1.0 KiB
1.0 KiB
Qwen30 Fixed-PD TTFT admission diagnosis
Verdict: Frontier overpredicts decode service time, so TP4's higher global arrival rate crosses the MNS admission cap only in simulation. The resulting first-scheduling queue reverses the TP topology ranking; without that queue Frontier itself predicts TP4 faster than TP2, matching real hardware.
| Config | Sim TPOT / real | Sim required slots / MNS | Real slots upper bound / MNS | Sim first-schedule wait | Real waiting max |
|---|---|---|---|---|---|
| tp2_mns64 | 8.09x | 58.0/64 | 7.5/64 | 38.2 ms | 0 |
| tp4_mns16 | 3.83x | 54.5/16 | 14.6/16 | 64906.3 ms | 0 |
| tp4_mns32 | 4.67x | 65.9/32 | 14.5/32 | 27180.7 ms | 0 |
| tp4_mns64 | 5.61x | 79.8/64 | 14.7/64 | 6367.3 ms | 0 |
TP4/MNS32 minus TP2/MNS64 TTFT
- Observed Frontier: +27087.0 ms.
- Frontier without first-scheduling queue: -55.4 ms.
- Real hardware: -71.4 ms.
Boundary: This identifies where the selection error is created, but the frozen real logs do not attribute service-time overprediction to one operator.