Empirical Runtime Investigation of Sequence and KV Cache Management in the vLLM 0.7.2 Inference Engine

This paper presents an empirical runtime investigation of the vLLM 0.7.2 inference engine, focusing on sequence management and Key-Value (KV) cache allocation during inference. Using standard Python runtime inspection techniques on live objects without modifying the underlying implementation, the study examines the internal organization of the engine from the top-level LLM interface to the scheduler and memory management subsystems, including the SelfAttnBlockSpaceManager. Runtime state transitions are characterized before inference, during active decoding, and after synchronous generation to distinguish observable runtime behavior from post-execution state. The investigation further examines scheduler-level sequence branching interfaces and the corresponding capabilities of the block management layer while clearly separating empirical observations from architectural interpretations and implementation hypotheses. The resulting study provides a reproducible empirical baseline for future investigations of sequence management and KV cache behavior in high-throughput LLM serving systems.