# Cohort private access 完整结果

FP32，W8，block=32，legacy task grain，predicated effects 开、full-packet 开、fusion 关、fast math 关。时间单位 µs。两种计时边界不能混用；配对比值是逐轮比值的中位数，不是表中中位时间相除。

## 单线程纯 native entry：RMSNorm

实际 ORC / TorchInductor 入口；6 个平衡顺序，每次 7 个样本。不含 Runtime、Python、分配；保留原生函数调用、Luisa launch record 重置及 LLVM 生成的 libc 调用。

| 尺寸 | baseline | candidate | Inductor | candidate/baseline | candidate/Inductor | 后者逐轮范围 |
|---|---:|---:|---:|---:|---:|---:|
| 17×65 | 1.364 | 0.249 | 0.729 | 0.182 | 0.340 | 0.339–0.342 |
| 257×1538 | 544.018 | 116.720 | 247.348 | 0.213 | 0.469 | 0.467–0.478 |
| 1024×4097 | 5784.021 | 1148.873 | 2593.370 | 0.198 | 0.442 | 0.426–0.478 |
| 129×768 | 26.311 | 26.238 | 63.246 | 0.998 | 0.416 | 0.364–0.429 |
| 137×1023 | 193.905 | 40.881 | 86.733 | 0.211 | 0.471 | 0.454–0.473 |
| 513×2051 | 1456.840 | 283.972 | 655.054 | 0.197 | 0.436 | 0.430–0.449 |

## Runtime E2E：全部 110 个固定映射对照

8 个请求的 CPU workers；2 个反向顺序，每次 7 个样本；小差异与相反方向的轮次不能视为稳定收益。local=1 为 whole-program，local=8 为 packet-local；attention 仅测 local=1。RoPE 要求偶数宽度，奇数测试点向上取偶，开关两侧相同。

