Skip to content

Low cost inference with AWS Inferentia

This demo uses Inf2 instance, read here for Inf1 demo.

AWS Inferentia is a high performance machine learning inference chip, custom designed by AWS. Amazon EC2 Inf2 instances are powered by AWS Inferentia chips, which provides you with the lowest cost per inference in the cloud and lower the barriers for everyday developers to use machine learning (ML) at scale. Customers using models such as YOLO v3 and YOLO v4 can get up to 1.85 times higher throughput and up to 40% lower cost per inference compared to the EC2 G4 GPU-based instances.

In the demo, you will learn how to run PyTorch model with DJL on Amazon EC2 Inf2 instances.

Setup environment

Launch Inf2 EC2 instance

Please launch Inf2 instance by following the Install Neuron Instructions

This demo tested on neuronx-tools and PyTorch 1.13.1 on Ubuntu DLAMI. Please make sure you have Neuron Runtime 2.x installed:

# Configure Linux for Neuron repository updates
. /etc/os-release

sudo tee /etc/apt/sources.list.d/neuron.list > /dev/null <<EOF
curl -L | sudo apt-key add -

# Update OS packages
sudo apt-get update -y

sudo apt-get install -y linux-headers-$(uname -r)
sudo apt-get install -y aws-neuronx-dkms=2.*
sudo apt-get install -y aws-neuronx-collectives=2.*
sudo apt-get install -y aws-neuronx-runtime-lib=2.*
sudo apt-get install -y aws-neuronx-tools=2.*

Install AWS Inferentia neuron SDK

The Inferentia neuron SDK is required for converting PyTorch pre-trained model into neuron traced model.

python3 -m venv myenv

source myenv/bin/activate
pip install -U pip

pip install neuronx-cc==2.* torch_neuronx==1.13.1.* torchvision --extra-index-url=

After installing the Inferentia neuron SDK, you will find is installed in myenv/lib/python3.8/site-packages/torch_neuronx/lib folder. You need configuration environment variable to enable Inferentia for DJL:

export PYTORCH_EXTRA_LIBRARY_PATH=$(python -m site | grep $VIRTUAL_ENV | awk -F"'" 'END{print $2}')/torch_neuronx/lib/

Compile your model into Neuron traced model

Use the following python script to trace a PyTorch resnet50 model. The script can also be found in the repo at

import torch
import os
import torch_neuronx
from torchvision import models
import logging

# Enable logging so we can see any important warnings
logger = logging.getLogger('Neuron')

# An example input you would normally provide to your model's forward() method.
image = torch.zeros([1, 3, 224, 224], dtype=torch.float32)

# Load a pretrained ResNet50 model
model = models.resnet50(pretrained=True)

# Tell the model we are using it for evaluation (not training)

# Use torch.jit.trace to generate a torch.jit.ScriptModule via tracing.
djl_traced_model = torch.jit.trace(model, image)

# Save the Regular TorchScript model for benchmarking
os.makedirs("models/djl/resnet50", exist_ok=True)"models/djl/resnet50/")

# Now compile the model - with logging set to "info" we will see
# what compiles for Neuron, and if there are any fallbacks
model_neuron = torch_neuronx.trace(model, image)

# Export to saved model
os.makedirs("models/inferentia/resnet50", exist_ok=True)"models/inferentia/resnet50/")
print("Compile success")
git clone
cd djl-demo/aws/inferentia


Execute above command, now you have a Neuron traced model ready for inference in models/inferentia/resnet50 folder.

Run example java program

Notes: in torch-neuron 1.11.x and 1.12.x, the stack size must be set to 2M and more. The default JVM stack size is 1M, you have to explicitly set the JVM arguments: -Xss2m

cd djl-demo/aws/inferentia

./gradlew run

[INFO ] - Number of inter-op threads is 4
[INFO ] - Number of intra-op threads is 8
Running inference with PyTorch: 1.13.1
        class: "n02124075 Egyptian cat", probability: 0.41596
        class: "n02123159 tiger cat", probability: 0.26856
        class: "n02123045 tabby, tabby cat", probability: 0.23701
        class: "n02123394 Persian cat", probability: 0.04384
        class: "n02127052 lynx, catamount", probability: 0.00612


You can use DJL benchmark tool to compare performance w/o Inferentia enabled:

Run multithreading benchmark with Inferentia enabled:

./gradlew benchmark

[INFO ] - Running inference with PyTorch: 1.13.1
[INFO ] - Loading from: /home/ubuntu/myenv/lib/python3.8/site-packages/torch_neuronx/lib/
[INFO ] - Multithreaded inference with 8 threads.
[INFO ] - Throughput: 288.59, completed 8000 iteration in 27721 ms.
[INFO ] - Latency P50: 27.697 ms, P90: 27.915 ms, P99: 28.426 ms

Run multi-thread benchmark with regular PyTorch model:

./gradlew benchmark --args="models/djl/resnet50"

[INFO ] - Running inference with PyTorch: 1.13.1
[INFO ] - Loading regular pytorch model ...
[INFO ] - Multithreaded inference with 8 threads.
[INFO ] - Throughput: 33.92, completed 8000 iteration in 235833 ms.
[INFO ] - Latency P50: 234.925 ms, P90: 257.252 ms, P99: 277.015 ms

Run single-threaded benchmark with Inferentia enabled:

./gradlew benchmark --args="models/inferentia/resnet50 1"

Run single-thread benchmark with regular PyTorch model:

./gradle benchmark --args="models/djl/resnet50 1"