Support Deepseekv2

[hongywei]

Description

The current TE version does not seem to support Deepseekv2, so some modifications have been made to attention.py.

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 refractor

Changes

Please list the changes introduced in this PR:
only modify attention.py

• 2292line modify _get_qkv_layout() function according to nvidia/Transformer_engine
• 4542line kv_channels:int - > kv_channels: Union[int, Tuple[int, int]] and other changes such as 4578-4583line
• 4602line modify softmax_scale computation
• 4881line comment out, because deepseekv2 does not require k and vshape to be consistent.

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

[wenchenvincent]

@hongywei Could you please describe which attention features that deepseek v2 need?

[hongywei]

@wenchenvincent Deepseekv2 does not require the shape of k and v to be consistent. I also modified it based on nvidia/transformer_engine. I think you can also look here to get the answer.
https://github.com/NVIDIA/TransformerEngine/blob/df94903704bd850977fdaf93e18fc73e7df58ef7/transformer_engine/pytorch/attention.py#L4987
https://github.com/alibaba/Pai-Megatron-Patch/blob/bf582d8f30d8ffbba51db3dcda984c9f0261d57d/megatron_patch/model/deepseek_v2/transformer/attention.py#L414