评价此页

★ ★ ★ ★ ★

recipes/torch_export_aoti_python

在 Google Colab 中运行

注意

跳转至页面底部下载完整示例代码。

`torch.export` 针对 Python 运行时的 AOTInductor 教程 (Beta)#

创建日期：2024 年 8 月 23 日 | 最后更新：2025 年 1 月 24 日 | 最后验证：2024 年 11 月 5 日

作者： Ankith Gunapal, Bin Bao, Angela Yi

警告

torch._inductor.aoti_compile_and_package 和 torch._inductor.aoti_load_package 目前处于 Beta 状态，未来可能会发生破坏向后兼容性的变更。本教程提供了使用这些 API 在 Python 运行时中部署模型的示例。

此前已经展示过如何使用 AOTInductor 对 PyTorch 导出模型进行提前（Ahead-of-Time）编译，从而创建一个可以在非 Python 环境中运行的工件。在本教程中，你将学习一个关于如何在 Python 运行时中使用 AOTInductor 的端到端示例。

内容

先决条件 #

PyTorch 2.6 或更高版本
对 torch.export 和 AOTInductor 的基本了解
完成 AOTInductor：Torch.Export 导出模型的提前编译教程

你将学到什么 #

如何在 Python 运行时中使用 AOTInductor。
如何结合使用 torch._inductor.aoti_compile_and_package() 和 torch.export.export() 来生成编译后的工件。
如何使用 torch._export.aot_load() 在 Python 运行时中加载并运行该工件。
何时在 Python 运行时中使用 AOTInductor

模型编译 #

我们将以 TorchVision 预训练的 ResNet18 模型为例。

第一步是使用 torch.export.export() 将模型导出为图表示。要了解有关此函数的更多信息，可以查看文档或教程。

一旦导出了 PyTorch 模型并获得了 ExportedProgram，我们就可以应用 torch._inductor.aoti_compile_and_package()，让 AOTInductor 将程序编译到指定设备，并将生成的内容保存为 “.pt2” 工件。

注意

此 API 支持与 torch.compile() 相同的可用选项，例如 mode 和 max_autotune（适用于想要启用 CUDA 图并利用基于 Triton 的矩阵乘法和卷积的用户）。

