[PyTorch] Debug checkpointing with operation-based API

[timmoon10]

Description

This PR debugs checkpointing with the operation-based API (see #707), in particular by adding logic to include FP8 scaling factors in the checkpoint. The checkpointing logic is very similar to the module-based API:

TransformerEngine/transformer_engine/pytorch/module/base.py

Line 555 in 5b6546c

def get_extra_state(self) -> torch.Tensor:

TransformerEngine/transformer_engine/pytorch/module/base.py

Line 587 in 5b6546c

def set_extra_state(self, state: torch.Tensor) -> None:

It is admittedly rather unintuitive, but I've added comments to justify the weird behavior.

I've also fixed an orthogonal bug where the linear op was including two copies of its params in its checkpoint.

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

• Add logic for FP8 scales when checkpointing with operation-based API
• Fix bug where linear op checkpoint saves params twice

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

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

These changes in the Linear op are orthogonal to the rest of this PR. PyTorch modules save checkpoints recursively, so checkpointing a fused op (e.g. Linear) will also checkpoint its constituent basic ops (e.g. BasicLinear, Bias). By registering the weight and bias params with the Linear op, the checkpoints were saving two copies of the params. Converting weight and bias into Python properties avoids this behavior while retaining the existing API.

[timmoon10]

/te-ci pytorch

[timmoon10]

Merging with approval from @ptrendx and @ksivaman.