Salvato in:
Dettagli Bibliografici
Autori principali: Liu, Qingxiu, He, Cyril Y., Jiang, Hanser, Wang, Zion, Zhao, Alan, Lee, Patrick P. C.
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:https://arxiv.org/abs/2604.02715
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915969914372096
author Liu, Qingxiu
He, Cyril Y.
Jiang, Hanser
Wang, Zion
Zhao, Alan
Lee, Patrick P. C.
author_facet Liu, Qingxiu
He, Cyril Y.
Jiang, Hanser
Wang, Zion
Zhao, Alan
Lee, Patrick P. C.
contents Mixture-of-Experts (MoE) models have become a dominant paradigm for scaling large language models, but their rapidly growing parameter sizes introduce a fundamental inefficiency during inference: most expert weights remain idle in GPU memory while competing with performance-critical runtime state such as the key-value (KV) cache. Since KV cache capacity directly determines serving throughput, this mismatch leads to underutilized memory and degraded performance. In this paper, we present FluxMoE, a new MoE inference system that decouples expert parameters from persistent GPU residency. FluxMoE introduces an expert paging abstraction that treats expert weights as streamed, transient resources, materializing them on demand and evicting them immediately after use, allowing GPU memory to be preferentially allocated to throughput-critical runtime state. We implement FluxMoE atop vLLM to enable efficient MoE inference under severe memory constraints. Experimental results demonstrate that FluxMoE achieves up to 3.0$\times$ throughput gains over vLLM in memory-intensive regimes, without compromising model fidelity.
format Preprint
id arxiv_https___arxiv_org_abs_2604_02715
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving
Liu, Qingxiu
He, Cyril Y.
Jiang, Hanser
Wang, Zion
Zhao, Alan
Lee, Patrick P. C.
Machine Learning
Mixture-of-Experts (MoE) models have become a dominant paradigm for scaling large language models, but their rapidly growing parameter sizes introduce a fundamental inefficiency during inference: most expert weights remain idle in GPU memory while competing with performance-critical runtime state such as the key-value (KV) cache. Since KV cache capacity directly determines serving throughput, this mismatch leads to underutilized memory and degraded performance. In this paper, we present FluxMoE, a new MoE inference system that decouples expert parameters from persistent GPU residency. FluxMoE introduces an expert paging abstraction that treats expert weights as streamed, transient resources, materializing them on demand and evicting them immediately after use, allowing GPU memory to be preferentially allocated to throughput-critical runtime state. We implement FluxMoE atop vLLM to enable efficient MoE inference under severe memory constraints. Experimental results demonstrate that FluxMoE achieves up to 3.0$\times$ throughput gains over vLLM in memory-intensive regimes, without compromising model fidelity.
title FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving
topic Machine Learning
url https://arxiv.org/abs/2604.02715