🚨重磅！DeepSeek 发布 DeepEP V2.5 版本！又把 MoE 通信库掏出来升级了！
This is a dev post classified by Jev as AI dev tools (a launch), kept by the Dev Radar because it carries real work, not commentary.
🚨重磅！DeepSeek 发布 DeepEP V2.5 版本！又把 MoE 通信库掏出来升级了！ 升级要点如下： 1️⃣缓冲区拆分 ElasticBuffer → EPBuffer / EngramBuffer / PPBuffer / BucketBuffer，共享同一套生命周期，职责更清晰。 2️⃣先规划再分配 新增 BufferAllocator，构建缓冲区前先规划对称张量怎么占显存，少浪费、少碎片。 3️⃣批量集合通信 BucketBuffer 支持批量 all-gather、reduce-scatter、all-reduce，普通 PyTorch 张量也能走同一套会话。 4️⃣动态冗余专家 lb_prefetch_weights：算专家前用 NVLink 预取权重和量化尺度 lb_reduce_grads：反向时把冗余专家的 FP32 梯度累加回原始专家 对齐 MoonEP / UltraEP 那套专家复制思路。 5️⃣训练细节补齐 支持延迟 EP 收尾、缓存扩展布局、专家之间零填充。 6️⃣多层 Engram 可在 GPU 或 CPU 上多层存储，每层一个等待钩子。 7️⃣大扫除 完全移除 V1 API、NVSHMEM 后端和旧文档，NVSHMEM 不再是依赖。 一句话：把原来那个大而全的 ElasticBuffer 拆开、把显存规划提前、把冗余专家和集
Posted by SuSu_酥酥👅 (16.8k followers) 1 h ago · 32 likes · 2.8k views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: deepep.
More dev work like this
- Google Antigravity SDK now runs Gemma 4 26B locally on GPU via Google AI Edge LiteRT. — @0x0SojalSec
- MiniMax H3 跑视频，最磨人的有时候不是生成，而是 VAE 解码太慢。 — @bkdgiffug
- This is one of the great example of using AI agent to do your works. 🚀 — @BkashJosi
- i didn't mean for this to sound like a dunk on github, they have pulled off some… — @dexhorthy
- Hardcoded agent URLs don’t scale. This repo gives agent discovery a DNS-native path. — @DanKornas
- 👀 ICYMI: @finkd opened his keynote for Connect this afternoon with a live demo of @Muse… — @claireszhou
- Okay, my first complete benchmark of Jev with my new Decision v1 eval suite on… — @morganlinton
- Browser Use 团队基于最近爆火的 Jev 模型做了个 jev-ultrafast，主打就是快，已经狂揽了 19000+ Star。 — @GitHub_Daily
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 23.1k posts from 5k X accounts over the last 21 days, 2.7k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-24 00:59 UTC. Full method.