注意
单击此处下载完整的示例代码
torch.export AOTInductor Python 运行时教程 (Beta)¶
创建于：2024 年 8 月 23 日 | 最后更新：2025 年 1 月 24 日 | 最后验证：2024 年 11 月 05 日
作者： Ankith Gunapal, Bin Bao, Angela Yi
警告
torch._inductor.aoti_compile_and_package 和 torch._inductor.aoti_load_package 处于 Beta 状态，可能会发生向后不兼容的更改。本教程提供了一个如何使用这些 API 进行模型部署的 Python 运行时示例。
先前已展示如何使用 AOTInductor 通过创建可在非 Python 环境中运行的工件来对 PyTorch 导出的模型进行提前编译。在本教程中，您将学习一个端到端的示例，了解如何将 AOTInductor 用于 Python 运行时。
目录
您将学到什么¶
如何将 AOTInductor 用于 Python 运行时。
如何使用
torch._inductor.aoti_compile_and_package()以及torch.export.export()生成编译后的工件如何在 Python 运行时中使用
torch._export.aot_load()加载和运行工件。何时将 AOTInductor 与 Python 运行时一起使用
模型编译¶
我们将使用 TorchVision 预训练的 ResNet18 模型作为示例。
第一步是使用 torch.export.export() 将模型导出到图形表示。要了解有关使用此函数的更多信息，您可以查看 文档 或 教程。
一旦我们导出了 PyTorch 模型并获得了 ExportedProgram，我们就可以应用 torch._inductor.aoti_compile_and_package() 到 AOTInductor，以将程序编译到指定的设备，并将生成的内容保存到 “.pt2” 工件中。
注意
此 API 支持与 torch.compile() 相同的可用选项，例如 mode 和 max_autotune（对于那些想要启用 CUDA 图并利用基于 Triton 的矩阵乘法和卷积的人）
import os
import torch
import torch._inductor
from torchvision.models import ResNet18_Weights, resnet18
model = resnet18(weights=ResNet18_Weights.DEFAULT)
model.eval()
with torch.inference_mode():
inductor_configs = {}
if torch.cuda.is_available():
device = "cuda"
inductor_configs["max_autotune"] = True
else:
device = "cpu"
model = model.to(device=device)
example_inputs = (torch.randn(2, 3, 224, 224, device=device),)
exported_program = torch.export.export(
model,
example_inputs,
)
path = torch._inductor.aoti_compile_and_package(
exported_program,
package_path=os.path.join(os.getcwd(), "resnet18.pt2"),
inductor_configs=inductor_configs
)
Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /var/lib/ci-user/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth
0%| | 0.00/44.7M [00:00<?, ?B/s]
77%|#######7 | 34.5M/44.7M [00:00<00:00, 361MB/s]
100%|##########| 44.7M/44.7M [00:00<00:00, 376MB/s]
AUTOTUNE convolution(2x3x224x224, 64x3x7x7)
convolution 0.0530 ms 100.0%
triton_convolution2d_4 0.1401 ms 37.8% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=7, KERNEL_W=7, PADDING_H=3, PADDING_W=3, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_0 0.1503 ms 35.2% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=7, KERNEL_W=7, PADDING_H=3, PADDING_W=3, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_3 0.1831 ms 28.9% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=7, KERNEL_W=7, PADDING_H=3, PADDING_W=3, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5 0.2471 ms 21.4% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=7, KERNEL_W=7, PADDING_H=3, PADDING_W=3, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2 0.5350 ms 9.9% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=7, KERNEL_W=7, PADDING_H=3, PADDING_W=3, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_1 0.8847 ms 6.0% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=7, KERNEL_W=7, PADDING_H=3, PADDING_W=3, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.9344 seconds and 0.0034 seconds precompiling for 7 choices
AUTOTUNE convolution(2x64x56x56, 64x64x3x3)
triton_convolution2d_10 0.0356 ms 100.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6 0.0368 ms 96.8% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_11 0.0373 ms 95.5% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9 0.0415 ms 85.8% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
convolution 0.0460 ms 77.5%
triton_convolution2d_12 0.0651 ms 54.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7 0.0768 ms 46.4% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_8 0.1257 ms 28.4% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9402 seconds and 0.0007 seconds precompiling for 8 choices
AUTOTUNE convolution(2x64x56x56, 128x64x3x3)
triton_convolution2d_38 0.0288 ms 100.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_39 0.0395 ms 72.9% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
