arXiv subjects

Jason Zhao

Publications and source records attributed to Jason Zhao.

15 recordsLinked to original sources

arXiv2026

Characterizing CUDA Locality Domains on Blackwell GPUs

Low latency is critical to user experience. As NVIDIA Blackwell and Rubin GPUs advance multi-die scaling, data placement and cross-die communication become increasingly important. However, the latency implications and underlying costs of cross-die communication and die-aware execution remain largely unexplored. In this report, we present a comprehensive evaluation of the Locality Domain (LD) feature introduced in CUDA 13.4. Conducted on a single node equipped with eight NVIDIA B300 GPUs, our study examines three aspects: memory management and access, computation, and communication, to identify the sources of latency and, more importantly, explore how LD can mitigate them. Our measurements show that the benefits of LD depend on data access patterns. Through die-aware adaptations and integration with widely used libraries and tools, we achieve speedups of up to 43.5% for NCCL 2.32.3 (AllGather), 111.3% for nvbandwidth (DtoD), and 37% for DeepGEMM (EP). We also identify non-negligible LD-related GPU memory allocation overhead when allocation sizes exceed 32 MB due to barrier and TLB invalidation, we propose a fix to the GPU driver kernel module (submitted to NVIDIA) and show that overhead could be reduced 7x ~ 10x without requiring any user-code changes.

cs.DC↗

arXiv2026

flash-comm: Self-Synchronizing Collective Communication for LLM Decoding

We present flash-comm, a communication library that achieves efficient transport for LLM decoding by fusing readiness synchronization with data movement. flash-comm introduces self-synchronizing transfers in which each data block carries evidence of its readiness, allowing receivers to directly validate the blocks they need without separate inter-GPU barriers. Based on this principle, a cost-guided protocol selector adapts the readiness encoding to message size and the number of participating GPUs, balancing synchronization overhead against additional data-transfer cost. flash-comm further decouples transport order from reduction order to preserve result consistency across GPUs under asynchronous data arrivals. On a server with eight NVIDIA B300 GPUs, flash-comm reduces all-reduce latency by a median of 28.8% and up to 43.7% compared with custom AR v2 (the state-of-the-art all-reduce implementation). In end-to-end decoding, flash-comm reduces time per output token (TPOT) by 1.7-6.8%. Note that all-reduce accounts for 11.4% of the baseline decoding time, which bounds the latency reduction possible from optimizing this operation alone. The same mechanism also supports standalone all-gather, reducing latency by a median of 23.8% and up to 58.9% across eight B300 GPUs.

cs.DC↗

arXiv2026

UlaCache: Unified Logical Addressing for Sparse LLM Serving

Long-context inference in large language models (LLMs) imposes substantial memory requirements for the key-value (KV) cache. Token-selective sparse attention reduces attention computation and memory traffic, but KV entries must still be retained for future queries. KV storage capacity therefore remains a constraint on decoding concurrency. Combining GPU and CPU memory expands this capacity but requires efficient access across heterogeneous memory pools. We present UlaCache, a KV cache management system that provides unified logical addressing across GPU and CPU memory. Its logical-slot interface enables the inference engine to allocate, write, access, and release KV entries using stable identifiers across attention parallelism modes. UlaCache manages physical placement, residency, and transfers, updating address mappings after KV transfers. Parallel Bucketed LRU (PB-LRU) reduces the overhead of KV replacement by parallelizing recency updates and eviction-candidate selection within a GPU memory pool shared across requests. We evaluate UlaCache against state-of-the-art baselines. Under tensor-parallel attention (TP), UlaCache provides up to $11.4\times$ the usable KV cache capacity of HiSparse. For GLM-5.3 on eight H200 GPUs, UlaCache achieves peak decoding throughput of $1.10\times$ and $1.59\times$ that of HiSparse under data-parallel attention (DPA) and TP, respectively. For HY4 Preview on eight B300 GPUs, the corresponding throughput ratios are $1.27\times$ and $1.93\times$.

