arXiv subjects

Jidong Zhai

Publications and source records attributed to Jidong Zhai.

At least 19 recordsLinked to original sources

arXiv2026

Decoupling Readiness from Release for Tail-Aware Scheduling of Agentic LLM Workflows

Agentic LLM workflows consist of sequences of model turns interleaved with tool interactions, so their end-to-end completion time depends not only on inference speed but also on when ready turns are released. Most runtimes release each turn immediately upon readiness. Under contention, this eager release policy can accumulate released but unfinished work; once submitted, those turns can no longer be reordered by the workflow-level policy, increasing tail latency. We present a tail-risk-aware turn release scheduling method that jointly decides which ready turn to release next and how much released but unfinished work to maintain. The method uses a mean--Conditional Value-at-Risk (CVaR) objective to capture the evolving tail risk of unfinished workflows, incorporates online estimates of turn work when prioritizing ready turns, and adapts the released work budget to observed queue pressure. We evaluate the method using real agent execution traces from software engineering tasks across multiple LLMs and workflow arrival rates. The method performs comparably to eager release under light load and substantially reduces the P95 of workflow flow time under contention, achieving up to a $3.50\times$ speedup.

cs.AI↗

arXiv2026

Epoch: Compiling Diffusion Blocks for Sparse MoE Serving

Diffusion language models generate text by refining a fixed-size block of token positions through many forward passes, a loop that does not match the per-forward execution unit used by most LLM serving systems. A dense MoE runtime binds all work to the refinement-iteration clock: it rebuilds similar routing structure on every forward, recomputes expert outputs for positions whose logits are already dead, and sends those positions through dense expert-parallel collectives. This paper presents \sys{}, a serving system that treats the diffusion block as a compilation unit. \sys{} compiles a small \emph{block plan} for the block-clock structure of one diffusion block and refreshes every value that can affect a live decode decision on the iteration clock. \sys{} realizes this plan along three dense axes of an MoE forward: \atlas{} compiles a coverage-driven active expert support per layer while recomputing gate logits every iteration; \lsp{} keeps full sequence shards as model state but routes only live, newly decoded, and refresh-required positions through fresh routed-expert computation; \freshlane{} carries this fresh token--expert worklist through expert-parallel dispatch, kernels, and combine, then restores the dense logical shard at the layer boundary. We implement \sys{} on 8 NVIDIA H100 GPUs and evaluate it on three open-weight block-diffusion MoE models (LLaDA-MoE, LLaDA2.0-mini, and LLaDA2.0-Flash, spanning 7B to 100B total parameters) across GSM8K, HumanEval, MGSM, and MT-Bench. \sys{} improves end-to-end execution time by up to 2.7$\times$ over the strongest surviving baseline under the same 8-GPU placement and remains feasible at the largest batch sizes where multiple baselines run out of memory, while preserving task quality relative to the dense reference.

cs.DC↗

arXiv2026

Odin: Primitive-Level Synchronization for Distributed Point-Based Neural Rendering

Point-based neural rendering (PBNR) represents 3D scenes as explicit, trainable primitives and underpins high-quality reconstruction and emerging embodied AI and world-model pipelines. Unlike layer-structured neural networks, PBNR has primitive-indexed dependencies: each view reads and updates only a sparse, view-dependent subset of mutable scene state. As large scenes require distributed training and optimized renderers reduce per-view computation, global task- or iteration-level barriers increasingly place synchronization, rather than rendering, on the critical path. We present Odin, a distributed PBNR training system that replaces global barriers with primitive-level synchronization. Its ahead-of-time scheduler uses stable locality and phase order to identify low-conflict overlap windows, while the runtime validates primitive publication before later work observes mutable state. Odin provides a quality-first path that preserves synchronized-training visibility and a throughput-first path that uses overlap and gradient evidence to admit only small, low-impact delayed reads; structural changes and high-impact cases remain synchronized. Across four existing PBNR pipelines and 13 non-city scenes on 8 GPUs, Odin improves throughput by 1.22 times on average and hides 82% of critical-path wait while preserving reconstruction quality. In a MatrixCity mixed-parallel case study scaling to 64 GPUs, Odin improves throughput over Grendel by up to 1.89 times without changing renderer kernels, optimizers, training budgets, or model capacity.

