Decagon shares speculative decoding lessons on DFlash and EAGLE

Decagon engineer Nicholas Liu shares lessons from evaluating DFlash and EAGLE, two speculative decoding methods, according to a post from the LMSYS Org account on X. The post says concurrency, the cost of verifying tokens on mixture-of-experts models, and structured outputs shape latency, cost and deployment decisions for these methods. The post does not report measured speedups or specific results for either method.