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]
 72%|███████▏  | 32.4M/44.7M [00:00<00:00, 339MB/s]
100%|██████████| 44.7M/44.7M [00:00<00:00, 336MB/s]
/usr/lib/python3.10/copyreg.py:101: FutureWarning: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
  return cls.__new__(cls, *args)
/usr/local/lib/python3.10/dist-packages/torch/_inductor/compile_fx.py:320: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
  warnings.warn(
/usr/local/lib/python3.10/dist-packages/torch/_inductor/select_algorithm.py:4628: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
  current_out_size = out_base.storage().size()
Autotune Choices Stats:
{"num_choices": 10, "num_triton_choices": 9, "best_kernel": "convolution", "best_time": 0.09830400347709656, "best_triton_pos": 1, "best_triton_time": 0.37785598635673523, "best_triton_kernel": "triton_convolution2d_8", "best_triton_kernel_desc": "ALLOW_TF32=False, 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=4, num_warps=4"}
AUTOTUNE convolution(2x3x224x224, 64x3x7x7)
strides: [150528, 1, 672, 3], [147, 1, 21, 3]
dtypes: torch.float32, torch.float32
  convolution 0.0983 ms 100.0%
  triton_convolution2d_8 0.3779 ms 26.0% ALLOW_TF32=False, 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=4, num_warps=4
  triton_convolution2d_7 0.4014 ms 24.5% ALLOW_TF32=False, 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=4, num_warps=4
  triton_convolution2d_0 0.4250 ms 23.1% ALLOW_TF32=False, 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_5 0.4516 ms 21.8% ALLOW_TF32=False, 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_6 0.4946 ms 19.9% ALLOW_TF32=False, 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=3, num_warps=8
  triton_convolution2d_3 0.5038 ms 19.5% ALLOW_TF32=False, 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_2 0.5407 ms 18.2% ALLOW_TF32=False, 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 0.6410 ms 15.3% ALLOW_TF32=False, 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
  triton_convolution2d_4 0.7035 ms 14.0% ALLOW_TF32=False, 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
SingleProcess AUTOTUNE benchmarking takes 0.4810 seconds and 0.0193 seconds precompiling for 10 choices
Autotune Choices Stats:
{"num_choices": 12, "num_triton_choices": 11, "best_kernel": "convolution", "best_time": 0.09417600184679031, "best_triton_pos": 1, "best_triton_time": 0.11366400122642517, "best_triton_kernel": "triton_convolution2d_13", "best_triton_kernel_desc": "ALLOW_TF32=False, 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"}
AUTOTUNE convolution(2x64x56x56, 64x64x3x3)
strides: [200704, 1, 3584, 64], [576, 1, 192, 64]
dtypes: torch.float32, torch.float32
  convolution 0.0942 ms 100.0%
  triton_convolution2d_13 0.1137 ms 82.9% ALLOW_TF32=False, 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_16 0.1178 ms 80.0% ALLOW_TF32=False, BLOCK_K=64, 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=3, num_warps=8
  triton_convolution2d_12 0.1208 ms 77.9% ALLOW_TF32=False, 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_20 0.1229 ms 76.6% ALLOW_TF32=False, BLOCK_K=64, 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_9 0.1280 ms 73.6% ALLOW_TF32=False, 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_17 0.1280 ms 73.6% ALLOW_TF32=False, BLOCK_K=64, 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=4, num_warps=4
  triton_convolution2d_15 0.1423 ms 66.2% ALLOW_TF32=False, 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_10 0.1495 ms 63.0% ALLOW_TF32=False, 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_19 0.1526 ms 61.7% ALLOW_TF32=False, BLOCK_K=64, 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
SingleProcess AUTOTUNE benchmarking takes 0.3546 seconds and 0.0162 seconds precompiling for 12 choices
Autotune Choices Stats:
{"num_choices": 13, "num_triton_choices": 12, "best_kernel": "convolution", "best_time": 0.09728000313043594, "best_triton_pos": 1, "best_triton_time": 0.11366400122642517, "best_triton_kernel": "triton_convolution2d_25", "best_triton_kernel_desc": "ALLOW_TF32=False, 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"}
AUTOTUNE convolution(2x64x56x56, 64x64x3x3)
strides: [200704, 1, 3584, 64], [576, 1, 192, 64]
dtypes: torch.float32, torch.float32
  convolution 0.0973 ms 100.0%
  triton_convolution2d_25 0.1137 ms 85.6% ALLOW_TF32=False, 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_28 0.1178 ms 82.6% ALLOW_TF32=False, BLOCK_K=64, 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=3, num_warps=8
  triton_convolution2d_24 0.1198 ms 81.2% ALLOW_TF32=False, 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_32 0.1229 ms 79.2% ALLOW_TF32=False, BLOCK_K=64, 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_21 0.1280 ms 76.0% ALLOW_TF32=False, 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_29 0.1280 ms 76.0% ALLOW_TF32=False, BLOCK_K=64, 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=4, num_warps=4
  triton_convolution2d_27 0.1423 ms 68.3% ALLOW_TF32=False, 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_22 0.1495 ms 65.1% ALLOW_TF32=False, 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_31 0.1526 ms 63.8% ALLOW_TF32=False, BLOCK_K=64, 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
SingleProcess AUTOTUNE benchmarking takes 0.2759 seconds and 0.0001 seconds precompiling for 13 choices
Autotune Choices Stats:
{"num_choices": 13, "num_triton_choices": 12, "best_kernel": "triton_convolution2d_61", "best_kernel_desc": "ALLOW_TF32=False, 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", "best_time": 0.07065600156784058, "best_triton_pos": 0}
AUTOTUNE convolution(2x64x56x56, 128x64x3x3)
strides: [200704, 1, 3584, 64], [576, 1, 192, 64]
dtypes: torch.float32, torch.float32
  triton_convolution2d_61 0.0707 ms 100.0% ALLOW_TF32=False, 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
  convolution 0.0942 ms 75.0%
  triton_convolution2d_57 0.0963 ms 73.4% ALLOW_TF32=False, 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_62 0.1198 ms 59.0% ALLOW_TF32=False, 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
  triton_convolution2d_63 0.1413 ms 50.0% ALLOW_TF32=False, 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_68 0.1413 ms 50.0% ALLOW_TF32=False, BLOCK_K=64, 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_58 0.1423 ms 49.6% ALLOW_TF32=False, 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_60 0.1454 ms 48.6% ALLOW_TF32=False, 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_59 0.2048 ms 34.5% ALLOW_TF32=False, 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
