Searcharxiv⌕ Search

arXiv subjects

Isaac Zhu

Publications and source records attributed to Isaac Zhu.

5 recordsLinked to original sources

The structure of the pion at large $N_c$

We compute the pion quark momentum fraction in the large-$N_c$ limit using the Twisted Eguchi-Kawai (TEK) reduced model. Gauge configurations for different values of the lattice spacing are generated with the pure-gauge TEK action, while the pion correlators are computed for different quark masses with maximally twisted mass fermions. The bare matrix element corresponding to the quark momentum fraction is obtained using the Feynman-Hellmann method, and we use perturbative renormalization factors to connect to the $\overline{\mathrm{MS}}$ scheme at $2\,\mathrm{GeV}$. A global analysis at several lattice spacings, pion masses, and values of $N_c$ yields $\langle x\rangle_q^{\overline{\mathrm{MS}}}(2\,\mathrm{GeV}) = 0.267(105)$ at the physical pion mass in the large-$N_c$ limit. This value is significantly smaller than existing $N_c=3$ determinations from phenomenology and lattice QCD, indicating that, in the large-$N_c$ limit, the pion's light-cone momentum is carried predominantly by gluons.

hep-lat↗

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↗

Knizhnik-Zamolodchikov equations in Deligne categories

We consider the Knizhnik-Zamolodchikov equations in Deligne Categories in the context of $(\mathfrak{gl}_m,\mathfrak{gl}_{n})$ and $(\mathfrak{so}_m,\mathfrak{so}_{2n})$ dualities. We derive integral formulas for the solutions in the first case and compute monodromy in both cases.

math.RT↗

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↗

Patterns in the Lattice Homology of Seifert Homology Spheres

In this paper, we study various homology cobordism invariants for Seifert fibered integral homology 3-spheres derived from Heegaard Floer homology. Our main tool is lattice homology, a combinatorial invariant defined by Ozsváth-Szabó and Némethi. We reprove the fact that the $d$-invariants of Seifert homology spheres $Σ(a_1,a_2,\dots,a_n)$ and $Σ(a_1,a_2,\dots,a_n+a_1a_2\cdots a_{n-1})$ are the same using an explicit understanding of the behavior of the numerical semigroup minimally generated by $a_1a_2\cdots a_n/a_i$ for $i\in[1,n]$. We also study the maximal monotone subroots of the lattice homologies, another homology cobordism invariant introduced by Dai and Manolescu. We show that the maximal monotone subroots of the lattice homologies of Seifert homology spheres $Σ(a_1,a_2,\dots,a_n)$ and $Σ(a_1,a_2,\dots,a_n+2a_1a_2\cdots a_{n-1})$ are the same.

math.GT↗