convolution 0.0469 ms 61.4%
triton_convolution2d_34 0.0477 ms 60.4% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_37 0.0606 ms 47.5% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_40 0.0622 ms 46.3% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_35 0.0714 ms 40.3% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_36 0.1327 ms 21.7% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9408 seconds and 0.0006 seconds precompiling for 8 choices
AUTOTUNE convolution(2x128x28x28, 128x128x3x3)
convolution 0.0450 ms 100.0%
triton_convolution2d_45 0.0492 ms 91.3% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_46 0.0694 ms 64.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_41 0.0841 ms 53.5% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_44 0.1041 ms 43.2% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_47 0.1172 ms 38.4% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_42 0.1395 ms 32.2% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_43 0.2411 ms 18.6% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9520 seconds and 0.0007 seconds precompiling for 8 choices
AUTOTUNE convolution(2x64x56x56, 128x64x1x1)
triton_convolution2d_52 0.0089 ms 100.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_53 0.0100 ms 88.5% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_48 0.0106 ms 84.0% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_51 0.0133 ms 66.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
convolution 0.0137 ms 65.1%
triton_convolution2d_54 0.0138 ms 64.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_49 0.0144 ms 61.8% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_50 0.0215 ms 41.4% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9232 seconds and 0.0007 seconds precompiling for 8 choices
AUTOTUNE convolution(2x128x28x28, 256x128x3x3)
convolution 0.0380 ms 100.0%
triton_convolution2d_73 0.0487 ms 78.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_74 0.1107 ms 34.3% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_72 0.1130 ms 33.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_75 0.1164 ms 32.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_69 0.1318 ms 28.8% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_70 0.1344 ms 28.3% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_71 0.2048 ms 18.6% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9544 seconds and 0.0006 seconds precompiling for 8 choices
AUTOTUNE convolution(2x256x14x14, 256x256x3x3)
convolution 0.0565 ms 100.0%
triton_convolution2d_80 0.0916 ms 61.6% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_79 0.2094 ms 27.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_81 0.2143 ms 26.4% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_82 0.2275 ms 24.8% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_76 0.2566 ms 22.0% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_77 0.2721 ms 20.8% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_78 0.3744 ms 15.1% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=512, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9747 seconds and 0.0007 seconds precompiling for 8 choices
AUTOTUNE convolution(2x128x28x28, 256x128x1x1)
triton_convolution2d_87 0.0103 ms 100.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_88 0.0189 ms 54.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_89 0.0190 ms 54.5% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_86 0.0192 ms 53.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_84 0.0216 ms 47.8% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_83 0.0217 ms 47.7% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
convolution 0.0250 ms 41.4%
triton_convolution2d_85 0.0263 ms 39.2% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9272 seconds and 0.0006 seconds precompiling for 8 choices
AUTOTUNE convolution(2x256x14x14, 512x256x3x3)
