DeepSpeed

Active
GitHub Python Apache-2.0

Description

Microsoft's open-source deep learning distributed training optimization library, featuring ZeRO memory optimization, 3D parallelism, and mixed-precision training for efficient training of trillion-parameter models.

Key Features

  • ZeRO optimizer — Partitions optimizer states, gradients, and parameters across GPUs for massive memory savings
  • 3D parallelism — Combines tensor, pipeline, and data parallelism simultaneously
  • Ultra-long sequence training — Ulysses Sequence Parallelism for training on extremely long context sequences
  • Mixture of Experts — Built-in DeepSpeed-MoE supports trillion-parameter model training
  • Inference acceleration — DeepSpeed Inference optimizes Transformer model inference performance
  • Multi-hardware support — Compatible with NVIDIA, AMD, Intel GPUs and CPU training

Use Cases

💡 Large-scale LLM pretraining: Train trillion-parameter models on thousands of GPUs with ZeRO-3/ZeRO-Infinity
💡 Instruction fine-tuning and RLHF: Efficient RLHF training with low resource requirements via DeepSpeed-Chat
💡 Scientific computing acceleration: DeepSpeed4Science optimizes molecular dynamics, weather forecasting, and other simulations
💡 Multimodal large model training: Train vision-language models with DeepSpeed-VisualChat

Strengths & Limitations

Strengths

  • Actively maintained, recent updates
  • High community interest (43.1k stars)
  • Permissive open-source license (Apache-2.0)

Quick Start

# Install DeepSpeed
pip install deepspeed

# Verify installation
ds_report

# Run training script with DeepSpeed
deepspeed --num_gpus=4 train_script.py

# Or use via PyTorch Lightning integration
pip install pytorch-lightning

Related Projects

Megatron-LM

17.7k · Python
Active A

NVIDIA's open-source GPU-optimized library for training transformer models at scale, providing tensor parallelism, pipeline parallelism, sequence parallelism, and mixed-precision (FP8/FP4) support — the core foundation for trillion-parameter LLM training.

transformerdistributed-traininggpu +3
  • · Tensor parallelism (TP) — Shard individual transformer layers across GPUs to reduce per-card memory
  • · Pipeline parallelism (PP) — Distribute model layers across GPUs for ultra-large models
  • · Context parallelism (CP) — Handle ultra-long sequences efficiently with million-token training

Agently

1.6k · Python
Active A

A GenAI application development framework that simplifies agent interaction with structured data and chained-calls syntax, using event-driven flow for complex logic.

agentpythonframework +1
  • · Structured output control with framework-guaranteed schemas, required field extraction, and retry validation
  • · Runtime Skills system for discovering, installing, and executing MCP/script capabilities on demand
  • · TriggerFlow event-driven workflows with fan-out, pause/resume, save/load, and sub-flow support

Atomic Agents

6.2k · Python
Active A+

Atomic Agents is a modular AI agent building framework with an atomic design philosophy, providing composable components including tools, pipelines, and memory management for constructing agent systems.

frameworkpythonagent +2
  • · Atomic design philosophy with single-purpose, reusable, composable components
  • · Built on Instructor and Pydantic for type-safe, predictable agent behavior
  • · Multi-provider support including OpenAI, Anthropic, Groq, Gemini, and more

AutoChain

1.9k · Python
Stale C

AutoChain is a lightweight, extensible, and testable LLM Agent framework by Forethought, providing clean abstractions for agent building with automatic tool selection, conversation history management, and automated testing workflows.

frameworkagentpython +3
  • · Lightweight and extensible generative agent pipeline with minimal abstraction layers
  • · Support for custom tools and OpenAI function calling with automatic spec conversion
  • · Simple memory tracking for conversation history and tool outputs with BufferMemory