10-08-2026, 02:16 PM
Interesting, I hadn't tried that. The pattern around speculative decoding has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
measure twice, benchmark three times