# skyzh/tiny-llm

learn LLM inference system on Apple Silicon for systems engineers: build a tiny vLLM + Qwen

Repository: https://github.com/skyzh/tiny-llm
Canonical: https://ross.abutalabs.com/products/tiny-llm
Homepage: https://skyzh.github.io/tiny-llm/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: course, large-language-model, llm, python, qwen, serving, vllm
Last push: 2026-08-26T09:55:57+00:00

## Health v2 (maintenance only)
Score: 64/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 35, longevity 35
- inputs: {"age_days": 501, "days_push": 7, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 4523, forks 369 (observed 2026-08-28T04:08:52.812406+00:00)

## What it is
A hands-on course for systems engineers to build an LLM inference and serving system from scratch on Apple Silicon using MLX, implementing a Qwen3 model, KV cache, and a mini vLLM-style server. It also covers building a coding agent, with exercises implemented in Python, C++, or Metal.

## Use cases
- learn how LLM inference works end to end
- build a tiny vLLM-style serving system
- implement attention, RoPE, and KV cache from scratch
- write Metal kernels for LLM decode on Apple Silicon
- understand continuous batching and paged KV cache
- study LLM serving without reading a huge CUDA codebase

## When to choose
- you are a systems engineer with PyTorch basics wanting to understand LLM serving internals
- you have an Apple Silicon Mac and no CUDA GPU
- you want a guided, test-verified path like CMU's Needle project

## When to avoid
- you need a production-ready inference server
- you want to learn on NVIDIA CUDA hardware
- you are looking for a high-level framework rather than a course

## Facets
- artifact type: learning-resource
- maturity: active
- function: llm-inference, machine-learning, gpu-computing, developer-tools
- domain: large-language-models, tutorials, developer-tools, machine-learning
- platform: python
- tags: llm-serving, mlx, apple-silicon, qwen, vllm, kv-cache, continuous-batching, metal-kernels, course, hands-on, macos, gpu

## Member repositories
- skyzh/tiny-llm (main) score 64

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:52.812406+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:20:16.493852+00:00, confidence not recorded.
  - readme: https://github.com/skyzh/tiny-llm (fetched 2026-08-28T04:08:52.812406+00:00, sha 28250e897483)
  - homepage: https://skyzh.github.io/tiny-llm/ (fetched 2026-08-29T09:06:06.302459+00:00, sha fa863229e78f)
- Data as of 2026-08-30T08:39:29.467469+00:00.