cs.DC↗

arXiv2026

DiTango: Cost-Effective Parallel Diffusion Generation with Selective Attention State Reuse

Recent advances in AI-generated content have driven widespread adoption of Diffusion Transformers (DiTs) for high-resolution, long-duration content generation. While parallelization techniques accelerate diffusion inference, they face significant scalability challenges due to excessive communication overhead in multi-node environments. We observe that sequence partitions in Context Parallelism (CP) exhibit distinct heterogeneity: spatially proximate partitions contribute more significantly to attention computation results. By mapping this heterogeneous pattern to hierarchical communication topology, we can access high-contribution partitions with reduced communication cost. This insight motivates our novel selective attention state mechanism that strategically balances partial attention computation and historical result reuse across denoising steps. We present DiTango, an efficient parallel framework for DiT generation. DiTango features an anchor-guided state selection planner that optimizes computation-reuse decisions for each partition, complemented by a runtime that orchestrates efficient state-centric operations. This design achieves superior system efficiency while preserving generation quality. Experimental evaluation on popular diffusion models demonstrates that DiTango achieves up to 1.9x end-to-end and 3.2x attention speedup with near-linear scaling in multi-node settings, while maintaining generation quality comparable to state-of-the-art approaches.

cs.CV↗

arXiv2026

Cordon: Semantic Transactions for Tool-Using LLM Agents

Tool-using LLM agents are shifting the unit of computation from explicit human-issued commands to model-driven tasks with stateful consequences. Yet today's agent runtimes still expose tools as isolated RPCs. This interface gives runtimes a convenient integration point, but it lacks a task-scoped execution boundary for commit, rollback, recovery, and audit across multi-step agent workflows. We argue that this mismatch calls for a runtime containment boundary rather than another per-call guardrail. This paper introduces Cordon, a transactional runtime system for staging and validating irreversible agent effects before commit. A semantic transaction is a task-level execution boundary that binds tool intents and runtime-tracked result lineage to reversible local state, staged external effects, delegated authority, and audit metadata. Cordon implements this abstraction with a transaction manager that tracks derived result objects, executes reversible mutations in shadow state, stages outward-facing actions in an effect outbox, and records recovery metadata. The runtime then validates the composed execution flow before it commits state or releases external effects. Our evaluation across adversarial and benign workflows shows that Cordon exposes cross-step violations missed by existing defenses. It also reduces irreversible-effect failures while preserving benign task completion with modest approval and latency overhead.

cs.OS↗

arXiv2026

DITRON: Distributed Multi-level Tiling Compiler for Parallel Tensor Programs

The scaling of large language models (LLMs) is currently bottlenecked by the rigidity of distributed programming. While high-performance libraries like CuBLAS and NCCL provide optimized primitives, they lack the flexibility required for rapidly evolving model architectures. Conversely, existing tensor compilers fail to address the complex memory hierarchy of distributed clusters effectively. To bridge this gap, we propose DITRON, a scalable tile-level compiler that democratizes high-performance distributed kernel development. DITRON introduces a novel hierarchical programming abstraction spanning Core, Device, and Task levels to map tensor programs efficiently onto heterogeneous distributed hardware. This abstraction allows DITRON to support diverse parallelism strategies while abstracting away the complexity of inter-node and intra-node communication. Evaluated across large-scale clusters, DITRON achieves performance parity with or exceeding expert-tuned CUDA libraries, delivering speedups of $6\%-30\%$ on isolated kernels and $5\%-30\%$ on end-to-end inference in vLLM. Furthermore, DITRON demonstrates strong portability, achieving significant speedups on both NVIDIA and AMD platforms. \ours{} has been deployed at the enterprise level for both training and inference. It achieves an MFU improvement of over 10\% in training tasks, saving approximately 500,000 GPU hours of training cost per month. For inference tasks, it delivers an end-to-end gain of over 20\% and has been applied to cloud service inference and edge inference scenarios.

