tiny-llm - LLM Serving in a Week

A course on LLM serving using MLX for system engineers. The codebase is solely (almost!) based on MLX array/matrix APIs without any high-level neural network APIs, so that we can build the model serving infrastructure from scratch and dig into the optimizations.

The goal is to learn the techniques behind efficiently serving a large language model (e.g., Qwen3 models).

In week 1, you will implement the necessary components in Python (only Python!) to use the Qwen3 model to generate responses (e.g., attention, RoPE, etc). In week 2, you will implement the inference system which is similar to but a much simpler version of vLLM (e.g., KV cache, continuous batching, flash attention, etc). In week 3, we will cover more advanced topics and how the model interacts with the outside world.

Why MLX: nowadays it's easier to get a macOS-based local development environment than setting up an NVIDIA GPU.

Why Qwen3: it keeps the dense decoder architecture small enough for a local MLX course while adding modern details such as QK norm and bfloat16 weights. The official MLX 4-bit model files also make the setup predictable on Apple Silicon.

Book

The tiny-llm book is available at https://skyzh.github.io/tiny-llm/. You can follow the guide and start building.

Community

You may join skyzh's Discord server and study with the tiny-llm community.

Roadmap

Week 1 and 2 is complete. Week 3 is in progress.

Other topics not covered: quantized/compressed kv cache, prefix/prompt cache; sampling, fine tuning; smaller kernels (softmax, silu, etc)

Star History