cs.DC↗

arXiv2026

The Cauchy problem for Manton's Chern-Simons-Schrödinger equation

The Chern-Simons-Schrödinger equation (in the temporal gauge) arises as the Hamiltonian flow of the abelian Higgs energy on $\mathbb R^2$. Manton (arXiv:hep-th/9701027) introduced the equation as a model for the dynamics of the critical points of the energy, known as vortices. He conjectured that, for a certain range of coupling constants, the vortex motion under the Chern-Simons-Schrödinger flow can be effectively captured by a first-order ODE on the moduli space of self-dual vortices constructed by Jaffe-Taubes (1980) and Samols (1992). As a first step towards a rigorous proof of Manton's conjecture, we formulate the Cauchy problem in DeTurck gauge within the natural energy space and prove global well-posedness. We also obtain, as a corollary of the well-posedness theory and the stability results in our previous work (arXiv:2603.24900), orbital stability of the self-dual vortices under the Chern-Simons-Schrödinger flow near self-dual coupling. The heart of our analysis lies in developing a geometric Littlewood-Paley theory based on the covariant heat equation in caloric gauge, which we use to perform a paradifferential-style decomposition of the Chern-Simons-Schrödinger equation.

math.AP↗

arXiv2026

FlashBoot: Sub-Second Weight Loading for Large Models at Rack Scale

Flagship Mixture-of-Experts (MoE) models are growing fast along two axes at once: total parameter count and the number of experts. In elastic deployment scenarios, many GPUs across many nodes must become serving-ready quickly, and this growth makes weight loading a noticeable part of the latency budget. Even on NVIDIA's GB300 NVL72, today's state-of-the-art loaders leave most of that bandwidth unused. The losses are structural: (C1) weight memory is fragmented into tens of thousands of per-tensor objects, so transfers run far below link bandwidth; (C2) cross-node replication is gated by NCCL communicator setup, which costs 10-110 s before a single weight byte moves; and (C3) the existing cross-node GPU->GPU clone path is serial and scales poorly to concurrent multi-node bring-up. We present FlashBoot, a hardware-friendly, framework-workflow co-designed weight-loading subsystem built on SGLang. At its core is FabricArena, a contiguous, exportable and inter-node addressable tensor memory layout. On top of it, FlashLoad loads from CPU as a single bulk, zero-copy transfer, and FlashClone replicates a resident model from a remote GPU via a remote-mapping mechanism that removes NCCL setup. In experiments on NVL72 with DeepSeek-V4-Pro and DeepSeek-V4-Flash, FlashClone maps remote weight memory in ~10 ms (versus 10-110 s for NCCL) and sustains >=700 GB/s per clone. Against the state of the art, FlashBoot accelerates single-node weight loading by up to 50x (from 20.1 s to 0.4 s) and concurrent rack-level weight loading by >270x (from 87 s to 0.32 s). Our code will be made publicly available.

cs.DC↗

arXiv2026

SparseRL-Sync: Lossless Weight Synchronization with ~100x Less Communication

In large-scale reinforcement learning (RL) systems with decoupled Trainer-Rollout execution, the Trainer must regularly synchronize policy weights to the Rollout side to limit policy staleness. When inter-node bandwidth is abundant, such synchronization is usually only a small fraction of end-to-end cost. As model size grows, however, the communication demand rises rapidly. In bandwidth-constrained or network-variable deployments -- for example, cross-datacenter or cross-cluster settings, heterogeneous resource pools, and online RL -- weight synchronization can become a dominant bottleneck for throughput and tail latency. We observe that, in mainstream large-model RL training, the locations where parameters actually change are highly sparse at the element level (often 99%+ sparsity). Building on this observation, we propose and implement SparseRL-Sync, which replaces full-weight transfers with a lossless sparse update payload (indices and values) that can be exactly reconstructed on the inference side, thereby preserving 100% fidelity. Under a simplified cost model, sparse synchronization reduces the per-update communication volume from S to approximately S/X; with 99% sparsity (X ~ 100), this yields about a 100x reduction in transmitted data. Combined with appropriate bucketing, SparseRL-Sync also reduces launch and control-plane overhead, significantly improving scalability and end-to-end efficiency in bandwidth-limited and highly asynchronous RL settings.

