← Browse tools

Training & fine-tuning

DeepSpeed

Makes model-state partitioning, offloading and distributed training optimizations available as reusable infrastructure.

43,172 stars

Overview

Research summary

DeepSpeed provides infrastructure for scaling model training and inference across accelerator resources. Its documented features include ZeRO optimizer-state partitioning, offloading, mixed precision and parallel execution strategies. Developers integrate it into a training program and configure the execution behavior to fit model size and available hardware. The repository also contains examples and research implementations for efficiency improvements. Its primary value is reducing the memory and systems overhead of large-model workloads rather than defining a particular model architecture.

Repository summary

Stars
43,172
Open issues
Unavailable
Last push
2026-10-02
Commits, 90 days
Unavailable
Repository activity
Not scored
Version
Unavailable

Recorded catalogue figures. View repository data and provenance →

Classification

Pricing & services

Paid services unknown

Whether the provider offers paid products or services has not been established.

Licence scope

Implementation

Recorded implementation details and interfaces for DeepSpeed.

Implementation details

Python library and supporting tools

Languages
Python
Repository type
source

Recorded interfaces and capabilities

Licence scope

Repository

Repository snapshots, release information and recorded maintenance signals.

Repository snapshot

deepspeedai/DeepSpeed ↗

Stars
43,172
Open issues
Unavailable
Last push
2026-10-02
Commits, 90 days
Unavailable
Repository activity
Not scored
Archived
Not recorded

Repository activity is a snapshot, not a quality or popularity ranking. It combines recent-push freshness (50%), 90-day commits (30%) and issue pressure (20%).

Maintenance and provenance

Catalogue snapshot
2026-10-06
Stars source
Recorded fallback
Last push source
Recorded fallback

Documentation

Recorded references and research provenance for this entry.

Recorded sources 5

Research metadata

Research date
2026-10-02
Catalogue snapshot
2026-10-06

Community & social 1 channels

Communities