TFL — Token Flow Laboratory

An interactive, deterministic simulation of how a language-model serving system carries one request from prompt to next token. Everything runs in your browser as a teaching model.

Scope: timings, capacities and token text are synthetic. No model, neural network or inference engine runs here, no hardware benchmark is executed, and tokenization is illustrative. Time-to-first-token is calculated from simulated events, not measured latency.

What it covers

Queueing, prefill, KV-cache reservations, decode steps and token delivery.
Nine experiments: one request, load storm, steady arrivals, 32K context, code generation, many short users, KV wall, overload and batching trade-off.
Ten guided checkpoints in Token Flow 101.
Assumptions and evidence shown beside every result, so a simulated number is never read as a measurement.