convolution 0.0580 ms 100.0%
triton_convolution2d_108 0.0929 ms 62.4% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_109 0.2173 ms 26.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_107 0.2214 ms 26.2% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_110 0.2252 ms 25.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_106 0.2421 ms 23.9% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=512, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_105 0.2596 ms 22.3% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_104 0.2606 ms 22.2% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.9720 seconds and 0.0006 seconds precompiling for 8 choices
AUTOTUNE convolution(2x512x7x7, 512x512x3x3)
convolution 0.0852 ms 100.0%
triton_convolution2d_115 0.1793 ms 47.5% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_113 0.2146 ms 39.7% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_117 0.2663 ms 32.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_112 0.3213 ms 26.5% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_114 0.4028 ms 21.1% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_116 0.4278 ms 19.9% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_111 0.5135 ms 16.6% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.9917 seconds and 0.0007 seconds precompiling for 8 choices
AUTOTUNE convolution(2x256x14x14, 512x256x1x1)
triton_convolution2d_122 0.0156 ms 100.0% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_120 0.0266 ms 58.5% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=512, BLOCK_N=16, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=1, num_warps=8
convolution 0.0286 ms 54.5%
triton_convolution2d_121 0.0307 ms 50.7% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_118 0.0308 ms 50.5% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_124 0.0310 ms 50.2% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_123 0.0315 ms 49.5% ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_119 0.0334 ms 46.6% ALLOW_TF32=True, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=2, STRIDE_W=2, UNROLL=True, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.9277 seconds and 0.0006 seconds precompiling for 8 choices
AUTOTUNE addmm(2x1000, 2x512, 512x1000)
triton_mm_146 0.0116 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=3, num_warps=4
triton_mm_141 0.0118 ms 98.6% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=5, num_warps=4
triton_mm_153 0.0120 ms 97.1% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=4, num_warps=4
triton_mm_142 0.0120 ms 96.8% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=5, num_warps=2
triton_mm_143 0.0123 ms 94.8% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=5, num_warps=2
triton_mm_152 0.0127 ms 91.7% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=3, num_warps=4
triton_mm_140 0.0132 ms 87.9% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=2, num_warps=2
triton_mm_149 0.0137 ms 84.8% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=16, BLOCK_N=128, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=4, num_warps=8
triton_mm_150 0.0137 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=3, num_warps=4
triton_mm_148 0.0144 ms 80.8% ACC_TYPE='tl.float32', ALLOW_TF32=True, BLOCK_K=32, BLOCK_M=16, BLOCK_N=128, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 1.7563 seconds and 0.0019 seconds precompiling for 18 choices
aoti_compile_and_package() 的结果是一个名为 “resnet18.pt2” 的工件，可以在 Python 和 C++ 中加载和执行。
工件本身包含大量 AOTInductor 生成的代码，例如生成的 C++ 运行器文件、从 C++ 文件编译的共享库以及 CUDA 二进制文件（又名 cubin 文件，如果针对 CUDA 进行优化）。
从结构上看，该工件是一个结构化的 .zip 文件，具有以下规范
我们可以使用以下命令来检查工件内容
$ unzip -l resnet18.pt2
Archive: resnet18.pt2
Length Date Time Name
--------- ---------- ----- ----
1 01-08-2025 16:40 version
3 01-08-2025 16:40 archive_format
10088 01-08-2025 16:40 data/aotinductor/model/cagzt6akdaczvxwtbvqe34otfe5jlorktbqlojbzqjqvbfsjlge4.cubin
