/ai-infra

AI Infra best practices

Creative Commons Zero v1.0 UniversalCC0-1.0

AI Infra best practices

This repo aims to provide a set of best practices for AI infrastructures based on consumer-grade hardware. The target audience is AI engineers who want to build their own AI infrastructures.

In this repo, we use Ubuntu 24.04 as the base system. The hardware is Amazon AWS G5 instances powered by NVIDIA A10G GPUs.

Specifically, we choose AWS G5 x.2large instances. The NVIDIA A10G is close to RTX3060, and it's the proper choice for consumer-grade AI infrastructures in Amazon AWS.

A10G vs RTX3060

Base system setup

Training

Demos

AI/ML Library

AI/ML Library Complexity Description
PyTorch High General-purpose AI/ML framework for building and deploying complex models across various domains.
tinygrad Low Light-weighted AI/ML framework
llm.c Low pure C/CUDA implementation