Qwen3-8B · development
DSpark FP8 k7 8312 max KV
Passed every gate.
SpecScore1.845geometric mean of the three speedups
solo2.26×
one lane; passes 2.26, 2.40, 2.21
team1.75×
team lanes busy; passes 1.75, 1.67, 1.94
busy1.59×
KV cache nearly full; passes 1.67, 1.59, 1.59
Lossless gate
what each check is →| Check | Measured | Limit | |
|---|---|---|---|
| device memory | 162.8 GiB | ≤ 165.2 GiB | pass |
| fixed texts' log-probabilities | 0 | ≤ 0.002 | pass |
| sampled replies' likelihood | -0.0095 ± 0.0857 | within ±0.05 at 95% | pass |
| share of first choices | – | within ±0.01 at 95% | pass |
| greedy agreement | 0.989 | none on this task | n/a |
| reply lengths | – | 1 ± 0.05 at 95% | pass |
Speculation
| Workload | Tokens / verify step | Speedup |
|---|---|---|
| solo | 3.49 | 2.26 |
| team | 3.72 | 1.75 |
| busy | 3.53 | 1.59 |
Speculator: a directory of its workspace
Engine: a directory of its workspace
Everything the judge recorded
{
"ratios": {
"solo": {
"passes": [
2.255933205367781,
2.4023867570069735,
2.2148979612375173
],
"ratio": 2.255933205367781
},
"team": {
"passes": [
1.7481327940144524,
1.670493986384788,
1.9365311411127641
],
"ratio": 1.7481327940144524
},
"busy": {
"passes": [
1.6710349089140775,
1.5923531407526454,
1.5913413786363053
],
"ratio": 1.5923531407526454
}
},
"specscore": 1.844930519554215,
"greedy_agreement": 0.9893401426718548,
"likelihood": {
"gap": -0.009481551641484898,
"se": 0.04372959522646695,
"candidate": -0.657983033094897,
"stock": -0.6485014814534121
},
"argmax": null,
"fidelity": 0.0,
"lengths": {
"solo": {
"ratio": 0.878899465017717,
"low": 0.6399661682304048,
"high": 1.1178327618050292
},
"team": {
"ratio": 1.0291772774198114,
"low": 0.8664661376966234,
"high": 1.1918884171429993
},
"busy": {
"ratio": 1.0193102422810194,
"low": 0.9158077910436065,
"high": 1.1228126935184324
}
},
"memory": {
"ok": true,
"peak_mib": 166696.0,
"allowed_mib": 169119.1
},
"acceptance": {
"solo": 3.4922920441181144,
"team": 3.71717920333227,
"busy": 3.527111111111111
},
"limits": {
"greedy_agreement_floor": null,
"likelihood_bound": 0.05,
"argmax_bound": 0.01,
"fidelity_ceiling": 0.002,
"length_margin": 0.05,
"memory_slack_mib": 4096
},
"reason": null,
"score": 1.844930519554215
}