  triton_convolution2d_67 0.4065 ms 17.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, 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
SingleProcess AUTOTUNE benchmarking takes 34.2879 seconds and 2.1925 seconds precompiling for 13 choices
Autotune Choices Stats:
{"num_choices": 17, "num_triton_choices": 16, "best_kernel": "convolution", "best_time": 0.09318400174379349, "best_triton_pos": 1, "best_triton_time": 0.13516800105571747, "best_triton_kernel": "triton_convolution2d_73", "best_triton_kernel_desc": "ALLOW_TF32=False, 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"}
AUTOTUNE convolution(2x128x28x28, 128x128x3x3)
strides: [100352, 1, 3584, 128], [1152, 1, 384, 128]
dtypes: torch.float32, torch.float32
  convolution 0.0932 ms 100.0%
  triton_convolution2d_73 0.1352 ms 68.9% ALLOW_TF32=False, 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_69 0.1864 ms 50.0% ALLOW_TF32=False, 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_74 0.2447 ms 38.1% ALLOW_TF32=False, 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_72 0.2693 ms 34.6% ALLOW_TF32=False, 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_75 0.2734 ms 34.1% ALLOW_TF32=False, 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_80 0.2765 ms 33.7% ALLOW_TF32=False, BLOCK_K=64, 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_70 0.2857 ms 32.6% ALLOW_TF32=False, 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_71 0.4219 ms 22.1% ALLOW_TF32=False, 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
  triton_convolution2d_79 0.7926 ms 11.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, 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
SingleProcess AUTOTUNE benchmarking takes 53.7784 seconds and 0.0003 seconds precompiling for 17 choices
Autotune Choices Stats:
{"num_choices": 13, "num_triton_choices": 12, "best_kernel": "triton_convolution2d_85", "best_kernel_desc": "ALLOW_TF32=False, 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", "best_time": 0.027648000046610832, "best_triton_pos": 0}
AUTOTUNE convolution(2x64x56x56, 128x64x1x1)
strides: [200704, 1, 3584, 64], [64, 1, 1, 1]
dtypes: torch.float32, torch.float32
  triton_convolution2d_85 0.0276 ms 100.0% ALLOW_TF32=False, 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_89 0.0276 ms 100.0% ALLOW_TF32=False, 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_90 0.0287 ms 96.4% ALLOW_TF32=False, 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_88 0.0317 ms 87.1% ALLOW_TF32=False, 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_91 0.0317 ms 87.1% ALLOW_TF32=False, 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 0.0338 ms 81.8% ALLOW_TF32=False, 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_93 0.0369 ms 75.0% ALLOW_TF32=False, BLOCK_K=64, 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=4, num_warps=4
  triton_convolution2d_94 0.0369 ms 75.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, 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=4, num_warps=4
  triton_convolution2d_87 0.0399 ms 69.2% ALLOW_TF32=False, 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
  triton_convolution2d_92 0.0410 ms 67.5% ALLOW_TF32=False, BLOCK_K=64, 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=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 16.7395 seconds and 0.0002 seconds precompiling for 13 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.08089599758386612, "best_triton_pos": 1, "best_triton_time": 0.13414399325847626, "best_triton_kernel": "triton_convolution2d_133", "best_triton_kernel_desc": "ALLOW_TF32=False, 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"}
AUTOTUNE convolution(2x128x28x28, 256x128x3x3)
strides: [100352, 1, 3584, 128], [1152, 1, 384, 128]
dtypes: torch.float32, torch.float32
  convolution 0.0809 ms 100.0%
  triton_convolution2d_133 0.1341 ms 60.3% ALLOW_TF32=False, 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_129 0.2724 ms 29.7% ALLOW_TF32=False, 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_135 0.2744 ms 29.5% ALLOW_TF32=False, 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_130 0.2765 ms 29.3% ALLOW_TF32=False, 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_132 0.2826 ms 28.6% ALLOW_TF32=False, 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_134 0.2877 ms 28.1% ALLOW_TF32=False, 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_131 0.3656 ms 22.1% ALLOW_TF32=False, 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
  triton_convolution2d_140 0.7700 ms 10.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, 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_139 0.7844 ms 10.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, 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
SingleProcess AUTOTUNE benchmarking takes 67.3476 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.08396799862384796, "best_triton_pos": 1, "best_triton_time": 0.26214399933815, "best_triton_kernel": "triton_convolution2d_150", "best_triton_kernel_desc": "ALLOW_TF32=False, 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"}
AUTOTUNE convolution(2x256x14x14, 256x256x3x3)
strides: [50176, 1, 3584, 256], [2304, 1, 768, 256]
dtypes: torch.float32, torch.float32
  convolution 0.0840 ms 100.0%
  triton_convolution2d_150 0.2621 ms 32.0% ALLOW_TF32=False, 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_148 0.4813 ms 17.4% ALLOW_TF32=False, 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
  triton_convolution2d_146 0.5294 ms 15.9% ALLOW_TF32=False, 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_149 0.5315 ms 15.8% ALLOW_TF32=False, 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_152 0.5356 ms 15.7% ALLOW_TF32=False, 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_151 0.5530 ms 15.2% ALLOW_TF32=False, 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_147 0.5663 ms 14.8% ALLOW_TF32=False, 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_157 1.5043 ms 5.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, 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_156 1.5534 ms 5.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, 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
SingleProcess AUTOTUNE benchmarking takes 67.3523 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_167", "best_kernel_desc": "ALLOW_TF32=False, 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", "best_time": 0.02969600073993206, "best_triton_pos": 0}
AUTOTUNE convolution(2x128x28x28, 256x128x1x1)
strides: [100352, 1, 3584, 128], [128, 1, 1, 1]
