참고
Go to the end to download the full example code.
torch.export
AOTInductor Tutorial for Python runtime (Beta)#
Author: Ankith Gunapal, Bin Bao, Angela Yi
경고
torch._inductor.aoti_compile_and_package
and
torch._inductor.aoti_load_package
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 an artifact 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.
Contents
Prerequisites#
PyTorch 2.6 or later
Basic understanding of
torch.export
and AOTInductorComplete 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.aoti_compile_and_package()
along withtorch.export.export()
to generate a compiled artifactHow to load and run the artifact in a Python runtime using
torch._export.aot_load()
.When to you use AOTInductor with a Python runtime
Model Compilation#
We will use the TorchVision pretrained ResNet18
model as an example.
The first step is to export the model to a graph representation using
torch.export.export()
. To learn more about using this function, you can
check out the docs or the
tutorial.
Once we have exported the PyTorch model and obtained an ExportedProgram
,
we can apply torch._inductor.aoti_compile_and_package()
to AOTInductor
to compile the program to a specified device, and save the generated contents
into a “.pt2” artifact.
참고
This API supports the same available options that torch.compile()
has, such as mode
and max_autotune
(for those who want to enable
CUDA graphs and leverage Triton based matrix multiplications and
convolutions)
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
)
/opt/conda/lib/python3.11/site-packages/torch/_inductor/compile_fx.py:282: UserWarning:
TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
AUTOTUNE convolution(2x3x224x224, 64x3x7x7)
strides: [150528, 1, 672, 3], [147, 1, 21, 3]
dtypes: torch.float32, torch.float32
triton_convolution2d_4 0.0531 ms 100.0% 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 0.0613 ms 86.7% 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_5 0.0788 ms 67.5% 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_3 0.0865 ms 61.4% 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
convolution 0.0988 ms 53.8%
triton_convolution2d_2 0.1032 ms 51.5% 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 0.1359 ms 39.1% 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 0.7883 seconds and 0.0008 seconds precompiling for 7 choices
AUTOTUNE convolution(2x64x56x56, 64x64x3x3)
strides: [200704, 1, 3584, 64], [576, 1, 192, 64]
dtypes: torch.float32, torch.float32
convolution 0.0109 ms 100.0%
triton_convolution2d_10 0.0189 ms 57.6% 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_11 0.0195 ms 55.8% 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=8
triton_convolution2d_9 0.0235 ms 46.3% 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 0.0286 ms 38.1% 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 0.0295 ms 36.9% 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_7 0.0413 ms 26.3% 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=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_8 0.0928 ms 11.7% 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=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.0850 seconds and 0.0010 seconds precompiling for 8 choices
AUTOTUNE convolution(2x64x56x56, 128x64x3x3)
strides: [200704, 1, 3584, 64], [576, 1, 192, 64]
dtypes: torch.float32, torch.float32
convolution 0.0106 ms 100.0%
triton_convolution2d_38 0.0191 ms 55.3% 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_39 0.0220 ms 48.0% 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 0.0269 ms 39.2% 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_40 0.0281 ms 37.5% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_34 0.0335 ms 31.5% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_35 0.0409 ms 25.8% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_36 0.0918 ms 11.5% 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=2, STRIDE_W=2, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.1004 seconds and 0.0001 seconds precompiling for 8 choices
AUTOTUNE convolution(2x128x28x28, 128x128x3x3)
strides: [100352, 1, 3584, 128], [1152, 1, 384, 128]
dtypes: torch.float32, torch.float32
convolution 0.0121 ms 100.0%
triton_convolution2d_45 0.0330 ms 36.7% 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 0.0387 ms 31.3% 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_44 0.0484 ms 25.0% 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_47 0.0500 ms 24.2% 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_41 0.0614 ms 19.7% 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_42 0.0764 ms 15.8% 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_43 0.1726 ms 7.0% 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=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.1230 seconds and 0.0002 seconds precompiling for 8 choices
AUTOTUNE convolution(2x64x56x56, 128x64x1x1)
strides: [200704, 1, 3584, 64], [64, 1, 1, 1]
dtypes: torch.float32, torch.float32
triton_convolution2d_52 0.0068 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 0.0071 ms 95.5% 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
convolution 0.0083 ms 82.2%
triton_convolution2d_51 0.0083 ms 81.9% 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_48 0.0084 ms 80.7% 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_54 0.0093 ms 73.4% 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 0.0103 ms 66.1% 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 0.0183 ms 37.3% 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 0.0777 seconds and 0.0002 seconds precompiling for 8 choices
AUTOTUNE convolution(2x128x28x28, 256x128x3x3)
strides: [100352, 1, 3584, 128], [1152, 1, 384, 128]
dtypes: torch.float32, torch.float32
convolution 0.0117 ms 100.0%
triton_convolution2d_73 0.0335 ms 34.9% 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 0.0452 ms 25.9% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_72 0.0473 ms 24.8% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_75 0.0492 ms 23.8% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_70 0.0761 ms 15.4% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_69 0.0762 ms 15.4% 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_71 0.1492 ms 7.9% 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 0.1230 seconds and 0.0001 seconds precompiling for 8 choices
AUTOTUNE convolution(2x256x14x14, 256x256x3x3)
strides: [50176, 1, 3584, 256], [2304, 1, 768, 256]
dtypes: torch.float32, torch.float32
