MegEngine/MegEngine

Health v2 · maintenance only

Adoption not part of the score

4808 stars · 546 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

MegEngine is a fast, scalable deep learning framework with automatic differentiation, developed in C++ with Python bindings. It unifies training and inference in a single model, supports quantization and dynamic shapes, and deploys efficiently across x86, Arm, CUDA, and ROCm platforms.

Use cases

• train neural networks with autograd in Python
• deploy deep learning models on mobile and embedded devices
• reduce GPU memory usage during training with DTR
• quantize models for faster inference
• run inference on CPU, CUDA, or ROCm hardware
• build models with a NumPy-like tensor API

When to choose

• you need a single framework for both training and inference
• you target diverse platforms including mobile, TEE, and embedded devices
• GPU memory is a bottleneck and you need DTR-style optimization
• you want quantization-aware training with a unified model

When to avoid

• you need the broadest ecosystem of pretrained models and community tools (PyTorch/TensorFlow)
• you require the latest Python versions (MegEngine supports Python 3.6-3.9)
• you need frequent releases and very active development

Facets

library · maturity maintenance

deep-learning machine-learning llm-training deep-learning machine-learning gpu-computing windows python cpp cross-platform autograd tensor training-and-inference quantization dtr-memory-optimization numpy-like-api linux macos android ios gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="MegEngine/MegEngine")