cs.PL↗

arXiv2026

UniEP: Unified Expert-Parallel MoE MegaKernel for LLM Training

The exponential growth in Large Language Model (LLM) parameters has transformed model training into an increasingly resource-intensive endeavor. With the stagnation of Moore's Law and the widening disparity between computation throughput and communication bandwidth, expert parallelism (EP) has emerged as a critical strategy for scaling mixture-of-experts (MoE) models. However, despite numerous proposals for optimizing EP, ranging from communication compression to computation-communication overlap, adoption within production-grade frameworks like Megatron-LM remains conservative. Existing solutions often rely on ad-hoc, complex kernels that lack adaptability across diverse optimization configurations and frequently neglect numerical stability, failing to meet the strict precision requirements of large-scale training. In this paper, we introduce UniEP, a novel system that unifies diverse EP optimization strategies into a cohesive abstraction. UniEP fuses the MoE communication and computation into MegaKernels, effectively transforming complex architectural tuning into a unified parameter search space for automated adaptability. Crucially, UniEP incorporates a deterministic token ordering mechanism that guarantees numerical consistency with sequential execution, even under aggressive overlap schedules. We evaluate UniEP on GPU clusters equipped with NVIDIA Hopper GPUs. Our results demonstrate that UniEP achieves 1.03$\times$-1.38$\times$ speedups over state-of-the-art work, effectively mitigating communication bottlenecks while maintaining the rigorous accuracy standards required for production LLM training.

cs.DC↗

arXiv2026

Jano: Adaptive Diffusion Generation with Early-stage Convergence Awareness

Diffusion models have achieved remarkable success in generative AI, yet their computational efficiency remains a significant challenge, particularly for Diffusion Transformers (DiTs) requiring intensive full-attention computation. While existing acceleration approaches focus on content-agnostic uniform optimization strategies, we observe that different regions in generated content exhibit heterogeneous convergence patterns during the denoising process. We present Jano, a training-free framework that leverages this insight for efficient region-aware generation. Jano introduces an early-stage complexity recognition algorithm that accurately identifies regional convergence requirements within initial denoising steps, coupled with an adaptive token scheduling runtime that optimizes computational resource allocation. Through comprehensive evaluation on state-of-the-art models, Jano achieves substantial acceleration (average 2.0 times speedup, up to 2.4 times) while preserving generation quality. Our work challenges conventional uniform processing assumptions and provides a practical solution for accelerating large-scale content generation. The source code of our implementation is available at https://github.com/chen-yy20/Jano.

cs.CV↗

arXiv2026

FlowPrefill: Decoupling Preemption from Prefill Scheduling Granularity to Mitigate Head-of-Line Blocking in LLM Serving

The growing demand for large language models (LLMs) requires serving systems to handle many concurrent requests with diverse service level objectives (SLOs). This exacerbates head-of-line (HoL) blocking during the compute-intensive prefill phase, where long-running requests monopolize resources and delay higher-priority ones, leading to widespread time-to-first-token (TTFT) SLO violations. While chunked prefill enables interruptibility, it introduces an inherent trade-off between responsiveness and throughput: reducing chunk size improves response latency but degrades computational efficiency, whereas increasing chunk size maximizes throughput but exacerbates blocking. This necessitates an adaptive preemption mechanism. However, dynamically balancing execution granularity against scheduling overheads remains a key challenge. In this paper, we propose FlowPrefill, a TTFT-goodput-optimized serving system that resolves this conflict by decoupling preemption granularity from scheduling frequency. To achieve adaptive prefill scheduling, FlowPrefill introduces two key innovations: 1) Operator-Level Preemption, which leverages operator boundaries to enable fine-grained execution interruption without the efficiency loss associated with fixed small chunking; and 2) Event-Driven Scheduling, which triggers scheduling decisions only upon request arrival or completion events, thereby supporting efficient preemption responsiveness while minimizing control-plane overhead. Evaluation on real-world production traces shows that FlowPrefill improves maximum goodput by up to 5.6$\times$ compared to state-of-the-art systems while satisfying heterogeneous SLOs.

