[Draft][JAX] E2E encoder sanity test with synthetic data

[jberchtold-nvidia]

Description

Currently we have the following:

1. We have some small e2e model training tests for a small model on a few epochs as a sanity integration test.
2. We also provide examples for users on an MNIST model and a single-layer encoder, with gradual increases in complexity as parallelism techniques are added.

Currently 1 and 2 are the same tests. To avoid this coupling, this PR introduces the following changes:

• Example tests are still present but moved to the L1 test suite
• A new test file tests/jax/test_distributed_sanity_e2e_train.py is introduced will be run in L0 and which trains a small model for a few epochs with the following features
  • All parallelisms including single-GPU are included in the same file. This simplifies the test code and allows us to keep track of all loss/accuracy tolerances in a single place.
  • Uses synthetic data so does not require pulling HuggingFace datasets

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Move examples/jax/... tests to L1 test scripts
• Introduce a new test file tests/jax/test_distributed_sanity_e2e_train.py to consolidate all parallelism types covered by the examples into a single test file that uses synthetic data and runs in L0

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[jberchtold-nvidia]

/te-ci L2 jax