Files
aituner/runs/simulator-tuning-latency-matrix-v0/qwen30-tracepd-u0p01-real-audit.md

1.4 KiB

Qwen3-30B-A3B TP-normalized Trace-PD: real vLLM audit

All 36 fresh-server trials passed the trace contract: 129/129 exact-usage requests per trial. The smoke run is excluded. Values below pool the three trials (387 requests/config); p90 uses nearest-rank order statistics.

Config TTFT mean/p90 (ms) TPOT mean/p90 (ms) E2E mean/p90 (ms)
tp1_mns8 88479.9/168185.5 14.3/16.3 137996.4/222136.0
tp1_mns16 17378.2/42739.8 20.6/25.3 87355.2/156177.2
tp1_mns32 620.5/1931.5 23.8/30.0 81051.2/162908.6
tp1_mns64 619.8/1956.7 24.1/30.4 82253.7/165355.3
tp2_mns8 95920.1/183878.8 9.2/10.2 128296.9/215848.3
tp2_mns16 34443.0/77922.4 14.1/16.1 82936.7/137726.4
tp2_mns32 1050.6/2521.8 18.0/21.3 62267.5/117239.3
tp2_mns64 375.8/1148.6 18.2/21.6 62353.6/118392.3
tp4_mns8 102605.9/193359.7 6.8/7.2 126405.4/209092.4
tp4_mns16 35042.6/74281.3 8.8/9.6 65779.1/110374.3
tp4_mns32 6296.2/19159.9 12.2/13.8 47944.5/83825.7
tp4_mns64 246.0/685.5 13.2/15.4 44985.1/83763.6

Per-metric winners

  • ttft_ms:pooled_mean_ms: tp4_mns64 (246.0 ms)
  • ttft_ms:pooled_p90_ms: tp4_mns64 (685.5 ms)
  • tpot_ms:pooled_mean_ms: tp4_mns8 (6.8 ms)
  • tpot_ms:pooled_p90_ms: tp4_mns8 (7.2 ms)
  • e2e_ms:pooled_mean_ms: tp4_mns64 (44985.1 ms)
  • e2e_ms:pooled_p90_ms: tp4_mns64 (83763.6 ms)