Optimize inference using torch.compile()

This guide aims to provide a benchmark on the inference speed-ups introduced with torch.compile() for computer vision models in 🤗 Transformers.

Benefits of torch.compile

Depending on the model and the GPU, torch.compile() yields up to 30% speed-up during inference. To use torch.compile(), simply install any version of torch above 2.0.

Compiling a model takes time, so it’s useful if you are compiling the model only once instead of every time you infer. To compile any computer vision model of your choice, call torch.compile() on the model as shown below:

from transformers import AutoModelForImageClassification

model = AutoModelForImageClassification.from_pretrained(MODEL_ID).to(DEVICE)
+ model = torch.compile(model)

compile() comes with multiple modes for compiling, which essentially differ in compilation time and inference overhead. max-autotune takes longer than reduce-overhead but results in faster inference. Default mode is fastest for compilation but is not as efficient compared to reduce-overhead for inference time. In this guide, we used the default mode. You can learn more about it here.

We benchmarked torch.compile with different computer vision models, tasks, types of hardware, and batch sizes on torch version 2.0.1.

Benchmarking code

Below you can find the benchmarking code for each task. We warm up the GPU before inference and take the mean time of 300 inferences, using the same image each time.

Image Classification with ViT

import torch
from PIL import Image
import requests
import numpy as np
from transformers import AutoImageProcessor, AutoModelForImageClassification
from accelerate.test_utils.testing import get_backend

device, _, _ = get_backend() # automatically detects the underlying device type (CUDA, CPU, XPU, MPS, etc.)
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)

processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
model = AutoModelForImageClassification.from_pretrained("google/vit-base-patch16-224").to(device)
model = torch.compile(model)

processed_input = processor(image, return_tensors='pt').to(device)

with torch.no_grad():
    _ = model(**processed_input)

Object Detection with DETR

from transformers import AutoImageProcessor, AutoModelForObjectDetection
from accelerate.test_utils.testing import get_backend

device, _, _ = get_backend() # automatically detects the underlying device type (CUDA, CPU, XPU, MPS, etc.)
processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50")
model = AutoModelForObjectDetection.from_pretrained("facebook/detr-resnet-50").to(device)
model = torch.compile(model)

texts = ["a photo of a cat", "a photo of a dog"]
inputs = processor(text=texts, images=image, return_tensors="pt").to(device)

with torch.no_grad():
    _ = model(**inputs)

Image Segmentation with Segformer

from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
from accelerate.test_utils.testing import get_backend

device, _, _ = get_backend() # automatically detects the underlying device type (CUDA, CPU, XPU, MPS, etc.)
processor = SegformerImageProcessor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")
model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512").to(device)
model = torch.compile(model)
seg_inputs = processor(images=image, return_tensors="pt").to(device)

with torch.no_grad():
    _ = model(**seg_inputs)

Below you can find the list of the models we benchmarked.

Image Classification

• google/vit-base-patch16-224
• microsoft/beit-base-patch16-224-pt22k-ft22k
• facebook/convnext-large-224
• microsoft/resnet-50

Image Segmentation

• nvidia/segformer-b0-finetuned-ade-512-512
• facebook/mask2former-swin-tiny-coco-panoptic
• facebook/maskformer-swin-base-ade
• google/deeplabv3_mobilenet_v2_1.0_513

Object Detection

• google/owlvit-base-patch32
• facebook/detr-resnet-101
• microsoft/conditional-detr-resnet-50

Below you can find visualization of inference durations with and without torch.compile() and percentage improvements for each model in different hardware and batch sizes.

Below you can find inference durations in milliseconds for each model with and without compile(). Note that OwlViT results in OOM in larger batch sizes.

A100 (batch size: 1)

A100 (batch size: 4)

A100 (batch size: 16)

V100 (batch size: 1)

V100 (batch size: 4)

V100 (batch size: 16)

T4 (batch size: 1)

T4 (batch size: 4)

T4 (batch size: 16)

PyTorch Nightly

We also benchmarked on PyTorch nightly (2.1.0dev, find the wheel here) and observed improvement in latency both for uncompiled and compiled models.

A100

T4

V100

Reduce Overhead

We benchmarked reduce-overhead compilation mode for A100 and T4 in Nightly.

A100

T4