cs.DC↗

arXiv2025

FlashFuser: Expanding the Scale of Kernel Fusion for Compute-Intensive Operators via Inter-Core Connection

The scaling of computation throughput continues to outpace improvements in memory bandwidth, making many deep learning workloads memory-bound. Kernel fusion is a key technique to alleviate this problem, but the fusion strategies of existing compilers and frameworks are limited to using local scratchpad memory. When the intermediate results exceed the limited capacity (such as FFN), the fusion fails. Although modern GPUs (like the NVIDIA H100) now incorporate an inter-core connection mechanism known as Distributed Shared Memory(DSM)--providing a larger, high-bandwidth, and low-latency on-chip memory pool--this hardware potential has yet to be exploited by software frameworks. To bridge this gap, we present FlashFuser, the first compiler framework to utilize inter-core connection for kernel fusion on modern GPUs. FlashFuser extends established fusion techniques to the DSM domain through three core contributions. First, we propose a powerful DSM-based communication abstraction that formalizes complex cluster-based data exchange patterns, such as reduce, shuffle and multiply. Second, we introduce a dataflow analyzer that generalizes loop scheduling, resource mapping, and tile selection to the distributed memory hierarchy; it determines the optimal execution order and tile sizes by quantifying data movement across memory levels. Finally, FlashFuser integrates these components into a unified search engine that employs analytical cost modeling and DSM-aware pruning strategies to efficiently discover the optimal execution plan. Our evaluation on an NVIDIA H100 GPU shows that FlashFuser reduces memory access by 58% and delivers kernel speedups of 3.3x against highly-tuned libraries and 4.1x against state-of-the-art compilers, resulting in a 1.24x end-to-end speedup.

cs.DC↗

arXiv2025

Lethe: Layer- and Time-Adaptive KV Cache Pruning for Reasoning-Intensive LLM Serving

Generative reasoning with large language models (LLMs) often involves long decoding sequences, leading to substantial memory and latency overheads from accumulating key-value (KV) caches. While existing KV compression methods primarily focus on reducing prefill memory from long input sequences, they fall short in addressing the dynamic and layer-sensitive nature of long-form generation, which is central to reasoning tasks. We propose Lethe, a dynamic KV cache management framework that introduces adaptivity along both the spatial and temporal dimensions of decoding. Along the spatial dimension, Lethe performs layerwise sparsity-aware allocation, assigning token pruning budgets to each transformer layer based on estimated attention redundancy. Along the temporal dimension, Lethe conducts multi-round token pruning during generation, driven by a Recency-Aware Selective Retention} (RASR) mechanism. RASR extends traditional recency-based heuristics by also considering token relevance derived from evolving attention patterns, enabling informed decisions about which tokens to retain or evict. Empirical results demonstrate that Lethe achieves a favorable balance between efficiency and generation quality across diverse models and tasks, increases throughput by up to 2.56x.

cs.LG↗

arXiv2025

A Flexible Programmable Pipeline Parallelism Framework for Efficient DNN Training

Pipeline parallelism is an essential distributed parallelism method. Increasingly complex and diverse DNN models necessitate meticulously customized pipeline schedules for performance. However, existing practices typically rely on predefined schedules, each with strengths, but fail to adapt automatically to the emerging model architectures. Exploring novel high-efficiency schedules is daunting due to the enormous and varying schedule space. Besides, manually implementing schedules can be challenging due to the onerous coding burdens and constantly changing needs. Unfortunately, existing frameworks have limitations in automated schedule exploration and lack flexibility and controllability. This paper presents FlexPipe, a programmable pipeline parallelism framework with enhanced productivity, programmability, debuggability, and ease of tuning. FlexPipe has two main components: a succinct domain-specific language (DSL) and an automated scheduler. FlexPipe enables automated schedule exploration for various parallel scenarios within a broad spectrum of schedule types at a small search cost. Besides, users can swiftly develop and customize schedules using the FlexPipe DSL, which embodies flexible controllability in the pipeline order of micro-batch computations over stages. It also provides convenient mechanisms to include new operations in schedules to meet changing demands. Our evaluation results demonstrate that FlexPipe achieves up to 2.28X performance speedup compared to the popular large-scale parallel framework Megtron-LM, and gains up to 1.49X performance speedup compared to the state-of-the-art automated pipeline parallelism framework.

