Cross-Heterogeneous-Model KV-Cache Transfer

Extend KV-cache transfer and selective recomputation beyond model families with shared tokenizers and vocabulary spaces to support completely heterogeneous model families with differing tokenizers.

Background

RaReCache is evaluated within the Qwen3 and Llama model families, where the source and target models share the same tokenizer and vocabulary space. This shared tokenization allows corresponding prompt tokens to produce aligned cache entries that can be mapped between models.

The framework does not currently address transfer between model families with different tokenizers. Supporting such settings would require resolving how token sequences and their associated key-value cache entries correspond across incompatible tokenization schemes while preserving the effectiveness of selective target-side recomputation. The paper explicitly identifies this extension as an unresolved challenge.