LLMs-from-scratch: Build LLM PyTorch Từng Bước

AI Engineering From Scratch is a free, MIT-licensed curriculum with 511 lessons across 20 phases. It takes you from linear algebra through LLMs, agent engineering, and autonomous systems. Every lesson follows a Build It / Use It / Ship It loop: implement the algorithm from raw math, see how it works in PyTorch, then produce a reusable artifact. Four languages: Python, TypeScript, Rust, Julia. The curriculum runs on your own laptop, with EPUB and PDF editions for offline reading. A placement quiz routes you to your starting phase, and learning skills integrate directly into Claude Code and Codex for guided study. The README claims 150,639 readers and 241,669 page views in the last 30 days.

Hugging Face Transformers (huggingface/transformers) is a Python library that centralizes model definitions for text, computer vision, audio, video, and multimodal machine learning, covering both inference and training. The README describes it as a pivot point compatible with training frameworks such as Axolotl, DeepSpeed, and PyTorch-Lightning, and inference engines such as vLLM, SGLang, and TGI, with more than 1M+ model checkpoints listed on the Hugging Face Hub.

LlamaFactory tackles the fragmented reality of LLM fine-tuning by consolidating 100+ model architectures, a dozen training methods (full fine-tuning, LoRA variants, QLoRA at various bit-widths via AWQ/GPTQ/AQLM/HQQ/EETQ), and multimodal support under a single unified codebase. The zero-code CLI and browser-based LlamaBoard (Gradio UI) lower the barrier enough that fine-tuning a 7B model on a free Colab T4 GPU is a documented, reproducible task, while FSDP+QLoRA extends that to 70B across two consumer GPUs. Used in production by Amazon, NVIDIA, and Aliyun according to the README, it backs inference with OpenAI-compatible APIs powered by vLLM or SGLang, so a fine-tuned model drops into existing applications without pipeline rewrites. The catch is that breadth breeds complexity: the configuration surface is wide enough that beginners can misconfigure training and spend time debugging YAML rather than iterating on data, and some advanced features like NPU or Ascend support lack thorough documentation. If you need to fine-tune anything from a small Phi variant to a 671B DeepSeek MoE across text, vision, or audio, this is the least-bad option for avoiding three separate toolchains — just budget time for the learning curve.

Repomix is a TypeScript CLI (plus a Docker image, web app, and browser extensions) that concatenates a repository's files into one XML, Markdown, JSON, or plain-text file sized for an LLM's context window, with git-aware filtering, Secretlint-based secret scanning, and Tree-sitter compression available via the --compress flag.