vLLM, from the source
Most material about vLLM explains what it does. These tracks show you the code that does it — the real files, classes and control flow at a pinned release, each paired with a simulation of that exact code. The point is to make the repository navigable, so you can open it yourself and know where you are.
Concepts live in the LLM Serving track, and the whole curriculum is on Tracks. Neither is a prerequisite.
Read the vLLM source without drowning in it — the real files, classes, and control flow behind every stage of a request, paired with simulations of that exact code.
- The request path, with coordinates
- The scheduler, both halves
- KV cache manager & block pool
- Executor, workers & the model runner
- Sampler & structured output
- Output processing & streaming
- Attention backends, chosen once at startup
- Model registry & weight loading
- Compilation & CUDA graph capture
- Distributed execution — TP, PP, DP as processes
- KV connectors & P/D disaggregation
- Capstone — predict a real PR from its diff
How a semantic router picks the right model for each request — intent classification, reasoning-mode control, semantic caching, and the safety classifiers sitting in front of them.
- Mixture-of-Models — the routing thesis
- Intent classification
- Reasoning-mode control per request
- Semantic caching
- PII & jailbreak classifiers
- Envoy ExtProc as the data plane
- The Rust core, called from Go
Questions
How is this different from the LLM Serving track?
LLM Serving teaches the concepts — what PagedAttention is, why continuous batching wins, what TTFT means. The vLLM tracks teach the code that implements them: which file, which class, which call, in what order. Concepts first is the easier path, but neither depends on the other.
Why are these separate from the other tracks?
The curriculum tracks teach mechanisms that stay true across releases. These read one project's source at a pinned release, so they age with that codebase rather than with the field. Keeping them on their own shelf means the curriculum stays a single path, and these can be re-pinned on their own schedule.
Which version of vLLM do they follow?
Every excerpt is distilled from a pinned vLLM release and labelled with that version, the file it came from, and the real size of that file. A checked-in symbol manifest and a drift job keep the citations honest as vLLM moves.
Do I need a GPU?
No. Everything runs as a browser simulation — you are reading and driving the control flow, not executing it.