Hierarchical CP implementation (Ulysses + Ring)

[xrennvidia]

Description

Add a hierarchical implementation of context parallelism. It applies A2A communications in low-level CP groups (e.g., via NVLink), and applies P2P communications in high-level CP groups (e.g., via IBLink). Refer details in LongVILA and USP.

This implementation supports:
dtype: BF16, FP16, FP8
mask_type: causal, no_mask
MHA and QGA

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 refractor

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

[xrennvidia]

/te-ci pytorch

[xrennvidia]

/te-ci pytorch

[xrennvidia]

/te-ci pytorch