ML Training Reference Architectures & Tests

This repository contains reference architectures and test cases for distributed model training with Amazon SageMaker Hyperpod, AWS ParallelCluster, AWS Batch, and Amazon EKS. The test cases cover different types and sizes of models as well as different frameworks and parallel optimizations (Pytorch DDP/FSDP, MegatronLM, NemoMegatron...).

The major components of this directory are:

reference-architectures/
|-- 1.architectures/               # CloudFormation templates for reference arch
|-- 2.ami_and_containers/          # Scripts to create AMIs and container images
|-- 3.test_cases/                  # Reference test cases and/or benchmark scripts
|-- 4.validation_observability/    # Tools to measure performance or troubleshoot
`-- ...

NOTE: the architectures are designed to work with the S3 bucket and VPC created using reference templates 1.architectures/0.s3/ and 1.architectures/1.vpc_network/. You're strongly recommended to deploy these two templates before deploying any of the reference architectures.

1. Architectures

Architectures are located in 1.architectures and consists of utilities and service related architectures.

More will come, feel free to add new ones (EKS, Ray?). You will also find documentation for EFA and the recommended environment variables.

2. Custom Amazon Machine Images

Custom machine images can be built using Packer for AWS ParallelCluster, Amazon EKS and plain EC2. These images are based are on Ansible roles and playbooks.

3. Test cases: support matrix

All test cases are under 3.test_cases/. You can go in each test case directory to learn how to run it.

4. Validation scripts

Utilities scripts and micro-benchmarks examples are set under 4.validation_scripts/. The EFA Prometheus exporter can be found in this directory

5. CI

Integration tests are written in pytest. Just run:

pytest .

Alternatively you can run tests with out capturing stdout and keeping all docker images an other artifacts.

pytest -s --keep-artifacts=t

6. Contributors

Thanks to all the contributors for building, reviewing and testing.