torch.export AOTInductor Tutorial for Python runtime (Beta)

Created On: Aug 23, 2024 | Last Updated: Sep 05, 2024 | Last Verified: Nov 05, 2024

Author: Ankith Gunapal, Bin Bao, Angela Yi

Warning

torch._inductor.aot_compile and torch._export.aot_load are in Beta status and are subject to backwards compatibility breaking changes. This tutorial provides an example of how to use these APIs for model deployment using Python runtime.

It has been shown previously how AOTInductor can be used to do Ahead-of-Time compilation of PyTorch exported models by creating a shared library that can be run in a non-Python environment.

In this tutorial, you will learn an end-to-end example of how to use AOTInductor for Python runtime. We will look at how to use torch._inductor.aot_compile() along with torch.export.export() to generate a shared library. Additionally, we will examine how to execute the shared library in Python runtime using torch._export.aot_load(). You will learn about the speed up seen in the first inference time using AOTInductor, especially when using max-autotune mode which can take some time to execute.

Contents

Prerequisites

• PyTorch 2.4 or later

• Basic understanding of torch.export and AOTInductor

• Complete the AOTInductor: Ahead-Of-Time Compilation for Torch.Export-ed Models tutorial

What you will learn

• How to use AOTInductor for python runtime.

• How to use torch._inductor.aot_compile() along with torch.export.export() to generate a shared library

• How to run a shared library in Python runtime using torch._export.aot_load().

• When do you use AOTInductor for python runtime

Model Compilation

We will use the TorchVision pretrained ResNet18 model and TorchInductor on the exported PyTorch program using torch._inductor.aot_compile().

Note

This API also supports torch.compile() options like mode This means that if used on a CUDA enabled device, you can, for example, set "max_autotune": True which leverages Triton based matrix multiplications & convolutions, and enables CUDA graphs by default.

We also specify dynamic_shapes for the batch dimension. In this example, min=2 is not a bug and is explained in The 0/1 Specialization Problem

import os
import torch
from torchvision.models import ResNet18_Weights, resnet18

model = resnet18(weights=ResNet18_Weights.DEFAULT)
model.eval()

with torch.inference_mode():

    # Specify the generated shared library path
    aot_compile_options = {
            "aot_inductor.output_path": os.path.join(os.getcwd(), "resnet18_pt2.so"),
    }
    if torch.cuda.is_available():
        device = "cuda"
        aot_compile_options.update({"max_autotune": True})
    else:
        device = "cpu"

    model = model.to(device=device)
    example_inputs = (torch.randn(2, 3, 224, 224, device=device),)

    # min=2 is not a bug and is explained in the 0/1 Specialization Problem
    batch_dim = torch.export.Dim("batch", min=2, max=32)
    exported_program = torch.export.export(
        model,
        example_inputs,
        # Specify the first dimension of the input x as dynamic
        dynamic_shapes={"x": {0: batch_dim}},
    )
    so_path = torch._inductor.aot_compile(
        exported_program.module(),
        example_inputs,
        # Specify the generated shared library path
        options=aot_compile_options
    )

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]

