Show HN: ElasticMM – 4.2 倍加速多模态 LLM 服务 (NeurIPS 2025 口头报告)
1 分•作者: PaperWeekly•5 天前
ElasticMM 是一个新发布的开源服务系统，专为现代多模态大型语言模型（MLLM）设计。该研究成果被选为 NeurIPS 2025 的口头报告。
与主要针对纯文本工作负载优化的现有服务栈（如 vLLM）不同，ElasticMM 引入了弹性多模态并行（EMP），这是一种新的执行范式，可在不同的推理阶段和模态之间调整并行度。
论文的主要发现：
* TTFT（首次令牌时间）降低高达 4.2 倍
* 在混合多模态工作负载下，吞吐量提高 3.2 倍至 4.5 倍
* 模态感知调度、弹性阶段划分、统一前缀缓存和非阻塞编码
论文（OpenReview PDF）：
[https://openreview.net/pdf?id=Zd6VyjmN1S](https://openreview.net/pdf?id=Zd6VyjmN1S)
GitHub 仓库：
[https://github.com/hpdps-group/ElasticMM](https://github.com/hpdps-group/ElasticMM)
很想听听 HN 社区的看法，特别是那些构建 LLM/MLLM 推理栈或处理生产环境中多模态服务的用户。
查看原文
ElasticMM is a newly released open-source serving system designed for modern multimodal large language models (MLLMs). The work was selected as an Oral presentation at NeurIPS 2025.<p>Unlike existing serving stacks such as vLLM—which are primarily optimized for text-only workloads—ElasticMM introduces Elastic Multimodal Parallelism (EMP), a new execution paradigm that adapts parallelism across different inference stages and modalities.<p>Key findings from the paper:<p>Up to 4.2× reduction in TTFT<p>3.2×–4.5× higher throughput under mixed multimodal workloads<p>Modality-aware scheduling, elastic stage partitioning, unified prefix caching, and non-blocking encoding<p>Paper (OpenReview PDF):
<a href="https://openreview.net/pdf?id=Zd6VyjmN1S" rel="nofollow">https://openreview.net/pdf?id=Zd6VyjmN1S</a><p>GitHub repo:
<a href="https://github.com/hpdps-group/ElasticMM" rel="nofollow">https://github.com/hpdps-group/ElasticMM</a><p>Curious to hear what the HN community thinks, especially from those building LLM/MLLM inference stacks or dealing with multimodal serving in production.