2.0 KiB
2.0 KiB
Qwen235 Fixed-PD state diagnosis
Verdict: State-composition mismatch is insufficient: after reweighting Frontier to measured real decode composition, it still predicts TP8 faster.
| 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% |
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 internal component contrast at real Running proxy
Positive means TP8 slower; negative means Frontier gives TP8 an advantage.
| Component | TP8 - TP4 (ms/step) |
|---|---|
| moe_compute | -16.0581 |
| total | -8.5517 |
| ep_communication | +6.8248 |
| tp_dp_communication | -5.4010 |
| moe_routing | +0.3109 |
| attention | -0.2407 |
| dense_mlp_compute | +0.0018 |
| pipeline_communication | +0.0000 |
| runtime_overhead | +0.0000 |
Interpretation boundary
- Real observed TPOT contrast: +6.9500 ms/token.
- 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.