Six findings and the mechanism behind each Six rows from the independent field reports. Each row names a workload, the result that was reported, and the SGLang mechanism that explains it: kernels, the radix cache, chunked prefill, the decode graph limit, attention data and expert parallelism, and the cost of draft tokens. WORKLOAD REPORTED RESULT MECHANISM unique prompts, dense tie within a few percent same kernel families repeated prefixes first token 583 vs 2,141 ms radix cache 16k-token prompts slower first token, faster TPOT chunked prefill 8,192 no concurrency cap ITL 100.92 vs 28.84 ms graph limit, a guess DeepSeek MoE, many GPUs reference engine, audited MLPerf attention DP and EP MTP on a B200 TPOT 18.43 to 9.92 ms draft cost, acceptance A dashed border marks our hypothesis, not a finding of the report.