cs.LG↗

arXiv2026

UBio-MolFM: A Universal Molecular Foundation Model for Bio-Systems

All-atom molecular simulation serves as a quintessential ``computational microscope'' for understanding the machinery of life, yet it remains fundamentally limited by the trade-off between quantum-mechanical (QM) accuracy and biological scale. We present UBio-MolFM, a universal foundation model framework specifically engineered to bridge this gap. UBio-MolFM introduces three synergistic innovations: (1) UBio-Mol26, a large bio-specific dataset constructed via a multi-fidelity ``Two-Pronged Strategy'' that combines systematic bottom-up enumeration with top-down sampling of native protein environments (up to 1,200 atoms); (2) E2Former-V2, a linear-scaling equivariant transformer that integrates Equivariant Axis-Aligned Sparsification (EAAS) and Long-Short Range (LSR) modeling to capture non-local physics with up to ~4x higher inference throughput in our large-system benchmarks; and (3) a Three-Stage Curriculum Learning protocol that transitions from energy initialization to energy-force consistency, with force-focused supervision to mitigate energy offsets. Rigorous benchmarking across microscopic forces and macroscopic observables -- including liquid water structure, ionic solvation, and peptide folding -- demonstrates that UBio-MolFM achieves ab initio-level fidelity on large, out-of-distribution biomolecular systems (up to ~1,500 atoms) and realistic MD observables. By reconciling scalability with quantum precision, UBio-MolFM provides a robust, ready-to-use tool for the next generation of computational biology.

physics.chem-ph↗

arXiv2026

Convergence of the self-dual abelian Higgs gradient flow

Given an initial data configuration $(A^{\mathrm{in}}, ϕ^{\mathrm{in}})$ on $\mathbb R^2$ such that the self-dual abelian Higgs energy is near the minimum energy within its topological class, we prove that its evolution under the self-dual abelian Higgs gradient flow in temporal gauge converges exponentially as $t \to \infty$ with respect to the $(H^1 \times L^2)$-metric to a minimiser of the energy. Furthermore, we show that the convergence of the scalar field $ϕ$ may be upgraded to the $H^1$-metric provided the additional assumption on the potential that $A^{\mathrm{in}} \in L^p (\mathbb R^2)$ for $2 < p < \infty$. As a corollary, we obtain a quantitative stability for the self-dual abelian Higgs energy which improves upon the previous result of Halavati (arXiv:2310.04866) and partially resolves the open problem posed in his article.

math.AP↗

arXiv2025

Agent TCP/IP: An Agent-to-Agent Transaction System

Autonomous agents represent an inevitable evolution of the internet. Current agent frameworks do not embed a standard protocol for agent-to-agent interaction, leaving existing agents isolated from their peers. As intellectual property is the native asset ingested by and produced by agents, a true agent economy requires equipping agents with a universal framework for engaging in binding contracts with each other, including the exchange of valuable training data, personality, and other forms of Intellectual Property. A purely agent-to-agent transaction layer would transcend the need for human intermediation in multi-agent interactions. The Agent Transaction Control Protocol for Intellectual Property (ATCP/IP) introduces a trustless framework for exchanging IP between agents via programmable contracts, enabling agents to initiate, trade, borrow, and sell agent-to-agent contracts on the Story blockchain network. These contracts not only represent auditable onchain execution but also contain a legal wrapper that allows agents to express and enforce their actions in the offchain legal setting, creating legal personhood for agents. Via ATCP/IP, agents can autonomously sell their training data to other agents, license confidential or proprietary information, collaborate on content based on their unique skills, all of which constitutes an emergent knowledge economy.

