# DeepSpeed

DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.

Repository: https://github.com/deepspeedai/DeepSpeed
Canonical: https://ross.abutalabs.com/products/deepspeed
Homepage: https://www.deepspeed.ai/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: deep-learning, pytorch, gpu, machine-learning, billion-parameters, data-parallelism, model-parallelism, inference, pipeline-parallelism, compression, mixture-of-experts, trillion-parameters, zero
Last push: 2026-08-26T22:18:46+00:00
Link (homepage): https://www.deepspeed.ai/
Link (site_page): https://www.deepspeed.ai/getting-started
Link (site_page): https://www.deepspeed.ai/docs/config-json

## Health v2 (maintenance only)
Score: 98/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 97, longevity 100
- inputs: {"age_days": 2414, "days_push": 7, "days_rel": 23, "gap_med": 18.5, "n_releases_24m": 37}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 43003, forks 4940 (observed 2026-08-28T04:12:08.903448+00:00)

## What it is
DeepSpeed is a deep learning optimization library from Microsoft that makes distributed training and inference of large models fast, memory-efficient, and easy via innovations like ZeRO, 3D-Parallelism, and MoE support. It integrates with PyTorch, HuggingFace Transformers, and PyTorch Lightning, and scales models from millions to trillions of parameters.

## Use cases
- train large language models across multiple GPUs
- fit billion-parameter models on limited GPU memory with ZeRO offloading
- speed up distributed PyTorch training with data and model parallelism
- run efficient inference for large transformer models
- train mixture-of-experts models at scale
- fine-tune LLMs with HuggingFace Transformers and DeepSpeed
- train models with very long sequences efficiently

## When to choose
- you need to train or fine-tune very large models that exceed single-GPU memory
- you want ZeRO-style optimizer state sharding and CPU/NVMe offloading
- you use PyTorch and want minimal code changes for distributed training
- you need proven large-scale training (BLOOM, MT-530B class models)

## When to avoid
- you train small models on a single GPU where plain PyTorch suffices
- you need a framework-agnostic solution outside the PyTorch ecosystem
- you want a simple high-level trainer without configuration complexity

## Facets
- artifact type: library
- maturity: stable
- function: deep-learning, llm-training, llm-inference, machine-learning, gpu-computing
- domain: deep-learning, large-language-models, machine-learning, gpu-computing
- platform: python, cloud
- tags: zero, distributed-training, model-parallelism, data-parallelism, pipeline-parallelism, mixture-of-experts, pytorch, memory-optimization, offloading, compression, gpu, linux, docker

## Member repositories
- deepspeedai/DeepSpeed (main) score 98
- deepspeedai/DeepSpeedExamples (examples) score 77

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:12:08.903448+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 2026-08-29T16:22:36.193477+00:00, confidence not recorded.
  - readme: https://github.com/deepspeedai/DeepSpeed (fetched 2026-08-28T04:12:08.903448+00:00, sha 22d407cb0dbe)
  - homepage: https://www.deepspeed.ai/ (fetched 2026-08-29T07:46:13.423109+00:00, sha 95efd6a17e95)
  - site_page: https://www.deepspeed.ai/getting-started (fetched 2026-08-29T07:46:13.432681+00:00, sha 844cd426fc60)
  - site_page: https://www.deepspeed.ai/docs/config-json (fetched 2026-08-29T07:46:13.435365+00:00, sha 2ed3774fd5ae)
  - registry_pypi: https://pypi.org/pypi/deepspeed/json (fetched 2026-08-29T07:46:13.441254+00:00, sha 490116f8450c)
- Data as of 2026-08-30T08:39:29.467469+00:00.