[jberchtold-nvidia]

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

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

Please list the changes introduced in this PR:

• Change A
• Change B

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

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The change reaches JAX expert-parallel MoE partitioning and its test setup.

The fully replicated EP=1 case should be fixed before merging.

Findings

1.  Replicated batches still fail ▶

Summary

The PR allows an omitted singleton ep axis when another batch axis remains.

• MoE partitioning accepts a missing EP axis when EP size is one.

_{Reviews (1) · Last reviewed commit: "Fix EP=1 case in TE MoEBlock" · Reviewed by Greptile}

[greptile-apps]

Replicated batches still fail

With ep=1, no dp_resource or fsdp_resource, and another mesh axis spanning multiple devices, MoE assigns the batch only to ep. Once JAX removes that singleton axis, _leading_axis_ok still rejects the fully replicated spec: a leading None fails the allowed-axis check, and an empty spec fails the length check. Dispatch then raises NotImplementedError, so this valid EP=1 case still cannot compile.

Allow fully replicated leading dimensions when ep_size == 1 and add a regression test without a retained outer axis.