Concepts

Five pages on what the kernel computes and why it is built the way it is. Read them in order the first time; each assumes the ones before it.

The kernel in BLIS

The five BLIS repositories, what this one takes from each, and why pricing lives apart from simulation and from data.

Pricing a step

How one forward pass is turned into time: the resources a step uses, the laws that price each, and how they compose into the two estimates a step returns.

Memory

What a GPU holds before any request arrives, and what each request adds as its context grows.

Parallelism

Tensor, data, expert, prefill-context and decode-context parallelism: what each divides, what each costs, and why it matters which GPUs end up in each group.

Refuse, assume, override

What the kernel does with a deployment it cannot price exactly, and how it says so.

The pages explain the model rather than list it. Where a statement rests on a detail of the engine, the page names the source file, and the comment there cites the vLLM lines.