ModelsOfficial5月13日 21:35

DeepSeek 开源 DeepEP：首个面向 MoE 模型的专家并行通信库

DeepEP

Primary sourceDeepSeek: GitHub 新仓库

News

Read the original

Why we picked it

MoE 模型的通信瓶颈是训练和推理的常见痛点，DeepEP 直接解决了这个问题。做大规模 MoE 训练或推理的团队，值得集成试试。

DeepSeek 发布了 DeepEP，这是一个专为 MoE（混合专家）模型设计的高效通信库，旨在优化专家并行场景下的 GPU 通信效率。

DeepSeek 开源 DeepEP：首个面向 MoE 模型的专家并行通信库 — Image · DeepSeek: GitHub 新仓库

Body

它支持低延迟推理和高吞吐量训练，并提供了节点内和节点间的通信优化。该库还引入了低精度操作和 FP8 调度，进一步提升了性能。

对于使用 MoE 架构的团队，DeepEP 可以直接集成到现有框架中，显著减少通信开销。

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include MoE, 通信库, DeepSeek, 开源/仓库.

Primary source verification: published directly by DeepSeek: GitHub 新仓库 with first-party authority.

#MoE #通信库 #DeepSeek #开源/仓库 #GPU优化

Keep reading

Recent stories with the same tag or section

Models Builders13:57

Chinese contentZONOS3-Mini 发布：886M 参数零样本语音克隆 TTS 可实时跑在手机上

Zyphra 开源了这个 886M 参数的语音克隆模型，100M 激活量就能在手机上实时跑，克隆效果还能对标更大模型。做端侧语音应用、多语言 TTS 的开发者可以直接拿权重和训练代码上手试。

#ZONOS3-Mini #TTS #语音克隆 #开源模型

arXiv cs.AI@Gabriel Clark et al. (4)Source

Models BuildersMulti-sourceSelected05:32

Inkling 2 Small Runs on vLLM From Day Zero

Day-zero vLLM support plus single-node GB200 fit means operational cost of serving this model is far lower than typical 594B-scale deployments. Engineering teams running self-hosted inference should benchmark the 630 tok/s TP4 figure against their current MoE serving setups, since the DSpark speculator and NVFP4 experts are the levers behind it. Teams needing 1M-token multimodal context without multi-node clusters have a concrete new option to evaluate.

#Inkling 2 Small #Thinking Machines #vLLM #open-source models

Event

Official accountvLLM@vllm_projectSource

Models BuildersMulti-source76°05:25

Chinese contentvLLM 第一时间支持 Thinking Machines 的 Inkling-2-Small 模型

Thinking Machines 发了个 500B 的多模态 MoE，vLLM 发布当天就能跑，还带了 FP4 量化和投机解码优化。想自部署大模型的人可以看看集成细节，不用等闭源 API。

#Inkling-2-Small #vLLM #Thinking Machines Lab #MoE

Event

Official accountvLLM@vllm_projectSource

Models BuildersMulti-source84°05:13

Chinese contentThinking Machines 的 inkling 2 small 上线 Together AI

Mira Murati 的 Thinking Machines 把 inkling 2 small 放上了 Together AI，594B 参数只激活 17B，还有 1M 上下文。写代码、跑长任务、看图都能覆盖，推理成本却按小模型算。做 agent 或长文档处理的开发者可以直接去 Together AI 试。

#inkling 2 small #Thinking Machines #Together AI #MoE

Event

Official accountTogether AI@togethercomputeSource

Models BuildersSelected00:50

Domyn trains 1T-parameter MoE on EuroHPC Jupiter using PyTorch

The non-obvious takeaway is that per-GPU throughput gains of this magnitude came from software stack optimisation, not hardware changes, and were upstreamed to PyTorch and TorchTitan for anyone to reuse. Teams training large MoE models on GH200 clusters should review the upstream contributions for immediate throughput wins. European companies and labs with EuroHPC access now have a demonstrated open-source path to trillion-parameter training.

#Domyn #PyTorch #TorchTitan #EuroHPC