| 算子 / 尺寸 | local | off | on | on/off 配对比 | 配对范围 | direct CFG off→on |
|---|---:|---:|---:|---:|---:|---|
| rmsnorm 3×63 | 1 | 0.569 | 0.563 | 0.990 | 0.978–1.002 | [True]→[True] |
| rmsnorm 3×63 | 8 | 0.195 | 0.198 | 1.014 | 1.009–1.018 | [True]→[True] |
| rmsnorm 17×65 | 1 | 1.371 | 1.335 | 0.975 | 0.942–1.007 | [True]→[True] |
| rmsnorm 17×65 | 8 | 30.472 | 32.095 | 1.053 | 1.053–1.054 | [True]→[True] |
| rmsnorm 64×256 | 1 | 36.435 | 35.976 | 0.988 | 0.985–0.990 | [True]→[True] |
| rmsnorm 64×256 | 8 | 38.509 | 38.037 | 0.988 | 0.983–0.992 | [True]→[True] |
| rmsnorm 129×769 | 1 | 67.689 | 67.746 | 1.001 | 0.986–1.016 | [True]→[True] |
| rmsnorm 129×769 | 8 | 76.905 | 57.396 | 0.746 | 0.743–0.749 | [True]→[True] |
| rmsnorm 257×1538 | 1 | 114.649 | 113.796 | 0.993 | 0.988–0.997 | [True]→[True] |
| rmsnorm 257×1538 | 8 | 133.717 | 80.799 | 0.604 | 0.598–0.610 | [True]→[True] |
| rmsnorm 1024×4097 | 1 | 780.405 | 609.955 | 0.817 | 0.652–0.981 | [True]→[True] |
| rmsnorm 1024×4097 | 8 | 889.739 | 307.457 | 0.346 | 0.342–0.349 | [True]→[True] |
| rmsnorm 17×16384 | 1 | 293.203 | 291.519 | 0.994 | 0.990–0.998 | [True]→[True] |
| rmsnorm 17×16384 | 8 | 72.128 | 71.891 | 0.997 | 0.988–1.006 | [True]→[True] |
| rmsnorm 137×1023 | 1 | 73.508 | 73.007 | 0.994 | 0.964–1.024 | [True]→[True] |
| rmsnorm 137×1023 | 8 | 84.595 | 63.663 | 0.753 | 0.751–0.754 | [True]→[True] |
| rmsnorm 513×2051 | 1 | 184.236 | 190.882 | 1.036 | 0.988–1.084 | [True]→[True] |
| rmsnorm 513×2051 | 8 | 259.897 | 116.422 | 0.448 | 0.447–0.449 | [True]→[True] |
| layernorm 3×63 | 1 | 0.670 | 0.668 | 0.996 | 0.995–0.997 | [True]→[True] |
| layernorm 3×63 | 8 | 0.228 | 0.230 | 1.007 | 1.004–1.010 | [True]→[True] |
| layernorm 17×65 | 1 | 1.589 | 1.601 | 1.008 | 1.006–1.010 | [True]→[True] |
| layernorm 17×65 | 8 | 34.391 | 31.641 | 0.921 | 0.882–0.960 | [True]→[True] |
| layernorm 64×256 | 1 | 40.409 | 41.327 | 1.023 | 1.022–1.023 | [True]→[True] |
| layernorm 64×256 | 8 | 45.780 | 47.899 | 1.046 | 1.042–1.049 | [True]→[True] |
| layernorm 129×769 | 1 | 71.517 | 72.043 | 1.007 | 0.999–1.016 | [True]→[True] |
| layernorm 129×769 | 8 | 99.994 | 63.510 | 0.636 | 0.610–0.662 | [True]→[True] |
| layernorm 257×1538 | 1 | 120.571 | 124.058 | 1.029 | 0.993–1.065 | [True]→[True] |
| layernorm 257×1538 | 8 | 216.179 | 87.412 | 0.404 | 0.404–0.405 | [True]→[True] |
| layernorm 1024×4097 | 1 | 753.993 | 735.210 | 0.976 | 0.955–0.996 | [True]→[True] |
| layernorm 1024×4097 | 8 | 1739.144 | 413.998 | 0.238 | 0.238–0.238 | [True]→[True] |
| layernorm 17×16384 | 1 | 397.520 | 393.387 | 0.990 | 0.984–0.995 | [True]→[True] |
| layernorm 17×16384 | 8 | 86.863 | 84.719 | 0.975 | 0.975–0.976 | [True]→[True] |
| layernorm 137×1023 | 1 | 80.946 | 80.834 | 0.999 | 0.993–1.004 | [True]→[True] |
| layernorm 137×1023 | 8 | 113.030 | 68.823 | 0.609 | 0.604–0.614 | [True]→[True] |
| layernorm 513×2051 | 1 | 225.012 | 213.674 | 0.951 | 0.915–0.986 | [True]→[True] |
| layernorm 513×2051 | 8 | 483.276 | 140.906 | 0.292 | 0.289–0.294 | [True]→[True] |
