ForceInjection/AI-fundamentals

AI 基础知识 - GPU 架构、CUDA 编程、大模型基础及AI Agent 相关知识。

What it solves

This repository provides a comprehensive, structured learning path for AI infrastructure. It bridges the gap between low-level hardware architecture and high-level AI application development, helping engineers and researchers understand the entire technical stack required to build, deploy, and optimize modern AI systems.

How it works

The content is organized into eight modular sections, ranging from hardware fundamentals (GPU/TPU architecture, interconnects like NVLink and CXL) to cloud-native AI platforms (Kubernetes, GPU virtualization, distributed storage). It includes deep dives into low-level programming (CUDA, DPU, NPU), LLM theory, and practical training pipelines, supported by documentation, code examples, and curated learning resources.

Who it’s for

This resource is designed for AI engineers, system architects, GPU programmers, LLM application developers, and technical researchers who want to master the underlying systems that power modern artificial intelligence.

Highlights

  • Hardware Deep Dives: Detailed analysis of GPU/TPU architectures, PCIe, NVLink, and CXL interconnects.
  • Cloud-Native AI: Guides on Kubernetes-based AI infrastructure, including GPU scheduling (Kueue/HAMi) and distributed storage (JuiceFS, 3FS).
  • Low-Level Programming: Extensive resources for CUDA, DPU, and Huawei NPU development.
  • LLM Engineering: Practical guides on LLM theory, training, fine-tuning, and agentic workflows using LangGraph and Spring AI.
  • Hands-on Demos: A centralized index of runnable code demonstrations for practical validation.

Related

  • Project
    ddz16/TSFpaper

    A comprehensive, curated reading list of over 400 research papers on time series and spatio-temporal forecasting, organized by model type and updated with the latest advancements in deep learning.

  • Project
    tslearn-team/tslearn

    A machine learning toolkit for time series analysis in Python that provides specialized tools for classification, clustering, and regression of sequential data.

  • Project
    radixark/miles

    Miles is a PyTorch‑native, enterprise‑grade reinforcement‑learning framework for post‑training trillion‑parameter models. It combines SGLang for ultra‑fast rollout with Megatron‑LM (or FSDP2/torchtitan) for scalable training, offering async RL, low‑precision support, LoRA adapters, fault‑tolerant rollout routing, and day‑0 compatibility with frontier LLMs.

  • Project
    Jittor/jittor

    Jittor is a high‑performance deep‑learning framework that JIT‑compiles the entire computation graph and uses meta‑operators to generate specialised CUDA/C++ kernels. It offers a PyTorch‑like Python API, supports CPU and GPU back‑ends (CUDA, ROCm, Hygon), and includes a model zoo for vision, rendering, geometric learning, and reinforcement learning.