exp: 多段推理代价实测(2段+5.97ms/4段+17.7ms/8段+42ms,B=1 K=8)——靠段数补偏差不进 20ms 预算

This commit is contained in:
cbb
2026-10-20 22:50:41 +08:00
parent b30c271896
commit c9344599af
2 changed files with 626 additions and 0 deletions
+611
View File
@@ -0,0 +1,611 @@
{
"device": "NVIDIA GeForce RTX 3090",
"B": 1,
"K": 8,
"rounds": 24,
"per_round": {
"1": [
{
"p50": 9.198,
"p95": 11.654,
"p99": 12.236,
"max": 12.236
},
{
"p50": 9.211,
"p95": 13.105,
"p99": 13.765,
"max": 13.765
},
{
"p50": 9.188,
"p95": 12.363,
"p99": 15.901,
"max": 15.901
},
{
"p50": 8.894,
"p95": 10.426,
"p99": 11.641,
"max": 11.641
},
{
"p50": 8.968,
"p95": 11.664,
"p99": 11.874,
"max": 11.874
},
{
"p50": 9.456,
"p95": 13.502,
"p99": 13.951,
"max": 13.951
},
{
"p50": 8.683,
"p95": 10.991,
"p99": 11.978,
"max": 11.978
},
{
"p50": 9.599,
"p95": 13.894,
"p99": 17.102,
"max": 17.102
},
{
"p50": 9.405,
"p95": 13.07,
"p99": 13.111,
"max": 13.111
},
{
"p50": 8.947,
"p95": 10.154,
"p99": 10.405,
"max": 10.405
},
{
"p50": 9.019,
"p95": 10.835,
"p99": 12.272,
"max": 12.272
},
{
"p50": 8.824,
"p95": 13.516,
"p99": 15.963,
"max": 15.963
},
{
"p50": 8.819,
"p95": 11.004,
"p99": 11.308,
"max": 11.308
},
{
"p50": 8.777,
"p95": 10.7,
"p99": 11.144,
"max": 11.144
},
{
"p50": 8.926,
"p95": 11.911,
"p99": 12.171,
"max": 12.171
},
{
"p50": 8.924,
"p95": 10.88,
"p99": 11.94,
"max": 11.94
},
{
"p50": 8.508,
"p95": 9.918,
"p99": 11.57,
"max": 11.57
},
{
"p50": 9.351,
"p95": 17.795,
"p99": 17.948,
"max": 17.948
},
{
"p50": 8.779,
"p95": 10.188,
"p99": 11.614,
"max": 11.614
},
{
"p50": 9.389,
"p95": 19.998,
"p99": 25.582,
"max": 25.582
},
{
"p50": 9.344,
"p95": 15.293,
"p99": 16.308,
"max": 16.308
},
{
"p50": 9.67,
"p95": 13.045,
"p99": 13.805,
"max": 13.805
},
{
"p50": 10.331,
"p95": 13.905,
"p99": 21.628,
"max": 21.628
},
{
"p50": 8.96,
"p95": 10.848,
"p99": 11.741,
"max": 11.741
}
],
"2": [
{
"p50": 15.201,
"p95": 18.17,
"p99": 19.766,
"max": 19.766
},
{
"p50": 14.902,
"p95": 17.614,
"p99": 18.457,
"max": 18.457
},
{
"p50": 15.324,
"p95": 18.549,
"p99": 19.078,
"max": 19.078
},
{
"p50": 15.021,
"p95": 19.99,
"p99": 21.951,
"max": 21.951
},
{
"p50": 14.689,
"p95": 18.919,
"p99": 19.926,
"max": 19.926
},
{
"p50": 16.051,
"p95": 32.815,
"p99": 33.077,
"max": 33.077
},
{
"p50": 14.717,
"p95": 17.324,
"p99": 17.522,
"max": 17.522
},
{
"p50": 15.032,
"p95": 18.216,
"p99": 19.497,
"max": 19.497
},
{
"p50": 15.555,
"p95": 19.893,
"p99": 21.052,
"max": 21.052
},
{
"p50": 14.499,
"p95": 16.796,
"p99": 16.882,
"max": 16.882
},
{
"p50": 14.645,
"p95": 17.099,
"p99": 17.543,
"max": 17.543
},
{
"p50": 14.358,
"p95": 21.149,
"p99": 24.103,
