Qwen3-8B · development
DSpark FP8 draft k7 dynamic
Lossless gate failed: held more device memory than allowed. SpecScore 0.
SpecScore0.000geometric mean of the three speedups
solo2.22×
one lane; passes 2.13, 2.22, 2.40
team1.81×
team lanes busy; passes 1.82, 1.81, 1.68
busy1.58×
KV cache nearly full; passes 1.51, 1.58, 1.61
Lossless gate
what each check is →| Check | Measured | Limit | |
|---|---|---|---|
| device memory | 166.2 GiB | ≤ 165.2 GiB | fail |
| fixed texts' log-probabilities | 0 | ≤ 0.002 | pass |
| sampled replies' likelihood | -0.0221 ± 0.0755 | within ±0.05 at 95% | pass |
| share of first choices | – | within ±0.01 at 95% | pass |
| greedy agreement | 1.000 | none on this task | n/a |
| reply lengths | – | 1 ± 0.05 at 95% | pass |
Speculation
| Workload | Tokens / verify step | Speedup |
|---|---|---|
| solo | 3.35 | 2.22 |
| team | 3.61 | 1.81 |
| busy | 3.55 | 1.58 |
Speculator: deepseek-ai/dspark_qwen3_8b_block7@03326e5
Engine: a directory of its workspace
Everything the judge recorded
{
"ratios": {
"solo": {
"passes": [
2.1311064776229878,
2.2180217317333137,
2.39518853198994
],
"ratio": 2.2180217317333137
},
"team": {
"passes": [
1.8235077749624644,
1.8088045515958697,
1.6809623781032121
],
"ratio": 1.8088045515958697
},
"busy": {
"passes": [
1.510425231011272,
1.5780494475578488,
1.608938760132175
],
"ratio": 1.5780494475578488
}
},
"specscore": 1.849947266727779,
"greedy_agreement": 1.0,
"likelihood": {
"gap": -0.02206277593036643,
"se": 0.03851892665644316,
"candidate": -0.6176142847648748,
"stock": -0.5955515088345084
},
"argmax": null,
"fidelity": 0.0,
"lengths": {
"solo": {
"ratio": 1.1211127783471635,
"low": 0.7804891908096245,
"high": 1.4617363658847025
},
"team": {
"ratio": 0.9472446169591768,
"low": 0.8013965139340106,
"high": 1.093092719984343
},
"busy": {
"ratio": 0.998905846376052,
"low": 0.8976046781968803,
"high": 1.1002070145552236
}
},
"memory": {
"ok": false,
"peak_mib": 170238.0,
"allowed_mib": 169119.1
},
"acceptance": {
"solo": 3.354454269038549,
"team": 3.6076071389225515,
"busy": 3.5477870947806056
},
"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": "held more device memory than allowed",
"score": 0.0
}