cs.DC↗

arXiv2025

HARMONY: A Scalable Distributed Vector Database for High-Throughput Approximate Nearest Neighbor Search

Approximate Nearest Neighbor Search (ANNS) is essential for various data-intensive applications, including recommendation systems, image retrieval, and machine learning. Scaling ANNS to handle billions of high-dimensional vectors on a single machine presents significant challenges in memory capacity and processing efficiency. To address these challenges, distributed vector databases leverage multiple nodes for the parallel storage and processing of vectors. However, existing solutions often suffer from load imbalance and high communication overhead, primarily due to traditional partition strategies that fail to effectively distribute the workload. In this paper, we introduce Harmony, a distributed ANNS system that employs a novel multi-granularity partition strategy, combining dimension-based and vector-based partition. This strategy ensures a balanced distribution of computational load across all nodes while effectively minimizing communication costs. Furthermore, Harmony incorporates an early-stop pruning mechanism that leverages the monotonicity of distance computations in dimension-based partition, resulting in significant reductions in both computational and communication overhead. We conducted extensive experiments on diverse real-world datasets, demonstrating that Harmony outperforms leading distributed vector databases, achieving 4.63 times throughput on average in four nodes and 58% performance improvement over traditional distribution for skewed workloads.

cs.DB↗

arXiv2025

Triton-distributed: Programming Overlapping Kernels on Distributed AI Systems with the Triton Compiler

In this report, we propose Triton-distributed, an extension of existing Triton compiler, to overcome the programming challenges in distributed AI systems. Triton-distributed is the first compiler that supports native overlapping optimizations for distributed AI workloads, providing a good coverage of existing optimizations from different frameworks. First, we integrate communication primitives compliant with the OpenSHMEM standard into the compiler. This enables programmers to utilize these primitives with a higher-level Python programming model. Second, we illustrate how to achieve complex joint optimization of computation, memory access, and communication with the assistance of the compiler. In particular, we show how to use overlapping techniques to hide latency and present our compiler-based programming methods in both single-node and multi-node scenarios. Finally, we showcase the performance of the code generated by our compiler. In a test environment with up to 64 devices, our compiler can fully utilize heterogeneous communication and computation resources to provide effective overlapping and high performance. In many cases, the performance of the generated code can even outperform hand-optimized code. Moreover, the development difficulty and the time cost for development using our compiler are far less than those of low-level programming such as CUDA/C++, which clearly demonstrates significant productivity advantages.

cs.DC↗

arXiv2025

SpecRouter: Adaptive Routing for Multi-Level Speculative Decoding in Large Language Models

Large Language Models (LLMs) present a critical trade-off between inference quality and computational cost: larger models offer superior capabilities but incur significant latency, while smaller models are faster but less powerful. Existing serving strategies often employ fixed model scales or static two-stage speculative decoding, failing to dynamically adapt to the varying complexities of user requests or fluctuations in system performance. This paper introduces \systemname{}, a novel framework that reimagines LLM inference as an adaptive routing problem solved through multi-level speculative decoding. \systemname{} dynamically constructs and optimizes inference "paths" (chains of models) based on real-time feedback, addressing the limitations of static approaches. Our contributions are threefold: (1) An \textbf{adaptive model chain scheduling} mechanism that leverages performance profiling (execution times) and predictive similarity metrics (derived from token distribution divergence) to continuously select the optimal sequence of draft and verifier models, minimizing predicted latency per generated token. (2) A \textbf{multi-level collaborative verification} framework where intermediate models within the selected chain can validate speculative tokens, reducing the verification burden on the final, most powerful target model. (3) A \textbf{synchronized state management} system providing efficient, consistent KV cache handling across heterogeneous models in the chain, including precise, low-overhead rollbacks tailored for asynchronous batch processing inherent in multi-level speculation. Preliminary experiments demonstrate the validity of our method.

