exp: 轮次33——限幅实测要付5.1pp到达(57.4→52.3%);GPU真实周期36.7ms≈27Hz,瓶颈是重复几何而非推理

- 四组单变量对照(同256题同seed):base / inc_cap / gap_margin 0.5 / 两者同开
  限幅后 到达 0.5742→0.5234、time_ok 0.3906→0.2656、2s误差比 0.210→0.279、超VMAX 1.0→0.0、刹停几乎不变
  → 之前那部分到达率是超速换来的;损失在逼近末段,不在刹停
- 0.5mm 余量门只付 3 题(0.5742→0.5625);但 plans['gaps'] 在 0.5mm 截顶,5mm 护栏可达性需抬 cap 重算
- 前提更正:eval_cl 不传 steps、rollout 默认 steps=1(policy.py:156 一步生成),先前按 steps=16 得到的
  "当前配方 116ms≈8.6Hz" 作废;真实配方 36.68ms P50 / 49.17ms P99 ≈ 27.3Hz
- 归因更正:空算子 0.028ms、空 sync 0.005ms、encode 2.27ms → 不是启动/同步地板;
  单步推理仅 7.2ms 且 K=1→8 免费,周期里 ~27ms 是 cum_xf+宽相位的重复计算
- certify_uniform 挂上后 36.7→45.3ms(n=8)/49.6ms(n=16)
- kin_torch 补 a/b 形状守卫(脚本里广播不一致会算错 view),测试 30→31 项
This commit is contained in:
cbb
2026-10-20 20:32:19 +08:00
parent 6ecd33db0c
commit 48cc566c3f
8 changed files with 610 additions and 1 deletions
+202
View File
@@ -0,0 +1,202 @@
{
"device": "NVIDIA GeForce RTX 3090",
"torch": "2.10.0+cu130",
"act_scale": 1.0,
"cfg": {},
"B": 1,
"commit_n": 2,
"fp64_note": "3090 fp64 \u541e\u5410\u7ea6 fp32 \u7684 1/54\uff1b\u8ba4\u8bc1\u76ee\u524d\u8981\u6c42 fp64",
"inference": {
"sample_K1_steps1": {
"p50": 7.453,
"p95": 9.912,
"p99": 10.766,
"max": 11.109,
"reps": 120
},
"sample_K8_steps1": {
"p50": 8.208,
"p95": 10.685,
"p99": 11.22,
"max": 15.224,
"reps": 120
},
"sample_K1_steps2": {
"p50": 13.149,
"p95": 17.621,
"p99": 18.438,
"max": 19.296,
"reps": 120
},
"sample_K8_steps2": {
"p50": 14.167,
"p95": 19.241,
"p99": 24.77,
"max": 25.703,
"reps": 120
},
"sample_K1_steps4": {
"p50": 19.193,
"p95": 22.338,
"p99": 23.666,
"max": 25.787,
"reps": 120
},
"sample_K8_steps4": {
"p50": 21.509,
"p95": 24.732,
"p99": 35.655,
"max": 41.514,
"reps": 120
},
"sample_K1_steps8": {
"p50": 40.658,
"p95": 55.162,
"p99": 60.394,
"max": 60.398,
"reps": 120
},
"sample_K8_steps8": {
"p50": 40.416,
"p95": 47.997,
"p99": 62.49,
"max": 71.976,
"reps": 120
},
"sample_K1_steps16": {
"p50": 72.429,
"p95": 90.188,
"p99": 107.536,
"max": 112.674,
"reps": 120
},
"sample_K8_steps16": {
"p50": 79.893,
"p95": 99.933,
"p99": 105.746,
"max": 106.078,
"reps": 120
}
},
"encode_once": {
"p50": 2.265,
"p95": 4.001,
"p99": 6.086,
"max": 6.136,
"reps": 120
},
"tiny_op_ms": 0.028,
"sync_only_ms": 0.005,
"components": {
"features_f32": {
"p50": 12.443,
"p95": 14.227,
"p99": 19.678,
"max": 20.531,
"reps": 120
},
"endpoint_score_K8_f32": {
"p50": 4.278,
"p95": 5.841,
"p99": 6.322,
"max": 8.883,
"reps": 120
},
"dls_fb_1step_f64": {
"p50": 7.191,
"p95": 12.26,
"p99": 14.642,
"max": 15.384,
"reps": 120
},
"dls_fb_1step_f32": {
"p50": 7.619,
"p95": 10.142,
"p99": 11.597,
"max": 12.022,
"reps": 120
},
"discrete_cert_4pts_f32": {
"p50": 8.214,
"p95": 13.365,
"p99": 14.235,
"max": 14.991,
"reps": 120
},
"cert_n8_f64_2roots": {
"p50": 11.265,
"p95": 14.64,
"p99": 16.788,
"max": 17.761,
"reps": 120
},
"cert_n16_f64_2roots": {
"p50": 14.574,
"p95": 16.731,
"p99": 18.459,
"max": 22.346,
"reps": 120
},
"cert_n32_f64_2roots": {
"p50": 21.528,
"p95": 24.746,
"p99": 28.037,
"max": 28.066,
"reps": 120
}
},
"cycle_K8_steps16_discrete": {
"p50": 116.474,
"p95": 134.474,
"p99": 142.593,
"max": 152.768,
"reps": 120
},
"cycle_K8_steps4_discrete": {
"p50": 55.919,
"p95": 65.767,
"p99": 74.809,
"max": 78.059,
"reps": 120
},
"cycle_K1_steps4_discrete": {
"p50": 51.087,
"p95": 58.614,
"p99": 62.326,
"max": 63.919,
"reps": 120
},
"cycle_K1_steps1_discrete": {
"p50": 34.493,
"p95": 42.078,
"p99": 43.906,
"max": 50.268,
"reps": 120
},
"cycle_K8_steps16_cert_n8": {
"p50": 132.027,
"p95": 150.463,
"p99": 160.83,
"max": 192.586,
"reps": 120
},
"cycle_K8_steps4_cert_n8": {
"p50": 64.812,
"p95": 79.192,
"p99": 89.629,
"max": 95.837,
"reps": 120
},
"cycle_K1_steps4_cert_n16": {
"p50": 61.967,
"p95": 73.219,
"p99": 79.053,
"max": 83.18,
"reps": 120
},
"f32_vs_f64": {
"points": 3,
"max_abs_err_mm": 0.0,
"sign_flips": 0
}
}