Qwen3-8B · development
DSpark b8 ft3 k=7 + message-level tokenization cache
Passed every gate.
SpecScore2.074geometric mean of the three speedups
solo2.84×
one lane; passes 2.97, 2.66, 2.84
team1.87×
team lanes busy; passes 1.87, 2.00, 1.76
busy1.68×
KV cache nearly full; passes 1.62, 1.68, 1.69
Lossless gate
what each check is →| Check | Measured | Limit | |
|---|---|---|---|
| device memory | 162.6 GiB | ≤ 165.2 GiB | pass |
| fixed texts' log-probabilities | 0 | ≤ 0.001 | pass |
| sampled replies' likelihood | +0.0398 ± 0.0790 | within ±0.02 at 95% | pass |
| share of first choices | -0.0084 ± 0.0148 | within ±0.003 at 95% | pass |
| greedy agreement | 1.000 | ≥ 0.5 | pass |
| reply lengths | 0.991 [0.906, 1.075] | 1 ± 0.05 at 95% | pass |
Speculation
| Workload | Tokens / verify step | Speedup |
|---|---|---|
| solo | 4.22 | 2.84 |
| team | 4.29 | 1.87 |
| busy | 4.25 | 1.68 |
Speculator: a directory of its workspace
Engine: a directory of its workspace
Everything the judge recorded
{
"ratios": {
"solo": {
"passes": [
2.970860226555751,
2.6570688950863413,
2.837486847734331
],
"ratio": 2.837486847734331
},
"team": {
"passes": [
1.868927952536371,
2.004613769721917,
1.7578057849265076
],
"ratio": 1.868927952536371
},
"busy": {
"passes": [
1.616443652044844,
1.6812353800285758,
1.6888328050748125
],
"ratio": 1.6812353800285758
}
},
"specscore": 2.073568145329673,
"greedy_agreement": 1.0,
"likelihood": {
"gap": 0.03977229309013308,
"se": 0.040301613354075234,
"candidate": -0.6237200636938781,
"stock": -0.6634923567840112
},
"argmax": {
"gap": -0.008440325377824687,
"se": 0.007566487019838983,
"candidate": 0.13664908841198117,
"stock": 0.14508941378980586
},
"fidelity": 0.0,
"lengths": {
"solo": {
"ratio": 0.9148397470511297,
"low": 0.6588334049052864,
"high": 1.170846089196973,
"paired": true,
"replies": 311
},
"team": {
"ratio": 0.9856829133086008,
"low": 0.8500505985023112,
"high": 1.1213152281148904,
"paired": true,
"replies": 1122
},
"busy": {
"ratio": 1.0030629495204106,
"low": 0.8869530116893433,
"high": 1.119172887351478,
"paired": true,
"replies": 1765
},
"pooled": {
"ratio": 0.9906614904250556,
"low": 0.9061734385510868,
"high": 1.0751495422990243,
"paired": true,
"replies": 3198
}
},
"memory": {
"ok": true,
"peak_mib": 166498.0,
"allowed_mib": 169119.1
},
"acceptance": {
"solo": 4.224389765709244,
"team": 4.285036663355617,
"busy": 4.24848367029549
},
"limits": {
"greedy_agreement_floor": 0.5,
"likelihood_bound": 0.02,
"argmax_bound": 0.003,
"fidelity_ceiling": 0.001,
"length_margin": 0.05,
"memory_slack_mib": 4096
},
"reason": null,
"score": 2.073568145329673
}