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.
The whole path with coordinates — every hop, and the file that owns it.
Both halves of one class — schedule() admits work, update_from_output closes the loop.
Four layers deep — cache-hit lookup, allocation, and the free-block queue underneath.
The persistent batch, the forward pass, and why sampling is a second call.
On-GPU sampling, the ordered logits-processor chain, and constrained decoding.
The return leg — across the boundary, incremental detokenization, and SSE.
Which backend gets picked at startup, what it demands every step, and how a new one plugs in.
How an architecture string resolves to a class, how checkpoint tensors find their parameters, and where a tensor-parallel shard gets decided.
CompilationConfig, why the graph gets split, the capture ladder, and what startup pays for it.
TP/PP/DP as three axes, who spawns what, which ranks can talk to which, and what one dead rank takes down.
One interface, two callers — the factory that builds it, the lookup that consults it, and what happens on a miss.
Three real, merged PRs — predict every file they touched from the title and description alone, then check yourself against the diff.