KV blocks · preserved progress · bounded recovery

LLM preemption policy lab

This is a local policy sandbox, not an LLM runtime or GPU benchmark. Every tick, token rate, and byte rate is declared synthetic arithmetic. Current vLLM V1 documents recomputation; swap here is a historical systems comparison.

Summary

Ready.

Per-request progress, delay, completion, and fairness state
ID	Prompt	Generated start → end	Preemptions	Queue ticks	Completion	TTFT-like	TPOT-like	Starvation

Event timeline

Canonical receipt

No receipt yet.