All headline throughput measurements cover decoding forward passes only and exclude prefill; requests are cyclically replayed until the reported measurements stabilize. The reported speedups therefore do not by themselves quantify end-to-end request throughput including prompt prefill. · CiteArk
Not assessedNo independent reproduction scheduledLimitationc29-decode-only-throughput-boundary
All headline throughput measurements cover decoding forward passes only and exclude prefill; requests are cyclically replayed until the reported measurements stabilize. The reported speedups therefore do not by themselves quantify end-to-end request throughput including prompt prefill.
