torch.export AOTInductor Tutorial for Python runtime (Beta)">

recipes/torch_export_aoti_python

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

torch.export AOTInductor Tutorial for Python runtime (Beta)

Created On: Aug 23, 2024 | Last Updated: Jan 24, 2025 | Last Verified: Nov 05, 2024

Author: Ankith Gunapal, Bin Bao, Angela Yi

Warning

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 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.aoti_compile_and_package() along with torch.export.export() to generate a compiled artifact

• How 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.

Note

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
    )

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]
 98%|█████████▊| 43.9M/44.7M [00:00<00:00, 459MB/s]
100%|██████████| 44.7M/44.7M [00:00<00:00, 459MB/s]
W1106 20:38:53.497000 96488 torch/_inductor/utils.py:2143] Not enough SMs to use max_autotune_gemm mode

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 using torch.export.export and torch._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 3.58 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")

/usr/local/lib/python3.12/dist-packages/triton/knobs.py:504: UserWarning: TRITON_DEFAULT_FP_FUSION is unset: defaulting to FP fusion ENABLED for backward compatibility in Triton 3.9. The default will change to DISABLED in Triton 3.10; set TRITON_DEFAULT_FP_FUSION=1 to keep fusion enabled or =0 to adopt the new default early and silence this warning.
  warn_once(FP_FUSION_DEFAULT_WARNING)
Time taken for first inference for torch.compile is 4332.70 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.