dtypes: torch.float32, torch.float32
  triton_convolution2d_167 0.0297 ms 100.0% ALLOW_TF32=False, 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_164 0.0399 ms 74.4% ALLOW_TF32=False, 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_168 0.0410 ms 72.5% ALLOW_TF32=False, 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_163 0.0430 ms 69.0% ALLOW_TF32=False, 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_166 0.0430 ms 69.0% ALLOW_TF32=False, 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_169 0.0430 ms 69.0% ALLOW_TF32=False, 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_165 0.0461 ms 64.4% ALLOW_TF32=False, 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
  convolution 0.0594 ms 50.0%
  triton_convolution2d_174 0.1085 ms 27.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, 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_173 0.1106 ms 26.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, 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
SingleProcess AUTOTUNE benchmarking takes 83.8629 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.08806400001049042, "best_triton_pos": 1, "best_triton_time": 0.26419198513031006, "best_triton_kernel": "triton_convolution2d_218", "best_triton_kernel_desc": "ALLOW_TF32=False, 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"}
AUTOTUNE convolution(2x256x14x14, 512x256x3x3)
strides: [50176, 1, 3584, 256], [2304, 1, 768, 256]
dtypes: torch.float32, torch.float32
  convolution 0.0881 ms 100.0%
  triton_convolution2d_218 0.2642 ms 33.3% ALLOW_TF32=False, 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_216 0.4710 ms 18.7% ALLOW_TF32=False, 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_214 0.5396 ms 16.3% ALLOW_TF32=False, 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_220 0.5560 ms 15.8% ALLOW_TF32=False, 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_217 0.5581 ms 15.8% ALLOW_TF32=False, 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_219 0.5693 ms 15.5% ALLOW_TF32=False, 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_215 0.5898 ms 14.9% ALLOW_TF32=False, 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_224 1.5012 ms 5.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, 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_225 1.5032 ms 5.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, 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
SingleProcess AUTOTUNE benchmarking takes 66.8366 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 17, "num_triton_choices": 16, "best_kernel": "convolution", "best_time": 0.09932799637317657, "best_triton_pos": 1, "best_triton_time": 0.5222399830818176, "best_triton_kernel": "triton_convolution2d_235", "best_triton_kernel_desc": "ALLOW_TF32=False, 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"}
AUTOTUNE convolution(2x512x7x7, 512x512x3x3)
strides: [25088, 1, 3584, 512], [4608, 1, 1536, 512]
dtypes: torch.float32, torch.float32
  convolution 0.0993 ms 100.0%
  triton_convolution2d_235 0.5222 ms 19.0% ALLOW_TF32=False, 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_233 0.6164 ms 16.1% ALLOW_TF32=False, 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_232 0.7506 ms 13.2% ALLOW_TF32=False, 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_237 0.9124 ms 10.9% ALLOW_TF32=False, 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_231 1.0639 ms 9.3% ALLOW_TF32=False, 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_234 1.0660 ms 9.3% ALLOW_TF32=False, 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_241 1.0762 ms 9.2% ALLOW_TF32=False, BLOCK_K=64, 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_236 1.1049 ms 9.0% ALLOW_TF32=False, 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_242 2.9839 ms 3.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, 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
SingleProcess AUTOTUNE benchmarking takes 48.6255 seconds and 0.0002 seconds precompiling for 17 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_251", "best_kernel_desc": "ALLOW_TF32=False, 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", "best_time": 0.043007999658584595, "best_triton_pos": 0}
AUTOTUNE convolution(2x256x14x14, 512x256x1x1)
strides: [50176, 1, 3584, 256], [256, 1, 1, 1]
dtypes: torch.float32, torch.float32
  triton_convolution2d_251 0.0430 ms 100.0% ALLOW_TF32=False, 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_249 0.0532 ms 80.8% ALLOW_TF32=False, 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
  triton_convolution2d_250 0.0645 ms 66.7% ALLOW_TF32=False, 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_253 0.0645 ms 66.7% ALLOW_TF32=False, 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_247 0.0666 ms 64.6% ALLOW_TF32=False, 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_252 0.0676 ms 63.6% ALLOW_TF32=False, 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_248 0.0686 ms 62.7% ALLOW_TF32=False, 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
  convolution 0.0748 ms 57.5%
  triton_convolution2d_257 0.1925 ms 22.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, 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_258 0.1925 ms 22.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, 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
SingleProcess AUTOTUNE benchmarking takes 65.7551 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_299", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=5, num_warps=2", "best_time": 0.027648000046610832, "best_triton_pos": 0}
AUTOTUNE addmm(2x1000, 2x512, 512x1000)
strides: [0, 1], [512, 1], [1, 512]
dtypes: torch.float32, torch.float32, torch.float32
  triton_mm_299 0.0276 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
  triton_mm_300 0.0307 ms 90.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
  triton_mm_297 0.0328 ms 84.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
  addmm 0.0410 ms 67.5%
  triton_mm_310 0.0420 ms 65.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
  triton_mm_309 0.0429 ms 64.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
  triton_mm_296 0.0430 ms 64.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=1, num_warps=2
  triton_mm_298 0.0430 ms 64.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
  triton_mm_303 0.0430 ms 64.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
  triton_mm_302 0.0440 ms 62.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float32', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 9.7501 seconds and 0.0002 seconds precompiling for 19 choices