cs.LG↗

arXiv2025

Jenga: Effective Memory Management for Serving LLM with Heterogeneity

Large language models (LLMs) are widely used but expensive to run, especially as inference workloads grow. To lower costs, maximizing the request batch size by managing GPU memory efficiently is crucial. While PagedAttention has recently been proposed to improve the efficiency of memory management, we find that the growing heterogeneity in the embeddings dimensions, attention, and access patterns of modern LLM architectures introduces new challenges for memory allocation. In this paper, we present Jenga, a novel memory allocation framework for heterogeneous embeddings in LLMs. Jenga tackles two key challenges: (1) minimizing memory fragmentation when managing embeddings of different sizes, and (2) enabling flexible caching and eviction policies tailored to the specific token-dependency patterns of various layers. Jenga employs a two-level memory allocator, leveraging the least common multiple (LCM) of embedding sizes to optimize memory usage and providing APIs to express layer-specific caching logic to enhance memory reuse. We implemente Jenga on vLLM, a state-of-the-art LLM inference engine, and evaluate it with diverse LLMs, datasets, and GPU configurations. Evaluations show that Jenga improves GPU memory utilization by up to 79.6%, and increases serving throughput by up to 4.92x (1.80x on average).

cs.DC↗

arXiv2025

FastCache: Optimizing Multimodal LLM Serving through Lightweight KV-Cache Compression Framework

Multi-modal Large Language Models (MLLMs) serving systems commonly employ KV-cache compression to reduce memory footprint. However, existing compression methods introduce significant processing overhead and queuing delays, particularly in concurrent serving scenarios. We present \texttt{FastCache}, a novel serving framework that effectively addresses these challenges through two key innovations: (1) a dynamic batching strategy that optimizes request scheduling across prefill, compression, and decode stages, and (2) an efficient KV-cache memory pool mechanism that eliminates memory fragmentation while maintaining high GPU utilization. Our comprehensive experiments on the GQA and MileBench datasets demonstrate that \texttt{FastCache} achieves up to 19.3$\times$ reduction in Time-To-First-Token (TTFT) and 12.1$\times$ improvement in throughput compared to state-of-the-art baselines. The system maintains stable performance under high-concurrency scenarios (up to 40 req/s) while reducing average memory consumption by 20\%. These results establish \texttt{FastCache} as an efficient solution for real-world LLM serving systems with KV-cache compression.

cs.MM↗

arXiv2025

GS-Cache: A GS-Cache Inference Framework for Large-scale Gaussian Splatting Models

Rendering large-scale 3D Gaussian Splatting (3DGS) model faces significant challenges in achieving real-time, high-fidelity performance on consumer-grade devices. Fully realizing the potential of 3DGS in applications such as virtual reality (VR) requires addressing critical system-level challenges to support real-time, immersive experiences. We propose GS-Cache, an end-to-end framework that seamlessly integrates 3DGS's advanced representation with a highly optimized rendering system. GS-Cache introduces a cache-centric pipeline to eliminate redundant computations, an efficiency-aware scheduler for elastic multi-GPU rendering, and optimized CUDA kernels to overcome computational bottlenecks. This synergy between 3DGS and system design enables GS-Cache to achieve up to 5.35x performance improvement, 35% latency reduction, and 42% lower GPU memory usage, supporting 2K binocular rendering at over 120 FPS with high visual quality. By bridging the gap between 3DGS's representation power and the demands of VR systems, GS-Cache establishes a scalable and efficient framework for real-time neural rendering in immersive environments.

cs.CV↗