Salvato in:
Dettagli Bibliografici
Autori principali: Oncescu, Costin-Andrei, Wu, Qingyang, Chung, Wai Tong, Wu, Robert, Gopal, Bryan, Wang, Junxiong, Dao, Tri, Athiwaratkun, Ben
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2511.02237
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908626674778112
author Oncescu, Costin-Andrei
Wu, Qingyang
Chung, Wai Tong
Wu, Robert
Gopal, Bryan
Wang, Junxiong
Dao, Tri
Athiwaratkun, Ben
author_facet Oncescu, Costin-Andrei
Wu, Qingyang
Chung, Wai Tong
Wu, Robert
Gopal, Bryan
Wang, Junxiong
Dao, Tri
Athiwaratkun, Ben
contents An increasing number of LLMs employ Mixture-of-Experts (MoE) architectures where the feed-forward layer is replaced by a pool of experts and each token only activates a small subset of them. During autoregressive generation, these models often enter a memory-bound regime even for moderate batch sizes because the average expert load grows more slowly than in an equivalent dense feedforward layer. Consequently, MoE latency is governed by the number of activated experts. We introduce a framework for dynamically re-routing token-to-expert mapping to lower this number (and thus, the decode latency) while preserving a comparable quality. Our best results use a batch-aware routing that works by having tokens piggyback experts that have already been loaded into memory due to being crucial to other tokens within the same batch. Empirically, we evaluate our method on the Qwen3-30B and Qwen3-235B models with a batch size of $16$. Without any statistically significant loss in accuracy, our approach achieves latency reductions of $39\%$ and $15\%$ in the MoE layer decode latency, respectively.
format Preprint
id arxiv_https___arxiv_org_abs_2511_02237
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Opportunistic Expert Activation: Batch-Aware Expert Routing for Faster Decode Without Retraining
Oncescu, Costin-Andrei
Wu, Qingyang
Chung, Wai Tong
Wu, Robert
Gopal, Bryan
Wang, Junxiong
Dao, Tri
Athiwaratkun, Ben
Machine Learning
An increasing number of LLMs employ Mixture-of-Experts (MoE) architectures where the feed-forward layer is replaced by a pool of experts and each token only activates a small subset of them. During autoregressive generation, these models often enter a memory-bound regime even for moderate batch sizes because the average expert load grows more slowly than in an equivalent dense feedforward layer. Consequently, MoE latency is governed by the number of activated experts. We introduce a framework for dynamically re-routing token-to-expert mapping to lower this number (and thus, the decode latency) while preserving a comparable quality. Our best results use a batch-aware routing that works by having tokens piggyback experts that have already been loaded into memory due to being crucial to other tokens within the same batch. Empirically, we evaluate our method on the Qwen3-30B and Qwen3-235B models with a batch size of $16$. Without any statistically significant loss in accuracy, our approach achieves latency reductions of $39\%$ and $15\%$ in the MoE layer decode latency, respectively.
title Opportunistic Expert Activation: Batch-Aware Expert Routing for Faster Decode Without Retraining
topic Machine Learning
url https://arxiv.org/abs/2511.02237