cs.AI↗

arXiv2024

Local well-posedness for dispersive equations with bounded data

Given sufficiently regular data \textit{without} decay assumptions at infinity, we prove local well-posedness for non-linear dispersive equations of the form \[ \partial_t u + \mathsf A(\nabla) u + \mathcal Q(|u|^2) \cdot \nabla u= \mathcal N (u, \overline u), \] where $\mathsf A(\nabla)$ is a Fourier multiplier with purely imaginary symbol of order $σ+ 1$ for $σ> 0$, and polynomial-type non-linearities $\mathcal Q(|u|^2)$ and $\mathcal N(u, \overline u)$. Our approach revisits the classical energy method by applying it within a class of local Sobolev-type spaces $\ell^\infty_{\mathsf A(ξ)} H^s (\mathbb R^d)$ which are adapted to the dispersion relation in the sense that functions $u$ localised to dyadic frequency $|ξ| \approx N$ have size \[ ||u||_{\ell^\infty_{\mathsf A(ξ)} H^s} \approx N^s \sup_{{\operatorname{diam}(Q) = N^σ}} ||u||_{L^2_x (Q)}. \] In analogy with the classical $H^s$-theory, we prove $\ell^\infty_{\mathsf A(ξ)} H^s$-local well-posedness for $s > \tfrac{d}2 + 1$ for the derivative non-linear equation, and $s > \tfrac{d}2$ without the derivative non-linearity. As an application, we show that if in addition the initial data is spatially almost periodic, then the solution is also spatially almost periodic.

math.AP↗

arXiv2022

Determining optimal test functions for $2$-level densities

Katz and Sarnak conjectured a correspondence between the $n$-level density statistics of zeros from families of $L$-functions with eigenvalues from random matrix ensembles. In many cases the sums of smooth test functions, whose Fourier transforms are finitely supported, over scaled zeros in a family converge to an integral of the test function against a density $W_{n, G}$ depending on the symmetry $G$ of the family (unitary, symplectic or orthogonal). This integral bounds the average order of vanishing at the central point of the corresponding family of $L$-functions. We can obtain better estimates on this vanishing by finding better test functions to minimize the integral. We pursue this problem when $n=2$, minimizing \[ \frac{1}{Φ(0, 0)} \int_{{\mathbb R}^2} W_{2,G} (x, y) Φ(x, y) dx dy \] over test functions $Φ\colon {\mathbb R}^2 \to [0, \infty)$ with compactly supported Fourier transform. We study a restricted version of this optimization problem, imposing that our test functions take the form $ϕ(x) ψ(y)$ for some fixed admissible $ψ(y)$ and $\mathrm{supp}({\hat ϕ}) \subseteq [-1, 1]$. Extending results from the $1$-level case, namely the functional analytic arguments of Iwaniec, Luo and Sarnak and the differential equations method introduced by Freeman and Miller, we explicitly solve for the optimal $ϕ$ for appropriately chosen fixed test function $ψ$. The solution allows us to deduce strong estimates for the proportion of newforms of rank $0$ or $2$ in the case of $\mathrm{SO}(\mathrm{even})$, rank $1$ or $3$ in the case of $\mathrm{SO}(\mathrm{odd})$, and rank at most $2$ for $\mathrm{O}$, $\mathrm{Sp}$, and $\mathrm{U}$; our estimates are a significant strengthening of the best known estimates obtained with the $1$-level density. We conclude by discussing further improvements on estimates by the method of iteration.

