Skip to content

DeepEP

A communication library for MoE and expert parallel

Visit official site

What DeepEP is

DeepEP is described as a communication library for mixture of experts models, with high throughput and low latency GPU kernels.

What you can do with it

  • Use high throughput kernels for dispatch and combine
  • Forward data from an NVLink domain to an RDMA domain
  • Use pure RDMA low latency kernels for decoding

Requirements

  • Hopper GPUs and RDMA network devices are named
  • Python 3.8 and CUDA 12.3 or newer are listed
  • NVSHMEM is installed before the library setup

Pros & cons

✓ What we like

  • Pure RDMA low latency kernels are listed
  • SM number control is supported

! What to watch out for

  • Requirements name Hopper GPUs and RDMA networking
  • The implementation differs slightly from the DeepSeek-V3 paper

FAQ

What does it target?

The material describes mixture of experts models and expert parallel communication.

What are the requirements?

The steps name Hopper GPUs, RDMA devices, Python 3.8, CUDA 12.3, PyTorch 2.1 and NVSHMEM.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review