Throughput on the trained Shakespeare model holds a flat ~650 tok/s with the cache while the naive path sags as context grows. A speedup is worthless if it changes the output. The benchmark uses the ...