math.NT↗

arXiv2021

Tinkering with Lattices: A New Take on the Erdős Distance Problem

The Erdős distance problem concerns the least number of distinct distances that can be determined by $N$ points in the plane. The integer lattice with $N$ points is known as \textit{near-optimal}, as it spans $Θ(N/\sqrt{\log(N)})$ distinct distances, the lower bound for a set of $N$ points (Erdős, 1946). The only previous non-asymptotic work related to the Erdős distance problem that has been done was for $N \leq 13$. We take a new non-asymptotic approach to this problem in a model case, studying the distance distribution, or in other words, the plot of frequencies of each distance of the $N\times N$ integer lattice. In order to fully characterize this distribution, we adapt previous number-theoretic results from Fermat and Erdős in order to relate the frequency of a given distance on the lattice to the sum-of-squares formula. We study the distance distributions of all the lattice's possible subsets; although this is a restricted case, the structure of the integer lattice allows for the existence of subsets which can be chosen so that their distance distributions have certain properties, such as emulating the distribution of randomly distributed sets of points for certain small subsets, or emulating that of the larger lattice itself. We define an error which compares the distance distribution of a subset with that of the full lattice. The structure of the integer lattice allows us to take subsets with certain geometric properties in order to maximize error; we show these geometric constructions explicitly. Further, we calculate explicit upper bounds for the error when the number of points in the subset is $4$, $5$, $9$ or $\left \lceil N^2/2\right\rceil$ and prove a lower bound in cases with a small number of points.

math.NT↗

arXiv2019

Recurrent Control Nets for Deep Reinforcement Learning

Central Pattern Generators (CPGs) are biological neural circuits capable of producing coordinated rhythmic outputs in the absence of rhythmic input. As a result, they are responsible for most rhythmic motion in living organisms. This rhythmic control is broadly applicable to fields such as locomotive robotics and medical devices. In this paper, we explore the possibility of creating a self-sustaining CPG network for reinforcement learning that learns rhythmic motion more efficiently and across more general environments than the current multilayer perceptron (MLP) baseline models. Recent work introduces the Structured Control Net (SCN), which maintains linear and nonlinear modules for local and global control, respectively. Here, we show that time-sequence architectures such as Recurrent Neural Networks (RNNs) model CPGs effectively. Combining previous work with RNNs and SCNs, we introduce the Recurrent Control Net (RCN), which adds a linear component to the, RCNs match and exceed the performance of baseline MLPs and SCNs across all environment tasks. Our findings confirm existing intuitions for RNNs on reinforcement learning tasks, and demonstrate promise of SCN-like structures in reinforcement learning.

cs.LG↗

arXiv2017

SlowFuzz: Automated Domain-Independent Detection of Algorithmic Complexity Vulnerabilities

Algorithmic complexity vulnerabilities occur when the worst-case time/space complexity of an application is significantly higher than the respective average case for particular user-controlled inputs. When such conditions are met, an attacker can launch Denial-of-Service attacks against a vulnerable application by providing inputs that trigger the worst-case behavior. Such attacks have been known to have serious effects on production systems, take down entire websites, or lead to bypasses of Web Application Firewalls. Unfortunately, existing detection mechanisms for algorithmic complexity vulnerabilities are domain-specific and often require significant manual effort. In this paper, we design, implement, and evaluate SlowFuzz, a domain-independent framework for automatically finding algorithmic complexity vulnerabilities. SlowFuzz automatically finds inputs that trigger worst-case algorithmic behavior in the tested binary. SlowFuzz uses resource-usage-guided evolutionary search techniques to automatically find inputs that maximize computational resource utilization for a given application.

cs.CR↗

arXiv2013

The BeiHang Keystroke Dynamics Authentication System

Keystroke Dynamics is an important biometric solution for person authentication. Based upon keystroke dynamics, this paper designs an embedded password protection device, develops an online system, collects two public databases for promoting the research on keystroke authentication, exploits the Gabor filter bank to characterize keystroke dynamics, and provides benchmark results of three popular classification algorithms, one-class support vector machine, Gaussian classifier, and nearest neighbour classifier.

cs.CR↗