# deepseek-ai/DualPipe

A bidirectional pipeline parallelism algorithm for computation-communication overlap in DeepSeek V3/R1 training.

Repository: https://github.com/deepseek-ai/DualPipe
Canonical: https://ross.abutalabs.com/products/dualpipe
Language: Python
License: MIT
License Family: permissive
Last push: 2026-01-14T06:34:59+00:00

## Health v2 (maintenance only)
Score: 48/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 62, release rhythm 35, longevity 39
- inputs: {"age_days": 553, "days_push": 231, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2998, forks 333 (observed 2026-08-28T04:07:37.259797+00:00)

## What it is
DualPipe is a Python library implementing a bidirectional pipeline parallelism algorithm that overlaps forward and backward computation with communication to reduce pipeline bubbles in large model training. It also includes DualPipeV, a V-shaped schedule variant, and was introduced in the DeepSeek-V3 technical report.

## Use cases
- reduce pipeline bubbles when training large language models
- overlap computation and communication in distributed PyTorch training
- implement bidirectional pipeline parallel schedules for MoE models
- compare pipeline parallelism strategies like 1F1B, ZB1P, and DualPipe
- train DeepSeek V3/R1 style models efficiently across many GPUs

## When to choose
- you are training very large models with pipeline parallelism and communication is a bottleneck
- you use PyTorch 2.0+ and can implement a custom overlapped_forward_backward for your module
- you want to reproduce DeepSeek-style training efficiency

## When to avoid
- you only train small models where pipeline bubbles are negligible
- you need a turnkey training framework rather than a scheduling algorithm
- your stack does not support PyTorch 2.0 or multi-GPU pipeline parallelism

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, gpu-computing
- domain: large-language-models, deep-learning, microservices, gpu-computing
- platform: python
- tags: pipeline-parallelism, distributed-training, pytorch, computation-communication-overlap, deepseek, gpu, linux

## Member repositories
- deepseek-ai/DualPipe (main) score 48

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:37.259797+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 2026-08-29T18:47:25.820218+00:00, confidence not recorded.
  - readme: https://github.com/deepseek-ai/DualPipe (fetched 2026-08-28T04:07:37.259797+00:00, sha e9bf8b690a07)
- Data as of 2026-08-30T08:39:29.467469+00:00.