 82%|########2 | 36.6M/44.7M [00:00<00:00, 383MB/s]
100%|##########| 44.7M/44.7M [00:00<00:00, 350MB/s]
AUTOTUNE convolution(2x3x224x224, 64x3x7x7)
  convolution 0.0460 ms 100.0%
  triton_convolution2d_0 0.1038 ms 44.3% 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_4 0.1065 ms 43.2% 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=8
  triton_convolution2d_3 0.1275 ms 36.1% 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_1 0.1404 ms 32.8% 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
  triton_convolution2d_5 0.1868 ms 24.6% 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 0.2194 ms 21.0% 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
SingleProcess AUTOTUNE benchmarking takes 0.7992 seconds and 0.0055 seconds precompiling
AUTOTUNE convolution(2x64x56x56, 64x64x3x3)
  convolution 0.0448 ms 100.0%
  triton_convolution2d_9 0.0747 ms 60.0% 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_6 0.0749 ms 59.9% 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_12 0.0774 ms 57.9% 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_10 0.0833 ms 53.8% 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_11 0.0844 ms 53.1% 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_7 0.1007 ms 44.5% 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 0.1422 ms 31.5% 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 0.9540 seconds and 0.0005 seconds precompiling
AUTOTUNE convolution(2x64x56x56, 128x64x3x3)
  convolution 0.0344 ms 100.0%
  triton_convolution2d_38 0.0633 ms 54.3% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
  triton_convolution2d_40 0.0818 ms 42.0% 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_34 0.0858 ms 40.0% 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_39 0.0909 ms 37.8% 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
  triton_convolution2d_37 0.1070 ms 32.1% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
  triton_convolution2d_35 0.1100 ms 31.3% 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 0.3052 ms 11.3% 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 0.9626 seconds and 0.0006 seconds precompiling
AUTOTUNE convolution(2x64x56x56, 128x64x1x1)
  triton_convolution2d_52 0.0111 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 0.0124 ms 89.7% 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 0.0128 ms 86.8% 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
  convolution 0.0132 ms 84.0%
  triton_convolution2d_54 0.0149 ms 74.6% 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 0.0155 ms 71.5% 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_51 0.0156 ms 71.0% 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_50 0.0457 ms 24.3% 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 0.9591 seconds and 0.0006 seconds precompiling
AUTOTUNE convolution(2x128x28x28, 128x128x3x3)
  convolution 0.0435 ms 100.0%
  triton_convolution2d_59 0.1174 ms 37.0% 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_61 0.1356 ms 32.0% 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_55 0.1665 ms 26.1% 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_60 0.1766 ms 24.6% 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_56 0.1901 ms 22.9% 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_58 0.1950 ms 22.3% 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_57 0.2653 ms 16.4% 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 0.9521 seconds and 0.0005 seconds precompiling
AUTOTUNE convolution(2x128x28x28, 256x128x3x3)
  convolution 0.0370 ms 100.0%
  triton_convolution2d_73 0.0999 ms 37.0% 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_75 0.1591 ms 23.3% 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_72 0.2034 ms 18.2% 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_70 0.2149 ms 17.2% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
  triton_convolution2d_71 0.2653 ms 14.0% 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
  triton_convolution2d_74 0.2844 ms 13.0% 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_69 0.3381 ms 11.0% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.9649 seconds and 0.0006 seconds precompiling
AUTOTUNE convolution(2x128x28x28, 256x128x1x1)
  triton_convolution2d_87 0.0119 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
  convolution 0.0202 ms 58.9%
  triton_convolution2d_88 0.0212 ms 56.2% 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_86 0.0214 ms 55.5% 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_89 0.0221 ms 53.8% 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_85 0.0330 ms 36.1% 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
  triton_convolution2d_83 0.0515 ms 23.1% 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_84 0.0644 ms 18.5% 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 0.9965 seconds and 0.0005 seconds precompiling
AUTOTUNE convolution(2x256x14x14, 256x256x3x3)
  convolution 0.0537 ms 100.0%
  triton_convolution2d_94 0.1852 ms 29.0% 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_92 0.2602 ms 20.6% 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
  triton_convolution2d_96 0.2625 ms 20.5% 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_91 0.3700 ms 14.5% 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_93 0.3734 ms 14.4% 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_95 0.5472 ms 9.8% 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_90 0.6530 ms 8.2% 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=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.9454 seconds and 0.0005 seconds precompiling
AUTOTUNE convolution(2x256x14x14, 512x256x3x3)
  convolution 0.0531 ms 100.0%
  triton_convolution2d_108 0.1923 ms 27.6% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
  triton_convolution2d_106 0.2812 ms 18.9% 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_110 0.2928 ms 18.1% 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_105 0.3824 ms 13.9% 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_107 0.3890 ms 13.6% 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_109 0.5597 ms 9.5% 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_104 0.6840 ms 7.8% 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
SingleProcess AUTOTUNE benchmarking takes 0.9501 seconds and 0.0005 seconds precompiling
AUTOTUNE convolution(2x256x14x14, 512x256x1x1)
  triton_convolution2d_122 0.0179 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
  convolution 0.0257 ms 69.6%
  triton_convolution2d_120 0.0330 ms 54.2% 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
  triton_convolution2d_124 0.0339 ms 52.7% 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_123 0.0350 ms 51.1% 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_121 0.0650 ms 27.5% 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 0.1953 ms 9.1% 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_119 0.2020 ms 8.8% 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
SingleProcess AUTOTUNE benchmarking takes 1.0082 seconds and 0.0004 seconds precompiling
AUTOTUNE convolution(2x512x7x7, 512x512x3x3)
  convolution 0.0855 ms 100.0%
  triton_convolution2d_127 0.2792 ms 30.6% 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_129 0.3600 ms 23.8% 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_131 0.4235 ms 20.2% 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_126 0.4841 ms 17.7% 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_128 0.7234 ms 11.8% 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_130 1.1000 ms 7.8% 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_125 1.4550 ms 5.9% 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 0.9574 seconds and 0.0005 seconds precompiling
AUTOTUNE addmm(2x1000, 2x512, 512x1000)
  addmm 0.0156 ms 100.0%
  triton_mm_142 0.0215 ms 72.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, 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_153 0.0303 ms 51.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, 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_152 0.0304 ms 51.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, 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_141 0.0305 ms 51.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, 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_146 0.0308 ms 50.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, 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_139 0.0344 ms 45.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=16, BLOCK_N=32, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=1, num_warps=2
  triton_mm_145 0.0374 ms 41.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=2, num_warps=4
  triton_mm_144 0.0456 ms 34.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=16, BLOCK_N=64, B_PROLOGUE_CAST_TYPE=None, EVEN_K=True, GROUP_M=8, num_stages=2, num_warps=4
  triton_mm_148 0.0500 ms 31.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, 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 1.8155 seconds and 0.0021 seconds precompiling

