Qwen3.5-27B · development
MTP k=3, probabilistic draft with top-k/top-p (fp32 fix)
Passed every gate.
SpecScore1.272geometric mean of the three speedups
solo1.72×
one lane; passes 1.81, 1.72, 1.57
team1.22×
team lanes busy; passes 0.93, 1.22, 1.42
busy0.98×
KV cache nearly full; passes 0.98, 1.08, 0.84
Lossless gate
what each check is →| Check | Measured | Limit | |
|---|---|---|---|
| device memory | 161.5 GiB | ≤ 165.2 GiB | pass |
| fixed texts' log-probabilities | 0.012 | ≤ 0.02 | pass |
| sampled replies' likelihood | -0.0223 ± 0.0566 | within ±0.05 at 95% | pass |
| share of first choices | – | within ±None at 95% | n/a |
| greedy agreement | 0.460 | none on this task | n/a |
| reply lengths | – | 1 ± 0.05 at 95% | pass |
Speculation
| Workload | Tokens / verify step | Speedup |
|---|---|---|
| solo | 3.09 | 1.72 |
| team | 3.12 | 1.22 |
| busy | 3.17 | 0.98 |
Speculator: none
Engine: a directory of its workspace
Everything the judge recorded
{
"ratios": {
"solo": {
"passes": [
1.8122482268922282,
1.7202078793514275,
1.5721111262143177
],
"ratio": 1.7202078793514275
},
"team": {
"passes": [
0.9313444932236264,
1.218325868821551,
1.4155165515640267
],
"ratio": 1.218325868821551
},
"busy": {
"passes": [
0.9830874321782946,
1.0805494754965062,
0.8368754430018319
],
"ratio": 0.9830874321782946
}
},
"specscore": 1.2724640288062634,
"greedy_agreement": 0.4601225407286068,
"likelihood": {
"gap": -0.022268614296407918,
"se": 0.02887142023606424,
"candidate": -0.4984058281978848,
"stock": -0.4761372139014769
},
"fidelity": 0.011955906862795603,
"lengths": {
"solo": {
"ratio": 0.7880888471867672,
"low": 0.5143668445167442,
"high": 1.0618108498567902
},
"team": {
"ratio": 0.9929615154156227,
"low": 0.7285915976402043,
"high": 1.257331433191041
},
"busy": {
"ratio": 0.919122380141161,
"low": 0.6863926636030365,
"high": 1.1518520966792856
}
},
"memory": {
"ok": true,
"peak_mib": 165416.0,
"allowed_mib": 169119.1
},
"acceptance": {
"solo": 3.085219256165698,
"team": 3.120080230093854,
"busy": 3.1715901098306554
},
"limits": {
"greedy_agreement_floor": null,
"likelihood_bound": 0.05,
"fidelity_ceiling": 0.02,
"length_margin": 0.05,
"memory_slack_mib": 4096
},
"reason": null,
"score": 1.2724640288062634
}