Inside vLLM

12 modules, interactive — the whole vLLM request path plus the startup internals underneath it: attention backends, model loading, compilation, distributed execution, KV connectors, and a capstone against real merged PRs. Free.

RQ
Module 1
Request In, Tokens Out

The whole path with coordinates — every hop, and the file that owns it.

Start module →
SC
Module 2
The Scheduler

Both halves of one class — schedule() admits work, update_from_output closes the loop.

Start module →
KV
Module 3
KV Cache Manager

Four layers deep — cache-hit lookup, allocation, and the free-block queue underneath.

Start module →
MR
Module 4
Model Runner

The persistent batch, the forward pass, and why sampling is a second call.

Start module →
SM
Module 5
Sampler

On-GPU sampling, the ordered logits-processor chain, and constrained decoding.

Start module →
OS
Module 6
Output & Streaming

The return leg — across the boundary, incremental detokenization, and SSE.

Start module →
AB
Module 7
Attention Backends

Which backend gets picked at startup, what it demands every step, and how a new one plugs in.

Start module →
RG
Module 8
Model Registry

How an architecture string resolves to a class, how checkpoint tensors find their parameters, and where a tensor-parallel shard gets decided.

Start module →
CC
Module 9
Compilation & CUDA Graphs

CompilationConfig, why the graph gets split, the capture ladder, and what startup pays for it.

Start module →
DX
Module 10
Distributed Execution

TP/PP/DP as three axes, who spawns what, which ranks can talk to which, and what one dead rank takes down.

Start module →
KC
Module 11
KV Connectors & Disaggregation

One interface, two callers — the factory that builds it, the lookup that consults it, and what happens on a miss.

Start module →
CP
Module 12
Capstone — Predict the PR

Three real, merged PRs — predict every file they touched from the title and description alone, then check yourself against the diff.

Start module →