17160 01-08-2025 16:40 data/aotinductor/model/c6oytfjmt5w4c7onvtm6fray7clirxt7q5xjbwx3hdydclmwoujz.cubin
16616 01-08-2025 16:40 data/aotinductor/model/c7ydp7nocyz323hij4tmlf2kcedmwlyg6r57gaqzcsy3huneamu6.cubin
17776 01-08-2025 16:40 data/aotinductor/model/cyqdf46ordevqhiddvpdpp3uzwatfbzdpl3auj2nx23uxvplnne2.cubin
10856 01-08-2025 16:40 data/aotinductor/model/cpzfebfgrusqslui7fxsuoo4tvwulmrxirc5tmrpa4mvrbdno7kn.cubin
14608 01-08-2025 16:40 data/aotinductor/model/c5ukeoz5wmaszd7vczdz2qhtt6n7tdbl3b6wuy4rb2se24fjwfoy.cubin
11376 01-08-2025 16:40 data/aotinductor/model/csu3nstcp56tsjfycygaqsewpu64l5s6zavvz7537cm4s4cv2k3r.cubin
10984 01-08-2025 16:40 data/aotinductor/model/cp76lez4glmgq7gedf2u25zvvv6rksv5lav4q22dibd2zicbgwj3.cubin
14736 01-08-2025 16:40 data/aotinductor/model/c2bb5p6tnwz4elgujqelsrp3unvkgsyiv7xqxmpvuxcm4jfl7pc2.cubin
11376 01-08-2025 16:40 data/aotinductor/model/c6eopmb2b4ngodwsayae4r5q6ni3jlfogfbdk3ypg56tgpzhubfy.cubin
11624 01-08-2025 16:40 data/aotinductor/model/chmwe6lvoekzfowdbiizitm3haiiuad5kdm6sd2m6mv6dkn2zk32.cubin
15632 01-08-2025 16:40 data/aotinductor/model/c3jop5g344hj3ztsu4qm6ibxyaaerlhkzh2e6emak23rxfje6jam.cubin
25472 01-08-2025 16:40 data/aotinductor/model/chaiixybeiuuitm2nmqnxzijzwgnn2n7uuss4qmsupgblfh3h5hk.cubin
139389 01-08-2025 16:40 data/aotinductor/model/cvk6qzuybruhwxtfblzxiov3rlrziv5fkqc4mdhbmantfu3lmd6t.cpp
27 01-08-2025 16:40 data/aotinductor/model/cvk6qzuybruhwxtfblzxiov3rlrziv5fkqc4mdhbmantfu3lmd6t_metadata.json
47195424 01-08-2025 16:40 data/aotinductor/model/cvk6qzuybruhwxtfblzxiov3rlrziv5fkqc4mdhbmantfu3lmd6t.so
--------- -------
47523148 18 files
Python 中的模型推理¶
要在 Python 中加载和运行工件，我们可以使用 torch._inductor.aoti_load_package()。
import os
import torch
import torch._inductor
model_path = os.path.join(os.getcwd(), "resnet18.pt2")
compiled_model = torch._inductor.aoti_load_package(model_path)
example_inputs = (torch.randn(2, 3, 224, 224, device=device),)
with torch.inference_mode():
output = compiled_model(example_inputs)
何时将 AOTInductor 与 Python 运行时一起使用¶
人们将 AOTInductor 与 Python 运行时一起使用主要有两个原因
torch._inductor.aoti_compile_and_package生成一个单一的序列化工件。这对于部署的模型版本控制和跟踪模型性能随时间变化非常有用。由于
torch.compile()是 JIT 编译器，因此首次编译存在预热成本。您的部署需要考虑首次推理所需的编译时间。使用 AOTInductor，编译是提前完成的，使用torch.export.export和torch._inductor.aoti_compile_and_package。在部署时，加载模型后，运行推理不会产生任何额外成本。
以下部分显示了 AOTInductor 在首次推理中实现的速度提升
我们定义一个实用函数 timed 来测量推理所用的时间
import time
def timed(fn):
# Returns the result of running `fn()` and the time it took for `fn()` to run,
# in seconds. We use CUDA events and synchronization for accurate
# measurement on CUDA enabled devices.
if torch.cuda.is_available():
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
else:
start = time.time()
result = fn()
if torch.cuda.is_available():
end.record()
torch.cuda.synchronize()
else:
end = time.time()
# Measure time taken to execute the function in miliseconds
if torch.cuda.is_available():
duration = start.elapsed_time(end)
else:
duration = (end - start) * 1000
return result, duration
让我们测量使用 AOTInductor 进行首次推理的时间
torch._dynamo.reset()
model = torch._inductor.aoti_load_package(model_path)
example_inputs = (torch.randn(1, 3, 224, 224, device=device),)
with torch.inference_mode():
_, time_taken = timed(lambda: model(example_inputs))
print(f"Time taken for first inference for AOTInductor is {time_taken:.2f} ms")
Time taken for first inference for AOTInductor is 4.13 ms
让我们测量使用 torch.compile 进行首次推理的时间
torch._dynamo.reset()
model = resnet18(weights=ResNet18_Weights.DEFAULT).to(device)
model.eval()
model = torch.compile(model)
example_inputs = torch.randn(1, 3, 224, 224, device=device)
with torch.inference_mode():
_, time_taken = timed(lambda: model(example_inputs))
print(f"Time taken for first inference for torch.compile is {time_taken:.2f} ms")
Time taken for first inference for torch.compile is 5613.87 ms
我们看到，与 torch.compile 相比，使用 AOTInductor 可以大幅提升首次推理速度