"max": 24.103
},
{
"p50": 14.984,
"p95": 18.661,
"p99": 20.683,
"max": 20.683
},
{
"p50": 14.502,
"p95": 17.139,
"p99": 17.798,
"max": 17.798
},
{
"p50": 14.947,
"p95": 16.647,
"p99": 16.766,
"max": 16.766
},
{
"p50": 14.786,
"p95": 17.521,
"p99": 19.179,
"max": 19.179
},
{
"p50": 14.749,
"p95": 19.165,
"p99": 25.119,
"max": 25.119
},
{
"p50": 16.221,
"p95": 28.768,
"p99": 29.582,
"max": 29.582
},
{
"p50": 14.652,
"p95": 28.668,
"p99": 30.074,
"max": 30.074
},
{
"p50": 17.139,
"p95": 32.269,
"p99": 32.377,
"max": 32.377
},
{
"p50": 14.907,
"p95": 20.386,
"p99": 20.81,
"max": 20.81
},
{
"p50": 15.882,
"p95": 17.677,
"p99": 18.192,
"max": 18.192
},
{
"p50": 15.359,
"p95": 18.135,
"p99": 18.531,
"max": 18.531
},
{
"p50": 15.207,
"p95": 17.382,
"p99": 18.529,
"max": 18.529
}
],
"4": [
{
"p50": 26.665,
"p95": 31.215,
"p99": 34.178,
"max": 34.178
},
{
"p50": 27.294,
"p95": 33.673,
"p99": 35.04,
"max": 35.04
},
{
"p50": 27.918,
"p95": 34.553,
"p99": 42.061,
"max": 42.061
},
{
"p50": 26.37,
"p95": 30.528,
"p99": 32.127,
"max": 32.127
},
{
"p50": 25.9,
"p95": 28.742,
"p99": 34.469,
"max": 34.469
},
{
"p50": 27.106,
"p95": 35.303,
"p99": 40.808,
"max": 40.808
},
{
"p50": 26.195,
"p95": 34.707,
"p99": 45.867,
"max": 45.867
},
{
"p50": 26.772,
"p95": 30.526,
"p99": 30.766,
"max": 30.766
},
{
"p50": 26.886,
"p95": 46.087,
"p99": 50.21,
"max": 50.21
},
{
"p50": 26.004,
"p95": 29.001,
"p99": 30.311,
"max": 30.311
},
{
"p50": 27.095,
"p95": 32.363,
"p99": 38.011,
"max": 38.011
},
{
"p50": 25.952,
"p95": 30.357,
"p99": 31.543,
"max": 31.543
},
{
"p50": 26.537,
"p95": 30.396,
"p99": 31.056,
"max": 31.056
},
{
"p50": 26.563,
"p95": 28.768,
"p99": 30.415,
"max": 30.415
},
{
"p50": 26.863,
"p95": 32.136,
"p99": 32.967,
"max": 32.967
},
{
"p50": 28.413,
"p95": 50.4,
"p99": 58.088,
"max": 58.088
},
{
"p50": 26.623,
"p95": 32.801,
"p99": 35.603,
"max": 35.603
},
{
"p50": 27.344,
"p95": 32.779,
"p99": 38.661,
"max": 38.661
},
{
"p50": 26.408,
"p95": 29.565,
"p99": 30.094,
"max": 30.094
},
{
"p50": 26.198,
"p95": 29.741,
"p99": 30.489,
"max": 30.489
},
{
"p50": 27.724,
"p95": 32.174,
"p99": 32.423,
"max": 32.423
},
{
"p50": 26.724,
"p95": 31.253,
"p99": 32.153,
"max": 32.153
},
{
"p50": 26.76,
"p95": 31.054,
"p99": 31.087,
"max": 31.087
},
{
"p50": 26.278,
"p95": 31.043,
"p99": 39.867,
"max": 39.867
}
],
"8": [
{
"p50": 52.266,
"p95": 68.035,
"p99": 74.885,
"max": 74.885
},
{
"p50": 54.125,
"p95": 85.092,
"p99": 96.265,
"max": 96.265
},
{
"p50": 56.176,
"p95": 85.733,
"p99": 87.312,
"max": 87.312
},
{
"p50": 51.069,
"p95": 60.145,
"p99": 62.654,
"max": 62.654
},
{
"p50": 51.07,
