Paper notes · ICML 2026
dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching
Caches more than keys and values: per layer it stores K, V, attention output and FFN output, separately for the prompt and the response. The prompt cache is refreshed every steps (long), the response cache fully every steps (short), and in between only the response tokens whose value vectors changed most are recomputed.
The PDF's license does not allow reposting it here; page and line numbers refer to the version linked above.