Fix cuda graph capture for grouped gemm

[xrennvidia]

Description

Cuda graph does not work with Grouped GEMM.
The saved forward activations are corrupted before bwd_graph is replayed. Explicitly setting retain_graph=True can hold the activations and fix the 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

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

[timmoon10]

Wouldn't we expect this to increase memory usage?

I see that torch.cuda.make_graphed_callables doesn't set retain_graph=True:
https://github.com/pytorch/pytorch/blob/c25b201583fc28243b87c460a2f18e2531a676e7/torch/cuda/graphs.py#L326-L336
We want to match plain PyTorch as much as possible unless there is a good reason to introduce divergence. If this is MoE-specific, perhaps we could add a kwarg like retain_graph_in_backward that is False by default.

[xrennvidia]

/te-ci pytorch

[timmoon10]

LGTM

[xrennvidia]

/te-ci pytorch