# deepseek-ai/DeepEP

DeepEP: an efficient expert-parallel communication library

Repository: https://github.com/deepseek-ai/DeepEP
Canonical: https://ross.abutalabs.com/products/deepep
Language: Cuda
License: MIT
License Family: permissive
Last push: 2026-08-20T08:04:39+00:00

## Health v2 (maintenance only)
Score: 57/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 98, release rhythm 15, longevity 40
- inputs: {"age_days": 563, "days_push": 13, "days_rel": 352, "gap_med": null, "n_releases_24m": 1}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 10066, forks 1413 (observed 2026-08-28T04:10:39.453986+00:00)

## What it is
DeepEP is a high-performance GPU communication library for expert parallelism (EP) in MoE training and inference, providing high-throughput and low-latency all-to-all dispatch/combine kernels with FP8 support. It uses runtime JIT compilation with an NCCL Gin backend and experimental primitives for pipeline, context, and remote-memory parallelism with zero or minimal SM occupation.

## Use cases
- dispatch tokens to experts across GPUs for MoE training
- combine expert outputs with low-latency all-to-all kernels
- run FP8 dispatching and BF16 combining for large-scale MoE models
- train mixture-of-experts models at EP2048 scale without auto-tuning
- overlap expert-parallel communication with minimal SM usage
- add pipeline or context parallelism primitives to a distributed training stack

## When to choose
- you are training or serving MoE models and need efficient expert-parallel all-to-all communication
- you need near-hardware-bandwidth dispatch/combine on NVLink and RDMA (InfiniBand/RoCE) clusters
- you want to minimize SM occupation so GPUs stay free for compute
- you prefer runtime JIT kernels with no CUDA compilation step at install time
- you are building on Hopper (SM90) or newer (SM100) NVIDIA hardware with NCCL available

## When to avoid
- your workload is single-GPU or uses only data/tensor parallelism with no expert parallelism
- you need the legacy 0-SM RDMA low-latency EP mode, which is no longer supported in V2
- you cannot accept the larger buffer memory footprint of V2
- you target non-NVIDIA GPUs or platforms without NCCL/RDMA support
- you need production-stable pipeline, context, or Engram remote-memory features, which are still experimental

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-training, llm-inference, gpu-computing, concurrency
- domain: machine-learning, deep-learning, large-language-models, gpu-computing, microservices, developer-tools
- platform: python, cpp
- tags: expert-parallelism, mixture-of-experts, moe, all-to-all-communication, rdma, nvlink, nccl, nvshmem, jit-compilation, fp8, low-precision, pipeline-parallelism, context-parallelism, cuda-kernels, high-performance-computing, sm-occupation, elastic-buffer, deepseek, linux, gpu

## Member repositories
- deepseek-ai/DeepEP (main) score 57

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:39.453986+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:19:53.227562+00:00, confidence not recorded.
  - readme: https://github.com/deepseek-ai/DeepEP (fetched 2026-08-28T04:10:39.453986+00:00, sha 85c900ff9178)
- Data as of 2026-08-30T08:39:29.467469+00:00.