arXiv subjects

Shannon Wang

Publications and source records attributed to Shannon Wang.

5 recordsLinked to original sources

arXiv2026

DeaMoE: Efficient MoE Structure for Fast Small-Batch Decoding

Mixture-of-Experts (MoE) models have been widely adopted in real-time interactive applications such as coding assistants, real-time audio-video interaction systems. To meet the extremely low response latency requirements of these scenarios, practitioners commonly employ small-batch decoding, under which MoE inference becomes memory-bound and is severely bottlenecked by expert weight loading. However, this bottleneck has received limited attention, and existing solutions such as post-training weight compression or fine-grained expert design during pre-training either degrade model accuracy or introduce additional computation and communication overhead. To tackle this issue, we propose DeaMoE, a decoding-efficient MoE architecture, in which the experts are grouped into several departments, and the experts belonging to the same department share most parameters since they come from the same professional field, and additionally each expert contains a few private parameters to reflect its uniqueness. Moreover, we design customized two-stage routing strategy for DeaMoE to avoid redundant loading, under which DeaMoE greatly improves the efficiency during LLM decoding. Compared with vanilla MoE, DeaMoE reduces per-step loaded weights by up to 50.9% and achieves up to 1.33 end-to-end TPOT speedup for the pre-trained 7B model on A40, and up to 2.00x and 1.97x peak speedup for DeepSeek-V3 on A40 and H100 in microbenchmarks.

cs.LG↗

arXiv2024

Multi-matrix correlators and localization

We study generating functions of $\frac{1}{4}$-BPS states in $\mathcal{N}=4$ super Yang-Mills at finite $N$ by attempting to generalize the Harish-Chandra-Itzykson-Zuber integral to multiple commuting matrices. This allows us to compute the overlaps of two or more generating functions; such calculations arise in the computation of two-point correlators in the free-field limit. We discuss the four-matrix HCIZ integral in the $U(2)$ context and lay out a prescription for finding a more general formula for $N>2$. We then discuss its connections with the restricted Schur polynomial operator basis. Our results generalize readily to arbitrary numbers of matrices, opening up the opportunity to study more generic BPS operators.

hep-th↗

arXiv2022

Giant Gravitons, Harish-Chandra integrals, and BPS states in symplectic and orthogonal $\mathcal{N}$= 4 SYM

We find generating functions for half BPS correlators in $\mathcal{N}=4$ SYM theories with gauge groups $Sp(2N)$, $SO(2N+1)$, and $SO(2N)$ by computing the norms of a class of BPS coherent states. These coherent states are built from operators involving Harish-Chandra integrals. Such operators have an interpretation as localized giant gravitons in the bulk of anti-de-Sitter space. This extends the analysis of \cite{Berenstein:2022srd} to $Sp(2N)$, $SO(2N+1)$, and $SO(2N)$ gauge theories. We show that we may use ordinary Schur functions as a basis for the sector of states with no cross-caps in these theories. This is consistent with the construction of these theories as orientifold projections of an $SU(2N)$ theory. We make note of some relations between the symmetric functions that appear in the expansion of these coherent states and symplectic Schur functions. We also comment on some connections to Schubert calculus and Gromov-Witten invariants, which suggest that the Harish-Chandra integral may be extended to such problems.

hep-th↗

arXiv2022

BPS coherent states and localization

We introduce coherent states averaged over a gauge group action to study correlators of half BPS states in ${\cal N}=4 $ SYM theory. The overlaps of these averaged coherent states are a generating function of correlators and can be written in terms of the Harish-Chandra-Itzykzon-Zuber (HCIZ) integral. We show that this formula immediately leads to a computation of the normalization of two point functions in terms of characters obtained originally in the work of Corley, Jevicki and Ramgoolam. We also find various generalizations for $A_{n-1}$ quivers that follow directly from other solvable integrals over unitary groups. All of these can be computed using localization methods. When we promote the parameters of the generating function to collective coordinates, there is a dominant saddle that controls the effective action of these coherent states in the regime where they describe single AdS giant gravitons. We also discuss how to add open strings to this formulation. These will produce calculations that rely on correlators of matrix components of unitaries in the ensemble that is determined by the HCIZ integral to determine anomalous dimensions. We also discuss how sphere giants arise from Grassman integrals, how one gets a dominant saddle and how open strings are added in that case. The fact that there is a dominant saddle helps to understand how a $1/N$ expansion arises for open strings. We generalize the coherent state idea to study $1/4$ and $1/8$ BPS states as more general integrals over unitary groups.

hep-th↗

arXiv2020

Probing phase transitions of holographic entanglement entropy with fixed area states

Recent results suggest that new corrections to holographic entanglement entropy should arise near phase transitions of the associated Ryu-Takayanagi (RT) surface. We study such corrections by decomposing the bulk state into fixed-area states and conjecturing that a certain `diagonal approximation' will hold. In terms of the bulk Newton constant $G$, this yields a correction of order $O(G^{-1/2})$ near such transitions, which is in particular larger than generic corrections from the entanglement of bulk quantum fields. However, the correction becomes exponentially suppressed away from the transition. The net effect is to make the entanglement a smooth function of all parameters, turning the RT `phase transition' into a crossover already at this level of analysis. We illustrate this effect with explicit calculations (again assuming our diagonal approximation) for boundary regions given by a pair of disconnected intervals on the boundary of the AdS$_3$ vacuum and for a single interval on the boundary of the BTZ black hole. In a natural large-volume limit where our diagonal approximation clearly holds, this second example verifies that our results agree with general predictions made by Murthy and Srednicki in the context of chaotic many-body systems. As a further check on our conjectured diagonal approximation, we show that it also reproduces the $O(G^{-1/2})$ correction found Penington et al for an analogous quantum RT transition. Our explicit computations also illustrate the cutoff-dependence of fluctuations in RT-areas.

hep-th↗