"p95": 61.397,
"p99": 65.437,
"max": 65.437
},
{
"p50": 52.133,
"p95": 69.749,
"p99": 79.426,
"max": 79.426
},
{
"p50": 49.955,
"p95": 60.339,
"p99": 60.455,
"max": 60.455
},
{
"p50": 50.468,
"p95": 64.167,
"p99": 66.843,
"max": 66.843
},
{
"p50": 49.244,
"p95": 52.489,
"p99": 54.475,
"max": 54.475
},
{
"p50": 49.644,
"p95": 57.36,
"p99": 58.934,
"max": 58.934
},
{
"p50": 52.186,
"p95": 59.122,
"p99": 68.662,
"max": 68.662
},
{
"p50": 51.393,
"p95": 75.776,
"p99": 83.858,
"max": 83.858
},
{
"p50": 51.953,
"p95": 67.169,
"p99": 69.287,
"max": 69.287
},
{
"p50": 50.324,
"p95": 60.185,
"p99": 68.603,
"max": 68.603
},
{
"p50": 50.893,
"p95": 54.805,
"p99": 58.967,
"max": 58.967
},
{
"p50": 53.696,
"p95": 65.629,
"p99": 72.674,
"max": 72.674
},
{
"p50": 48.672,
"p95": 62.026,
"p99": 62.276,
"max": 62.276
},
{
"p50": 50.777,
"p95": 65.002,
"p99": 66.318,
"max": 66.318
},
{
"p50": 48.835,
"p95": 57.514,
"p99": 69.18,
"max": 69.18
},
{
"p50": 50.785,
"p95": 73.072,
"p99": 75.93,
"max": 75.93
},
{
"p50": 50.684,
"p95": 65.827,
"p99": 75.478,
"max": 75.478
},
{
"p50": 50.341,
"p95": 57.511,
"p99": 58.992,
"max": 58.992
},
{
"p50": 51.567,
"p95": 66.644,
"p99": 67.094,
"max": 67.094
},
{
"p50": 52.094,
"p95": 71.313,
"p99": 73.767,
"max": 73.767
}
]
},
"summary": {
"p501": 9.019,
"p951": 11.911,
"p991": 12.272,
"p502": 14.984,
"p952": 18.549,
"p992": 19.766,
"p504": 26.724,
"p954": 31.253,
"p994": 34.178,
"p508": 51.069,
"p958": 65.002,
"p998": 68.662
},
"delta_vs_1step_ms": {
"d502": 5.965,
"d504": 17.705,
"d508": 42.05
}
}
+15
View File
@@ -107,6 +107,21 @@
- **对 `backoff` 的预期与判据**:全拒意味着"连第 1 步的端点或中点已经穿进障碍",缩步能不能救取决于**穿得有多深**(穿 0.1 mm 救得回,穿 3 mm 救不回)。这个数转储里没记(只记被选中那条),所以直接由 `backoff_mean`/`reject_all_mean` 来量:
`reject_all_mean ≈ 原全拒次数`(=backoff_mean+brakes_mean,与轮次 33 cap 的 19.55 可比),`backoff_mean/(backoff_mean+brakes_mean)` 就是"缩步能救回的比例"。这是补跑(#22)要读的第一个数。
## 多段推理的代价实测(B=1、K=8、逐轮交替、奇数轮反序)
- `docs/experiment_round38_steps_cost.json`(24 轮 × 每轮 40 次,取每档的中位):
| 段数 | 推理 P50 | 相对 1 段增量 |
|---|---|---|
| 1 | 9.02 ms | — |
| 2 | 14.98 ms | **+5.97 ms** |
| 4 | 26.72 ms | +17.71 ms |
| 8 | 51.07 ms | +42.05 ms |
- 每多一段 ≈ 多一次解码器前向(编码器 `mem` 只算一次),所以增量近似线性;9.02 ms 这一档比轮次 33 报的 7.2 ms 高,是不同会话的时钟/上下文差别,**只用同批内的差值,不用绝对值跨轮比**。
- 合到整周期上:轮次 34 的现口径 P50 ≈ 33.5 ms,两段就是 ≈39.5 ms(25 Hz),四段 ≈51 ms(19.6 Hz)——**离 20 ms 预算更远**,所以"靠多段推理把一步的偏差补回来"在部署上不成立,只能作为开环诊断口径。
- 与 K 这条轴叠加会更糟(K=16 认证点翻倍),所以 `K=8/16 × cert_batch=0/1` 的四格周期微基准排在当前批次里(`ml/runs/_kcost_probe.py`),这是判断"候选数能不能买到部署里的到达率"的关键一列。
## 候选为什么会"塌":两段积分把 K 条候选吸成一条(轮次 39 探针,机理定案)
- `ml/runs/_cand_div.py`(256 题特征、同一批噪声 z、K=8,只改段数):