modern AI applications, ranging from multi-turn conversation to document analysis. Balancing fairness and efficiency is critical for managing diverse client workloads with varying prefix patterns. Unfortunately, existing fair scheduling algorithms for LLM serving, such as Virtual Token Counter (VTC), fail to take prefix locality into consideration and thus suffer from poor performance. On the other hand, locality-aware scheduling algorithms in existing LLM serving frameworks tend to maximize the prefix cache hit rate without considering fair sharing among clients. This paper introduces the first locality-aware fair scheduling algorithm, Deficit Longest Prefix Match (DLPM), which can maintain a high degree of prefix locality with a fairness guarantee. We also introduce a novel algorithm, Double Deficit LPM (D$^2$LPM), extending DLPM for the distributed setup that can find a balance point among fairness, locality, and load-balancing. Our extensive evaluation demonstrates the superior performance of DLPM and D$^2$LPM in ensuring fairness while maintaining high throughput (up to 2.87$\times$ higher than VTC) and low per-client (up to 7.18$\times$ lower than state-of-the-art distributed LLM serving system) latency.

Frank Sifei Luan, Ron Yifeng Wang, Yile Gu et al.

While ML model training and inference are both GPU-intensive, CPU-based data processing is often the bottleneck. Distributed data processing systems based on the batch or stream processing models assume homogeneous resource requirements. They excel at CPU-based computation but either under-utilize heterogeneous resources or impose high overheads on failure and reconfiguration. We introduce the streaming batch model, a hybrid of batch and streaming that enables efficient and fault-tolerant heterogeneous execution. The key idea is to use partitions as the unit of execution to achieve elasticity, but to allow partitions to be dynamically created and streamed between heterogeneous operators for memory-efficient pipelining. We present Ray Data, a streaming batch system that improves throughput on heterogeneous batch inference pipelines by 2.5-12$\times$ compared to traditional batch and stream processing systems. By leveraging heterogeneous clusters, Ray Data improves training throughput for multimodal models such as Stable Diffusion by 31% compared to single-node ML data loaders.

Yichuan Wang, Shu Liu, Zhifei Li et al.

Embedding-based search is widely used in applications such as recommendation and retrieval-augmented generation (RAG). Recently, there is a growing demand to support these capabilities over personal data stored locally on devices. However, maintaining the necessary data structure associated with the embedding-based search is often infeasible due to its high storage overhead. For example, indexing 100 GB of raw data requires 150 to 700 GB of storage, making local deployment impractical. Reducing this overhead while maintaining search quality and latency becomes a critical challenge. In this paper, we present LEANN, a storage-efficient approximate nearest neighbor (ANN) search index optimized for resource-constrained personal devices. LEANN combines a compact graph-based structure with an efficient on-the-fly recomputation strategy to enable fast and accurate retrieval with minimal storage overhead. Our evaluation shows that LEANN reduces index size to under 5% of the original raw data, achieving up to 50 times smaller storage than standard indexes, while maintaining 90% top-3 recall in under 2 seconds on real-world question answering benchmarks.

Ziming Mao, Jia Xu, Zeqi Zheng et al.

We present SAE-BrainMap, a novel framework that directly aligns deep learning visual model representations with voxel-level fMRI responses using sparse autoencoders (SAEs). First, we train layer-wise SAEs on model activations and compute the correlations between SAE unit activations and cortical fMRI signals elicited by the same natural image stimuli with cosine similarity, revealing strong activation correspondence (maximum similarity up to 0.76). Depending on this alignment, we construct a voxel dictionary by optimally assigning the most similar SAE feature to each voxel, demonstrating that SAE units preserve the functional structure of predefined regions of interest (ROIs) and exhibit ROI-consistent selectivity. Finally, we establish fine-grained hierarchical mapping between model layers and the human ventral visual pathway, also by projecting voxel dictionary activations onto individual cortical surfaces, we visualize the dynamic transformation of the visual information in deep learning models. It is found that ViT-B/16$_{CLIP}$ tends to utilize low-level information to generate high-level semantic information in the early layers and reconstructs the low-dimension information later. Our results establish a direct, downstream-task-free bridge between deep neural networks and human visual cortex, offering new insights into model interpretability.

Ziming Mao, Chen Henry Wu, Ansong Ni et al.

Transformer-based models have achieved state-of-the-art performance on short-input summarization. However, they still struggle with summarizing longer text. In this paper, we present DYLE, a novel dynamic latent extraction approach for abstractive long-input summarization. DYLE jointly trains an extractor and a generator and treats the extracted text snippets as the latent variable, allowing dynamic snippet-level attention weights during decoding. To provide adequate supervision, we propose simple yet effective heuristics for oracle extraction as well as a consistency loss term, which encourages the extractor to approximate the averaged dynamic weights predicted by the generator. We evaluate our method on different long-document and long-dialogue summarization tasks: GovReport, QMSum, and arXiv. Experiment results show that DYLE outperforms all existing methods on GovReport and QMSum, with gains up to 6.1 ROUGE, while yielding strong results on arXiv. Further analysis shows that the proposed dynamic weights provide interpretability of our generation process.

Xiangru Tang, Alexander Fabbri, Haoran Li et al.

Current pre-trained models applied to summarization are prone to factual inconsistencies which either misrepresent the source text or introduce extraneous information. Thus, comparing the factual consistency of summaries is necessary as we develop improved models. However, the optimal human evaluation setup for factual consistency has not been standardized. To address this issue, we crowdsourced evaluations for factual consistency using the rating-based Likert scale and ranking-based Best-Worst Scaling protocols, on 100 articles from each of the CNN-Daily Mail and XSum datasets over four state-of-the-art models, to determine the most reliable evaluation framework. We find that ranking-based protocols offer a more reliable measure of summary quality across datasets, while the reliability of Likert ratings depends on the target dataset and the evaluation design. Our crowdsourcing templates and summary evaluations will be publicly available to facilitate future research on factual consistency in summarization.

Linyong Nan, Chiachun Hsieh, Ziming Mao et al.

Existing table question answering datasets contain abundant factual questions that primarily evaluate the query and schema comprehension capability of a system, but they fail to include questions that require complex reasoning and integration of information due to the constraint of the associated short-form answers. To address these issues and to demonstrate the full challenge of table question answering, we introduce FeTaQA, a new dataset with 10K Wikipedia-based {table, question, free-form answer, supporting table cells} pairs. FeTaQA yields a more challenging table question answering setting because it requires generating free-form text answers after retrieval, inference, and integration of multiple discontinuous facts from a structured knowledge source. Unlike datasets of generative QA over text in which answers are prevalent with copies of short text spans from the source, answers in our dataset are human-generated explanations involving entities and their high-level relations. We provide two benchmark methods for the proposed task: a pipeline method based on semantic-parsing-based QA systems and an end-to-end method based on large pretrained text generation models, and show that FeTaQA poses a challenge for both methods.