convolution 0.0151 ms 100.0%
triton_convolution2d_80 0.0657 ms 23.0% 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_81 0.0830 ms 18.2% 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_79 0.0887 ms 17.0% 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_82 0.0972 ms 15.5% 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_76 0.1458 ms 10.4% 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 0.1468 ms 10.3% 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=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_78 0.1793 ms 8.4% 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 0.1573 seconds and 0.0002 seconds precompiling for 8 choices
AUTOTUNE convolution(2x128x28x28, 256x128x1x1)
strides: [100352, 1, 3584, 128], [128, 1, 1, 1]
dtypes: torch.float32, torch.float32
triton_convolution2d_87 0.0078 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 0.0086 ms 91.1%
triton_convolution2d_88 0.0104 ms 75.6% 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 0.0106 ms 73.8% 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 0.0111 ms 70.8% 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_83 0.0124 ms 63.3% 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 0.0141 ms 55.6% 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
triton_convolution2d_85 0.0209 ms 37.5% 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 0.0802 seconds and 0.0002 seconds precompiling for 8 choices
AUTOTUNE convolution(2x256x14x14, 512x256x3x3)
strides: [50176, 1, 3584, 256], [2304, 1, 768, 256]
dtypes: torch.float32, torch.float32
convolution 0.0148 ms 100.0%
triton_convolution2d_108 0.0658 ms 22.5% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_109 0.0815 ms 18.2% 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 0.0870 ms 17.0% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_110 0.0903 ms 16.4% 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=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_105 0.1384 ms 10.7% 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 0.1407 ms 10.5% 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_106 0.1625 ms 9.1% 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=2, STRIDE_W=2, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.1528 seconds and 0.0001 seconds precompiling for 8 choices
AUTOTUNE convolution(2x512x7x7, 512x512x3x3)
strides: [25088, 1, 3584, 512], [4608, 1, 1536, 512]
dtypes: torch.float32, torch.float32
convolution 0.0190 ms 100.0%
triton_convolution2d_115 0.1307 ms 14.5% 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_117 0.1463 ms 13.0% 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_116 0.1561 ms 12.2% 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_114 0.1742 ms 10.9% 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_113 0.2061 ms 9.2% 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_112 0.2327 ms 8.2% 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_111 0.2752 ms 6.9% 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 0.2082 seconds and 0.0002 seconds precompiling for 8 choices
AUTOTUNE convolution(2x256x14x14, 512x256x1x1)
strides: [50176, 1, 3584, 256], [256, 1, 1, 1]
dtypes: torch.float32, torch.float32
convolution 0.0090 ms 100.0%
triton_convolution2d_122 0.0110 ms 82.2% 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_123 0.0143 ms 62.9% 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 0.0147 ms 61.3% 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_124 0.0152 ms 59.5% 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_118 0.0193 ms 46.8% 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 0.0206 ms 43.9% 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_120 0.0211 ms 42.9% 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
SingleProcess AUTOTUNE benchmarking takes 0.0830 seconds and 0.0002 seconds precompiling for 8 choices
AUTOTUNE addmm(2x1000, 2x512, 512x1000)
strides: [0, 1], [512, 1], [1, 512]
dtypes: torch.float32, torch.float32, torch.float32
addmm 0.0107 ms 100.0%
triton_mm_143 0.0163 ms 65.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_142 0.0175 ms 61.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_140 0.0180 ms 59.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_147 0.0260 ms 41.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_146 0.0260 ms 41.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_141 0.0262 ms 40.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_153 0.0262 ms 40.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=4, num_warps=4
triton_mm_152 0.0264 ms 40.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_139 0.0341 ms 31.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=1, num_warps=2
SingleProcess AUTOTUNE benchmarking takes 0.3771 seconds and 0.0001 seconds precompiling for 18 choices
The result of aoti_compile_and_package()
is an artifact “resnet18.pt2”
which can be loaded and executed in Python and C++.
The artifact itself contains a bunch of AOTInductor generated code, such as a generated C++ runner file, a shared library compiled from the C++ file, and CUDA binary files, aka cubin files, if optimizing for CUDA.
Structure-wise, the artifact is a structured .zip
file, with the following
specification:
We can use the following command to inspect the artifact contents:
$ 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
Model Inference in Python#
To load and run the artifact in Python, we can use 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)
When to use AOTInductor with a Python Runtime#
There are mainly two reasons why one would use AOTInductor with a Python Runtime:
torch._inductor.aoti_compile_and_package
generates a singular serialized artifact. 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 ahead of time usingtorch.export.export
andtorch._inductor.aoti_compile_and_package
. At deployment time, after loading the model, running inference does not have any additional cost.
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._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 1.94 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 2575.98 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. This process
demonstrates the practical application of generating a compiled artifact and
running it within a Python environment. We also looked at the advantage of using
AOTInductor in model deployments, with regards to speed up in first inference time.
Total running time of the script: (0 minutes 33.595 seconds)