arXiv subjects

Chris Wu

Publications and source records attributed to Chris Wu.

6 recordsLinked to original sources

arXiv2026

LeanSparse: Memory-Efficient Serving for Native Sparse-Attention LLMs

Sparse attention limits the history read by each query, but serving long contexts still requires substantial GPU memory. Existing KV offloading systems retain the full history of indexer keys on GPU, limiting decode concurrency. Prefill's temporary memory demand can further delay admission even when enough memory remains for the same request's subsequent decode. We present LeanSparse, a serving system that reduces both decode's persistent memory use and prefill's temporary memory demand. Our design exploits two observations: prefill consumes historical data one layer at a time, and decode can identify promising candidates before comparing their original indexer scores. During prefill, LeanSparse streams layer histories through two reusable GPU buffers, overlapping next-layer loading with ongoing model computation. During decode, coarse-to-fine indexing searches compact representations and retrieves original indexer keys only for candidate refinement. Separate indexer key and KV caches reuse data across generation steps. We implement LeanSparse in SGLang and evaluate it on DeepSeek-V4-Flash. At 950K context length, LeanSparse supports 2.95$\times$ as many concurrent requests and achieves 2.14$\times$ decode throughput over HiSparse. On multi-turn Open-SWE traces, it reduces median time to first token from 21.6 to 0.6s and median session completion time by 17.7% relative to ECHO.

cs.DC↗

arXiv2026

Strong $\mathfrak{sl}_2$ Actions and Symmetric Chain Decompositions

We introduce a notion of a strong $\mathfrak{sl}_2$-action on a poset, and conjecture the existence of such an action is equivalent to the existence of a symmetric chain decomposition of that poset. We explain the intuition behind the conjecture, go through a number of explicit examples, and prove some special cases and relaxations of the conjecture. Then we explain a computational experiment where we tested the conjecture on millions of randomly constructed rank-symmetric, rank-unimodal, and strongly Sperner posets.

math.CO↗

arXiv2026

RR-Evict: Fine-Grained Prefix Cache Eviction beyond LRU for Agentic LLM Serving

LLM-based agents execute long-horizon tasks through repeated model calls interleaved with tool execution and user interaction. As each call extends the history accumulated in previous turns, prefix caching avoids repeated prefill of the agent's entire context. However, the aggregate cache footprint grows with context length and concurrency, forcing serving systems to reclaim cached KV tensors. We identify recency synchronization: accesses to an agent's cached history refresh its cache nodes together, causing least recently used (LRU) eviction to concentrate on a few agents' private histories. When a fully evicted agent returns, it must recompute nearly its entire accumulated context, producing a large time-to-first-token (TTFT) outlier even if most other requests retain substantial cache reuse. We present RR-EVICT, a fine-grained prefix-cache eviction strategy that distributes reclamation across idle agent trajectories. RR-EVICT visits agents in round-robin order and evicts a tail chunk from each, preserving reusable prefixes for more agents when capacity remains for idle state. Returning agents can reuse these partial histories and recompute smaller missing suffixes. The policy requires no prediction of future arrivals or tool latency. We implement RR-EVICT in SGLang and evaluate conversational and coding-agent workloads under colocated and prefill-decode-disaggregated serving. Compared with LRU, RR-EVICT reduces P99 TTFT by up to 75.4% and P99 uncached prompt tokens by up to 65.7%.

cs.DC↗

arXiv2026

AgentReplay: Token-Wise Trace Replay Is Essential for Fair Serving System Performance Benchmarking

LLM-based agents execute multi-turn workflows with interleaved model inference and tool calls, making efficient serving increasingly important. However, evaluating serving optimizations is challenging because identical tasks can produce different execution trajectories. Changes in generated tokens can alter subsequent prompts, tool calls, and reasoning turns, making it difficult to distinguish system improvements from workload variation. Greedy decoding does not guarantee identical outputs, while replaying only sequence lengths loses token information that affects prefix caching and mixture-of-experts (MoE) routing. To address these problems, we propose AgentReplay, a configurable trace record-and-replay framework for agent serving. AgentReplay records input/output tokens, expert selections, request dependencies, and tool durations. During replay, it forces the recorded output tokens while performing normal autoregressive computation, with optional controls for MoE expert selection and tool delays. This allows different systems to execute the same recorded workload while retaining their own batching, scheduling, and parallelization decisions. We further separate trajectory generation from performance evaluation, enabling compatible smaller models to replay long-horizon traces collected with more capable models. Our experiments show that token-wise replay in AgentReplay effectively eliminates workload variation that greedy decoding and length-wise replay cannot avoid, enabling fairer performance comparisons across serving configurations.

cs.DC↗

arXiv2022

Visual anomaly detection in video by variational autoencoder

Video anomalies detection is the intersection of anomaly detection and visual intelligence. It has commercial applications in surveillance, security, self-driving cars and crop monitoring. Videos can capture a variety of anomalies. Due to efforts needed to label training data, unsupervised approaches to train anomaly detection models for videos is more practical An autoencoder is a neural network that is trained to recreate its input using latent representation of input also called a bottleneck layer. Variational autoencoder uses distribution (mean and variance) as compared to latent vector as bottleneck layer and can have better regularization effect. In this paper we have demonstrated comparison between performance of convolutional LSTM versus a variation convolutional LSTM autoencoder

cs.CV↗

arXiv2017

Rapid point-of-care Hemoglobin measurement through low-cost optics and Convolutional Neural Network based validation

A low-cost, robust, and simple mechanism to measure hemoglobin would play a critical role in the modern health infrastructure. Consistent sample acquisition has been a long-standing technical hurdle for photometer-based portable hemoglobin detectors which rely on micro cuvettes and dry chemistry. Any particulates (e.g. intact red blood cells (RBCs), microbubbles, etc.) in a cuvette's sensing area drastically impact optical absorption profile, and commercial hemoglobinometers lack the ability to automatically detect faulty samples. We present the ground-up development of a portable, low-cost and open platform with equivalent accuracy to medical-grade devices, with the addition of CNN-based image processing for rapid sample viability prechecks. The developed platform has demonstrated precision to the nearest $0.18[g/dL]$ of hemoglobin, an R^2 = 0.945 correlation to hemoglobin absorption curves reported in literature, and a 97% detection accuracy of poorly-prepared samples. We see the developed hemoglobin device/ML platform having massive implications in rural medicine, and consider it an excellent springboard for robust deep learning optical spectroscopy: a currently untapped source of data for detection of countless analytes.

physics.med-ph↗