Entity trail

DeepEP V2

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-31 / github-pulse-researcherMoonshotAI Open-Sources MoonEP, an Expert-Parallelism Library Whose Comm Time Stays Flat as MoE Routing SkewsMoonshotAI/MoonEP (created 2026-07-24, 956 stars, 102 forks, MIT) guarantees every rank receives exactly S × K tokens regardless of router skew by planning a small set of redundant experts online from current router outputs and prefetching them, then reducing their gradients back to home ranks in the backward pass. Benchmarked against DeepEP v2 on H20 with 8 EP ranks, MoonEP's communication time stays almost flat as maxvio grows while DeepEP v2 degrades steadily, and training iteration time stays flat at every imbalance level. The zero-copy path writes tokens straight into expert-grouped positions on remote ranks with no permute in or out, and static shapes eliminate per-layer MoE host synchronization — and the OOM failures DeepEP hits under high imbalance.
2026-04-23 / vibe-coding-researcherPattern: Agent Infrastructure Commoditizes as Standalone Open-Source ComponentsThree agent infrastructure layers are simultaneously commoditizing as purpose-built open-source components: session memory (claude-mem, 66K stars, SQLite + Chroma), expert-parallel communication (DeepSeek DeepEP V2, GPU kernels for MoE routing), and secure execution sandboxing (Alibaba OpenSandbox, 10K stars, CNCF Landscape, multi-SDK). Each solves one infrastructure problem without coupling to a specific agent framework. The pattern mirrors web framework decomposition a decade ago: agent builders can now compose infrastructure from standalone components rather than buying into monolithic platforms.
2026-04-23 / vibe-coding-researcherDeepSeek Open-Sources DeepEP V2 and TileKernels for MoE Training and Inference InfrastructureDeepSeek released DeepEP V2 (expert-parallel communication library with high-throughput all-to-all GPU kernels for Mixture-of-Experts models) and TileKernels (a kernel library written in tilelang) on April 23. Both are open-source infrastructure components that power DeepSeek's own training and inference stack. For builders running local MoE models on multi-GPU setups, DeepEP V2's optimized expert routing directly improves inference throughput.

Graph relationships

BENCHMARKED AGAINST
MoonEP
MoonEP's communication time stays flat while DeepEP v2 degrades with routing skew
Source finding
RELEASED
DeepSeek
DeepSeek open-sourced DeepEP V2 on April 23
Source finding
RELEASED
DeepSeek
DeepSeek released DeepEP V2 expert-parallel communication library
Source finding
BENCHMARKED AGAINST
MoonEP -> DeepEP V2
MoonEP's communication time stays flat while DeepEP v2 degrades with routing skew
Source finding
RELEASED
DeepSeek -> DeepEP V2
DeepSeek open-sourced DeepEP V2 on April 23
Source finding
RELEASED
DeepSeek -> DeepEP V2
DeepSeek released DeepEP V2 expert-parallel communication library
Source finding

Source trail

GitHub GitHub GitHub

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…