| masked_softmax 3×63 | 1 | 1.273 | 1.282 | 1.007 | 1.000–1.014 | [True]→[True] |
| masked_softmax 3×63 | 8 | 0.603 | 0.597 | 0.989 | 0.982–0.996 | [True]→[True] |
| masked_softmax 17×65 | 1 | 4.520 | 4.540 | 1.005 | 1.003–1.006 | [True]→[True] |
| masked_softmax 17×65 | 8 | 34.719 | 32.841 | 0.946 | 0.921–0.971 | [True]→[True] |
| masked_softmax 64×256 | 1 | 63.431 | 63.490 | 1.001 | 0.994–1.008 | [True]→[True] |
| masked_softmax 64×256 | 8 | 60.602 | 60.914 | 1.005 | 1.004–1.007 | [True]→[True] |
| masked_softmax 129×769 | 1 | 125.156 | 130.964 | 1.045 | 0.989–1.101 | [True]→[True] |
| masked_softmax 129×769 | 8 | 135.920 | 95.817 | 0.707 | 0.673–0.740 | [True]→[True] |
| masked_softmax 257×1538 | 1 | 255.160 | 246.570 | 0.966 | 0.960–0.972 | [True]→[True] |
| masked_softmax 257×1538 | 8 | 359.889 | 234.017 | 0.650 | 0.626–0.675 | [True]→[True] |
| masked_softmax 1024×4097 | 1 | 1994.611 | 2021.999 | 1.014 | 0.987–1.041 | [True]→[True] |
| masked_softmax 1024×4097 | 8 | 3661.417 | 1913.912 | 0.531 | 0.449–0.613 | [True]→[True] |
| masked_softmax 17×16384 | 1 | 1149.102 | 1148.863 | 1.000 | 0.999–1.001 | [True]→[True] |
| masked_softmax 17×16384 | 8 | 216.129 | 217.340 | 1.006 | 0.997–1.014 | [True]→[True] |
| masked_softmax 137×1023 | 1 | 146.938 | 147.386 | 1.003 | 0.985–1.021 | [True]→[True] |
| masked_softmax 137×1023 | 8 | 161.336 | 112.254 | 0.696 | 0.684–0.708 | [True]→[True] |
| masked_softmax 513×2051 | 1 | 537.747 | 540.222 | 1.005 | 1.002–1.007 | [True]→[True] |
| masked_softmax 513×2051 | 8 | 854.188 | 468.058 | 0.548 | 0.546–0.550 | [True]→[True] |
| swiglu 3×63 | 1 | 1.366 | 1.367 | 1.000 | 0.999–1.002 | [True]→[True] |
| swiglu 3×63 | 8 | 0.405 | 0.406 | 1.002 | 1.001–1.003 | [True]→[True] |
| swiglu 17×65 | 1 | 3.616 | 3.594 | 0.994 | 0.992–0.996 | [True]→[True] |
| swiglu 17×65 | 8 | 35.837 | 31.476 | 0.889 | 0.783–0.994 | [True]→[True] |
| swiglu 64×256 | 1 | 54.796 | 54.802 | 1.000 | 0.981–1.020 | [True]→[True] |
| swiglu 64×256 | 8 | 51.536 | 49.587 | 0.962 | 0.946–0.979 | [True]→[True] |
| swiglu 129×769 | 1 | 100.392 | 104.502 | 1.041 | 1.009–1.073 | [True]→[True] |
| swiglu 129×769 | 8 | 93.787 | 78.695 | 0.839 | 0.835–0.843 | [True]→[True] |
| swiglu 257×1538 | 1 | 225.489 | 213.243 | 0.950 | 0.880–1.020 | [True]→[True] |
| swiglu 257×1538 | 8 | 189.021 | 141.178 | 0.747 | 0.721–0.773 | [True]→[True] |
| swiglu 1024×4097 | 1 | 1512.170 | 1556.264 | 1.029 | 0.996–1.062 | [True]→[True] |
| swiglu 1024×4097 | 8 | 1566.149 | 934.094 | 0.597 | 0.588–0.606 | [True]→[True] |
| swiglu 17×16384 | 1 | 829.639 | 828.661 | 0.999 | 0.996–1.002 | [True]→[True] |
| swiglu 17×16384 | 8 | 147.506 | 141.886 | 0.962 | 0.955–0.969 | [True]→[True] |
| swiglu 137×1023 | 1 | 127.789 | 123.263 | 0.965 | 0.953–0.976 | [True]→[True] |
| swiglu 137×1023 | 8 | 101.404 | 83.709 | 0.826 | 0.814–0.837 | [True]→[True] |
| swiglu 513×2051 | 1 | 448.634 | 446.432 | 0.998 | 0.939–1.057 | [True]→[True] |
| swiglu 513×2051 | 8 | 420.742 | 273.622 | 0.650 | 0.649–0.652 | [True]→[True] |
| gelu_residual 3×63 | 1 | 1.317 | 1.319 | 1.002 | 0.998–1.005 | [True]→[True] |