aoti_compile_and_package() 的结果是一个名为“resnet18.pt2”的工件，它可以在 Python 和 C++ 中加载并执行。

该工件本身包含大量由 AOTInductor 生成的代码，例如生成的 C++ 运行程序文件、从 C++ 文件编译而来的共享库，以及如果针对 CUDA 进行优化则包含的 CUDA 二进制文件（即 cubin 文件）。

从结构上讲，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 中加载并运行该工件，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)

何时在 Python 运行时中使用 AOTInductor #

在 Python 运行时中使用 AOTInductor 主要有两个原因：

torch._inductor.aoti_compile_and_package 会生成一个单一的序列化工件。这对于模型部署的版本控制以及跟踪模型性能随时间的变化非常有用。
由于 torch.compile() 是一个 JIT 编译器，首次编译存在预热成本。你的部署需要考虑首次推理所需的编译时间。使用 AOTInductor，编译通过 torch.export.export 和 torch._inductor.aoti_compile_and_package 提前完成。在部署时，加载模型后，运行推理不会产生任何额外的额外开销。

以下部分展示了使用 AOTInductor 在首次推理时实现的加速效果。

我们定义了一个实用函数 timed 来测量推理所花费的时间。

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 3.75 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 4372.93 ms

我们发现，与 torch.compile 相比，使用 AOTInductor 的首次推理时间有了显著的提升。

结论 #

在本配方中，我们学习了如何通过编译和加载预训练的 ResNet18 模型来有效地在 Python 运行时中使用 AOTInductor。此过程展示了生成编译工件并在 Python 环境中运行它的实际应用。我们还探讨了在模型部署中使用 AOTInductor 在缩短首次推理时间方面的优势。

脚本总运行时间： (9 分钟 8.232 秒)

torch.export 针对 Python 运行时的 AOTInductor 教程 (Beta)#

先决条件#

你将学到什么#

模型编译#

Python 中的模型推理#

何时在 Python 运行时中使用 AOTInductor#

结论#

文档

教程

资源