Model Inference in Python

Typically, the shared object generated above is used in a non-Python environment. In PyTorch 2.3, we added a new API called torch._export.aot_load() to load the shared library in the Python runtime. The API follows a structure similar to the torch.jit.load() API . You need to specify the path of the shared library and the device where it should be loaded.

Note

In the example above, we specified batch_size=1 for inference and it still functions correctly even though we specified min=2 in torch.export.export().

import os
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
model_so_path = os.path.join(os.getcwd(), "resnet18_pt2.so")

model = torch._export.aot_load(model_so_path, device)
example_inputs = (torch.randn(1, 3, 224, 224, device=device),)

with torch.inference_mode():
    output = model(example_inputs)

When to use AOTInductor for Python Runtime

One of the requirements for using AOTInductor is that the model shouldn’t have any graph breaks. Once this requirement is met, the primary use case for using AOTInductor Python Runtime is for model deployment using Python. There are mainly two reasons why you would use AOTInductor Python Runtime:

• torch._inductor.aot_compile generates a shared library. This is useful for model versioning for deployments and tracking model performance over time.

• With torch.compile() being a JIT compiler, there is a warmup cost associated with the first compilation. Your deployment needs to account for the compilation time taken for the first inference. With AOTInductor, the compilation is done offline using torch.export.export & torch._indutor.aot_compile. The deployment would only load the shared library using torch._export.aot_load and run inference.

The section below shows the speedup achieved with AOTInductor for first inference

We define a utility function timed to measure the time taken for inference

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

Lets measure the time for first inference using AOTInductor

torch._dynamo.reset()

model = torch._export.aot_load(model_so_path, device)
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 2.90 ms

Lets measure the time for first inference using torch.compile

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 6908.31 ms

We see that there is a drastic speedup in first inference time using AOTInductor compared to torch.compile

Conclusion

In this recipe, we have learned how to effectively use the AOTInductor for Python runtime by compiling and loading a pretrained ResNet18 model using the torch._inductor.aot_compile and torch._export.aot_load APIs. This process demonstrates the practical application of generating a shared library and running it within a Python environment, even with dynamic shape considerations and device-specific optimizations. We also looked at the advantage of using AOTInductor in model deployments, with regards to speed up in first inference time.