A couple of weeks ago, Hugging Face and AWS announced they will partner to make AI open and more accessible. Part of this partnership is to develop tools that make it easier for practitioners to leverage AWS purpose-built instance AWS Trainium and AWS Inferentia to train, fine-tune, and deploy Transformer and Diffusion models on AWS.

“AWS Trainium is the second-generation machine learning (ML) chip that AWS purpose-built for deep learning training. […] Trainium-based EC2 Trn1 instances solve this challenge by delivering faster time to train while offering up to 50% cost-to-train savings over comparable GPU-based instances.” - AWS

We are super excited to bring these price-performance advantages to Transformers and Diffusers. 🚀

In this hands-on post, We'll show you how to quickly set up an AWS Trainium using the Hugging Face Neuron Deep Learning AMI to fine-tune a BERT model for text classification.

Let's get started! 🔥

Setting up an AWS Trainium instance on AWS

The simplest way to work with AWS Trainium and Hugging Face Transformers is the Hugging Face Neuron Deep Learning AMI (DLAMI). The DLAMI comes with all required libraries pre-packaged for you, including the Neuron Drivers, Transformers, Datasets, and Accelerate.

To create an EC2 Trainium instance, you can start from the console or the Marketplace. This guide will start from the EC2 console.

Starting from the EC2 console in the us-east-1 region, you first click on Launch an instance and define a name for the instance (trainium-huggingface-demo).

Next you search the Amazon Marketplace for Hugging Face AMIs. Entering “Hugging Face” in the search bar for “Application and OS Images” and hitting “https://www.philschmid.de/static/blog/setup-aws-trainium/02-search-ami.pngrainium/02-search-ami.png" alt="02-search-ami">

This should now open the “Choose an Amazon Machine Image” view with the search. You can now navigate to “AWS Marketplace AMIs” and find the Hugging Face Neuron Deep Lehttps://www.philschmid.de/static/blog/setup-aws-trainium/03-select-ami.png"/static/blog/setup-aws-trainium/03-select-ami.png" alt="03-select-ami">

You will be asked to subscribe if you aren’t. The AMI is completely free of charge, and you will only pay for the EC2 compute.

Then you need to define a key pair, which will be used to connect to the instance via ssh. https://www.philschmid.dehttps://www.philschmid.de/static/blog/setup-aws-trainium/04-select-key.pngey pair.

After that, create or select a

AWS will now provision the instance using the Hugging Face Neuron Deep Learning AMI. Additional configurations can be made by increasing the disk space or creating an instance profile to access othhttps://www.philschmid.dehttps://www.philschmid.de/static/blog/setup-aws-trainium/07-copy-dns.png, you can view and copy the public IPv4 address to ssh into the machine.

Replace the empty strings "" in the snippet below with the IP address of your instances and the path to the key pair you created/selected when launching the instance.

PUBLIC_DNS="" # IP address, e.g. ec2-3-80-....KEY_PATH="" # local path to key, e.g. ssh/trn.pem ssh -i $KEY_PATH ubuntu@$PUBLIC_DNS

After you are connected, you can run neuron-ls to ensure you have access to the Trainium accelerators. You should see a similar output than below.

ubuntu@ip-172-31-79-164:~$ neuron-lsinstance-type: trn1.2xlargeinstance-id: i-0570615e41700a481+--------+--------+--------+---------+| NEURON | NEURON | NEURON |   PCI   || DEVICE | CORES  | MEMORY |   BDF   |+--------+--------+--------+---------+| 0      | 2      | 32 GB  | 00:1e.0 |+--------+--------+--------+---------+

Fine-tune a BERT model for text classification

The Hugging Face Neuron Deep Learning AMI not only comes with all libraries pre-packaged, it also includes all supported huggingface-neuron-samples/ scripts from the optimum-neuron library. This means you can directly launch our training job using the text-classification script.

The training scripts use the new TrainiumTrainer, a purpose-built Transformers Trainer for AWS Trainium. The TrainiumTrainer comes with several benefits, including a compilation cache. This means we can skip the compilation step (~10-15 min) for your model + configuration if it is cached already. Learn more about the cache in the documentation.

The training script will download the bert-base-uncased model from the Hugging Face hub and fine-tunes it on the emotion dataset, which consists of 10 000 Twitter messages with six labels: anger, fear, joy, love, sadness, and surprise.

The trn1.2xlarge instance comes with 2 Neuron Cores. Therefore torchrun is used to leverage both and launch our training.

torchrun --nproc_per_node=2 huggingface-neuron-samples/text-classification/run_glue.py \--model_name_or_path bert-base-uncased \--dataset_name philschmid/emotion \--do_train \--do_eval \--bf16 True \--per_device_train_batch_size 16 \--learning_rate 5e-5 \--num_train_epochs 3 \--output_dir ./bert-emotion

Note: If you see bad, bad accuracy, you might want to deactivate bf16 for now.

After 2 minutes and 42 seconds, the training was completed and achieved an excellent accuracy of 0.925.

***** train metrics *****  epoch                    =        3.0  train_loss               =     0.3032  train_runtime            = 0:02:42.34  train_samples            =      16000  train_samples_per_second =    295.663  train_steps_per_second   =      9.239 ***** eval metrics *****  epoch                   =        3.0  eval_accuracy           =      0.925  eval_loss               =     0.2057  eval_runtime            = 0:00:07.41  eval_samples            =       2000  eval_samples_per_second =    269.779  eval_steps_per_second   =     16.861

Last but not least, terminate the EC2 instance to avoid unnecessary charges. Looking at the price-performance, the training only cost 7ct (1.34$/h * 0.05h = 0.07$)

Conclusion

In conclusion, the combination of AWS Trainium and Hugging Face Transformers provides a powerful, simple, and cost-effective solution for training state-of-the-art natural language processing models. By leveraging the purpose-built Trainium instances, practitioners can achieve faster training times and cost savings over comparable GPU-based instances. With the Hugging Face Neuron Deep Learning AMI, all required libraries are pre-packaged, making it easy to fine-tune Transformer models for extractive or generative use cases.

Give it a try, and let us know what you think. We welcome your questions and feedback on the Hugging Face Forum.

Thanks for reading! If you have any questions, feel free to contact me on Twitter or LinkedIn.