Qwen3-8B · development
N-gram CPU k=4 n=1..4
Passed every gate.
SpecScore0.835geometric mean of the three speedups
solo1.11×
one lane; passes 1.39, 1.02, 1.11
team0.77×
team lanes busy; passes 0.78, 0.77, 0.70
busy0.68×
KV cache nearly full; passes 0.68, 0.70, 0.68
Lossless gate
what each check is →| Check | Measured | Limit | |
|---|---|---|---|
| device memory | 161.2 GiB | ≤ 165.2 GiB | pass |
| fixed texts' log-probabilities | 1.1e-06 | ≤ 0.002 | pass |
| sampled replies' likelihood | +0.0419 ± 0.0884 | within ±0.05 at 95% | pass |
| share of first choices | – | within ±None at 95% | n/a |
| greedy agreement | 0.761 | none on this task | n/a |
| reply lengths | – | 1 ± 0.05 at 95% | pass |
Speculation
| Workload | Tokens / verify step | Speedup |
|---|---|---|
| solo | 1.96 | 1.11 |
| team | 1.91 | 0.77 |
| busy | 1.92 | 0.68 |
Speculator: none
Engine: a directory of its workspace
Everything the judge recorded
{
"ratios": {
"solo": {
"passes": [
1.3939944757145466,
1.0226287616373306,
1.1129390209850192
],
"ratio": 1.1129390209850192
},
"team": {
"passes": [
0.7839924565947793,
0.7704861474487533,
0.6999014730854541
],
"ratio": 0.7704861474487533
},
"busy": {
"passes": [
0.6794338937559374,
0.7040627136411703,
0.6759319433062221
],
"ratio": 0.6794338937559374
}
},
"specscore": 0.8352076640034652,
"greedy_agreement": 0.7614555860892295,
"likelihood": {
"gap": 0.04189350811168446,
"se": 0.04510789905267865,
"candidate": -0.6376950507984388,
"stock": -0.6795885589101233
},
"fidelity": 1.0738140872371999e-06,
"lengths": {
"solo": {
"ratio": 0.8621914985455672,
"low": 0.5983891587109743,
"high": 1.12599383838016
},
"team": {
"ratio": 0.9177590006154016,
"low": 0.7768183446549689,
"high": 1.0586996565758344
},
"busy": {
"ratio": 1.053547510111891,
"low": 0.9174334225599646,
"high": 1.1896615976638172
}
},
"memory": {
"ok": true,
"peak_mib": 165084.0,
"allowed_mib": 169119.1
},
"acceptance": {
"solo": 1.9575545171339563,
"team": 1.9112284235751744,
"busy": 1.9224493281968524
},
"limits": {
"greedy_agreement_floor": null,
"likelihood_bound": 0.05,
"fidelity_ceiling": 0.002,
"length_margin": 0.05,
"memory_slack_mib": 4096
},
"reason": null,
"score": 0.8352076640034652
}