---
id: 20261119-T0-15
title: "MoE扩散解码的隐藏成本：采样器拖慢了生成"
title_en: "The Hidden Cost of MoE Diffusion Decoding: The Sampler"
url: https://ai.daily.yangsir.net/daily/20261119-T0-15
issue_date: 2026-11-19
publish_date: 2026-11-18T05:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2611.21852
---

# MoE扩散解码的隐藏成本：采样器拖慢了生成

论文剖析掩码扩散语言模型（dLLM）的解码开销。这类模型通过迭代去噪整块token画布来生成，每次前向传播都要在整个画布乘词表的logits张量上跑逐token采样和停止判断流水线。作者对单用户解码进行了性能剖析，定位到采样器环节的成本，并提出修复方案。对推理部署而言，这指向一个常被忽视的延迟来源。

## English Version

**The Hidden Cost of MoE Diffusion Decoding: The Sampler**

The paper profiles decoding overhead in masked diffusion language models (dLLMs). These models generate by iteratively denoising a full canvas of tokens, so every forward pass runs a per-token sampling and stopping pipeline over the entire canvas-by-vocabulary logits tensor. The authors profile single-user decoding, identify the sampler as a cost center, and propose a fix. For inference deployment, this points to an often-overlooked source of latency.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2611.21852)

**详情页**：https://ai.daily.yangsir.net/daily/20261119-T0-15

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*