| gelu_residual 3×63 | 8 | 0.650 | 0.644 | 0.991 | 0.990–0.992 | [True]→[True] |
| gelu_residual 17×65 | 1 | 5.899 | 5.922 | 1.004 | 1.000–1.008 | [True]→[True] |
| gelu_residual 17×65 | 8 | 34.095 | 32.047 | 0.940 | 0.920–0.960 | [True]→[True] |
| gelu_residual 64×256 | 1 | 67.364 | 66.974 | 0.995 | 0.967–1.024 | [True]→[True] |
| gelu_residual 64×256 | 8 | 60.331 | 58.110 | 0.963 | 0.938–0.988 | [True]→[True] |
| gelu_residual 129×769 | 1 | 135.381 | 135.277 | 0.999 | 0.983–1.015 | [True]→[True] |
| gelu_residual 129×769 | 8 | 108.631 | 90.349 | 0.832 | 0.828–0.835 | [True]→[True] |
| gelu_residual 257×1538 | 1 | 305.665 | 304.004 | 0.999 | 0.888–1.110 | [True]→[True] |
| gelu_residual 257×1538 | 8 | 272.326 | 199.230 | 0.732 | 0.716–0.747 | [True]→[True] |
| gelu_residual 1024×4097 | 1 | 2290.612 | 2300.396 | 1.004 | 1.000–1.009 | [True]→[True] |
| gelu_residual 1024×4097 | 8 | 2513.825 | 1566.668 | 0.624 | 0.607–0.640 | [True]→[True] |
| gelu_residual 17×16384 | 1 | 1426.344 | 1425.334 | 0.999 | 0.991–1.008 | [True]→[True] |
| gelu_residual 17×16384 | 8 | 218.803 | 223.659 | 1.022 | 1.009–1.035 | [True]→[True] |
| gelu_residual 137×1023 | 1 | 171.427 | 171.347 | 1.000 | 0.988–1.011 | [True]→[True] |
| gelu_residual 137×1023 | 8 | 133.672 | 103.449 | 0.774 | 0.771–0.777 | [True]→[True] |
| gelu_residual 513×2051 | 1 | 699.372 | 710.828 | 1.016 | 0.998–1.035 | [True]→[True] |
| gelu_residual 513×2051 | 8 | 676.205 | 451.931 | 0.668 | 0.668–0.669 | [True]→[True] |
| rope 3×64 | 1 | 0.603 | 0.598 | 0.992 | 0.992–0.992 | [True]→[True] |
| rope 3×64 | 8 | 0.182 | 0.182 | 1.000 | 0.997–1.003 | [True]→[True] |
| rope 17×66 | 1 | 1.923 | 1.916 | 0.996 | 0.993–1.000 | [True]→[True] |
| rope 17×66 | 8 | 35.584 | 36.076 | 1.014 | 0.996–1.033 | [True]→[True] |
| rope 64×256 | 1 | 39.056 | 39.258 | 1.005 | 0.988–1.022 | [True]→[True] |
| rope 64×256 | 8 | 39.599 | 38.811 | 0.980 | 0.963–0.997 | [True]→[True] |
| rope 129×770 | 1 | 70.274 | 70.907 | 1.009 | 0.989–1.029 | [True]→[True] |
| rope 129×770 | 8 | 79.278 | 52.976 | 0.668 | 0.662–0.675 | [True]→[True] |
| rope 257×1538 | 1 | 150.513 | 139.503 | 0.931 | 0.844–1.018 | [True]→[True] |
| rope 257×1538 | 8 | 148.122 | 82.762 | 0.559 | 0.548–0.570 | [True]→[True] |
| rope 1024×4098 | 1 | 850.139 | 857.141 | 1.008 | 0.999–1.017 | [True]→[True] |
| rope 1024×4098 | 8 | 1179.285 | 450.866 | 0.383 | 0.365–0.401 | [True]→[True] |
| rope 17×16384 | 1 | 358.378 | 357.159 | 0.997 | 0.990–1.003 | [True]→[True] |
| rope 17×16384 | 8 | 70.925 | 72.241 | 1.019 | 1.010–1.028 | [True]→[True] |
| rope 137×1024 | 1 | 78.305 | 81.034 | 1.035 | 0.978–1.092 | [True]→[True] |
| rope 137×1024 | 8 | 63.087 | 66.295 | 1.048 | 1.018–1.078 | [True]→[True] |
| rope 513×2052 | 1 | 250.757 | 239.681 | 0.958 | 0.908–1.009 | [True]→[True] |
| rope 513×2052 | 8 | 322.948 | 138.468 | 0.429 | 0.424–0.433 | [True]→[True] |
| attention 1×4×2×16×32×16×16 | 1 | 42.067 | 42.945 | 1.022 | 0.996–1.047 | [True]→[True] |
| attention 1×4×1×64×128×32×32 | 1 | 108.037 | 111.292 | 1.030 | 0.997–1.064 | [True]→[True] |
