# Qwen235 Fixed-PD state diagnosis **Verdict:** Exact real state composition flips Frontier to the correct TP4 winner; the ranking error is caused by closed-loop state/composition mismatch, not a conditional stage reversal. | Config | Real Running proxy mean | Sim decode batch mean | Proxy coverage | |---|---:|---:|---:| | tp4_ep1_mns64 | 4.667 | 13.537 | 100.0% | | tp8_ep8_mns64 | 12.238 | 18.938 | 100.0% | | Config | Real exact decode batch mean | Real token-weighted iteration ms | Matched simulator ms | Exact coverage | |---|---:|---:|---:|---:| | tp4_ep1_mns64 | 3.890 | 18.4710 | 34.2575 | 100.0% | | tp8_ep8_mns64 | 10.921 | 27.5887 | 45.1603 | 100.0% | ## Frontier internal component contrast over its own executed composition Decode-token-weighted over both pure-decode and mixed prefill/decode steps. | Component | TP8 - TP4 (ms/decoded token step) | |---|---:| | moe_compute | -32.4544 | | total | -20.0663 | | attention | -7.4619 | | tp_dp_communication | -6.5711 | | ep_communication | +6.2354 | | moe_routing | -0.2862 | | dense_mlp_compute | -0.1231 | | pipeline_communication | +0.0000 | | runtime_overhead | +0.0000 | ## Frontier component contrast at exact real token composition Positive means TP8 slower; negative means Frontier gives TP8 an advantage. | Component | TP8 - TP4 (ms/step) | |---|---:| | total | +10.9028 | | ep_communication | +6.5355 | | tp_dp_communication | -6.4718 | | attention | -2.1304 | | moe_compute | -1.9774 | | moe_routing | +0.1384 | | dense_mlp_compute | -0.0308 | | pipeline_communication | +0.0000 | | runtime_overhead | +0.0000 | ## Interpretation boundary - Real observed TPOT contrast: +6.9500 ms/token. - Real exact-iteration contrast: +9.1177 ms/decode-token-weighted step. - vLLM 10-second Running is active requests, not per-iteration decode batch; context lengths and graph buckets are unavailable in frozen logs. - Frontier state outputs have no direct graph bucket/padding/launch-overhead field; graph effects remain folded into predictors. - For TP8 shared-domain sync, total is ledger stage_end-start (critical path); op categories are serialized work estimates and are not additive because overlap and lane-summed wait diagnostics are represented separately. - Component deltas identify where Frontier creates its own TP8 advantage; without real per-stage measurements they are not yet root-cause proof.