Files
Robot_Arm/docs/experiment_round33_gpu_cycle_steps1.json
T
cbb 48cc566c3f 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
2026-10-20 20:32:19 +08:00

143 lines
2.9 KiB
JSON

{
"device": "NVIDIA GeForce RTX 3090",
"B": 1,
"steps_used_by_eval": 1,
"note": "eval_cl \u4ece\u4e0d\u4f20 steps\uff0crollout \u9ed8\u8ba4 steps=1\uff08MeanFlow \u4e00\u6b65\u751f\u6210\uff09\uff1b\u6240\u4ee5\u771f\u5b9e\u5468\u671f\u6309 steps=1 \u6d4b",
"parts": {
"features_total": {
"p50": 12.589,
"p95": 18.334,
"p99": 20.217,
"max": 23.649,
"reps": 120
},
"features_gm_groups_only": {
"p50": 6.817,
"p95": 8.771,
"p99": 10.254,
"max": 17.169,
"reps": 120
},
"features_gaps40_all_groups": {
"p50": 7.245,
"p95": 10.501,
"p99": 13.794,
"max": 14.508,
"reps": 120
},
"features_fk_only": {
"p50": 3.708,
"p95": 5.024,
"p99": 5.991,
"max": 6.301,
"reps": 120
},
"sample_k_K8_steps1": {
"p50": 7.799,
"p95": 11.655,
"p99": 13.342,
"max": 16.212,
"reps": 120
},
"sample_k_K1_steps1": {
"p50": 7.205,
"p95": 8.935,
"p99": 9.615,
"max": 13.92,
"reps": 120
},
"sample_k_K4_steps1": {
"p50": 7.834,
"p95": 10.354,
"p99": 14.744,
"max": 15.241,
"reps": 120
},
"dls_fb_f64": {
"p50": 6.729,
"p95": 9.297,
"p99": 11.186,
"max": 12.064,
"reps": 120
},
"discrete_cert_4pts": {
"p50": 8.603,
"p95": 11.181,
"p99": 12.464,
"max": 15.799,
"reps": 120
},
"cert_uniform_n8_2roots_f64": {
"p50": 11.479,
"p95": 19.221,
"p99": 20.942,
"max": 21.183,
"reps": 120
},
"cert_uniform_n16_2roots_f64": {
"p50": 16.138,
"p95": 21.972,
"p99": 23.898,
"max": 25.793,
"reps": 120
}
},
"cycles": {
"cycle_K8_full_discrete": {
"p50": 40.676,
"p95": 51.742,
"p99": 59.782,
"max": 62.074,
"reps": 120
},
"cycle_K4_full_discrete": {
"p50": 43.997,
"p95": 56.145,
"p99": 59.905,
"max": 65.882,
"reps": 120
},
"cycle_K1_full_discrete": {
"p50": 37.87,
"p95": 45.754,
"p99": 49.394,
"max": 53.139,
"reps": 120
},
"cycle_K8_noscore_discrete": {
"p50": 37.427,
"p95": 49.602,
"p99": 60.777,
"max": 70.139,
"reps": 120
},
"cycle_K8_full_cert_n8": {
"p50": 43.984,
"p95": 65.795,
"p99": 75.53,
"max": 95.452,
"reps": 120
},
"cycle_K8_full_cert_n16": {
"p50": 51.054,
"p95": 74.862,
"p99": 95.768,
"max": 106.48,
"reps": 120
},
"cycle_K8_fused_estimate": {
"p50": 24.054,
"p95": 26.595,
"p99": 30.17,
"max": 33.284,
"reps": 120
},
"cycle_K8_fused_cert_n8": {
"p50": 34.274,
"p95": 46.329,
"p99": 72.737,
"max": 79.022,
"reps": 120
}
}
}