TPP-SD is a fast and distributionally consistent sampling framework for Transformer-based temporal point processes, achieving 2–6× speedup over autoregressive methods by adapting speculative decoding with a lightweight draft model.