[PyTorch] Avoid parameters function in op backward pass

[timmoon10]

Description

We have recently experienced some esoteric errors in the LayerNorm backward pass:

[rank0]:   File "/usr/local/lib/python3.12/dist-packages/torch/_tensor.py", line 626, in backward
[rank0]:     torch.autograd.backward(
[rank0]:   File "/usr/local/lib/python3.12/dist-packages/torch/autograd/__init__.py", line 347, in backward
[rank0]:     _engine_run_backward(
[rank0]:   File "/usr/local/lib/python3.12/dist-packages/torch/autograd/graph.py", line 823, in _engine_run_backward
[rank0]:     return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/usr/local/lib/python3.12/dist-packages/torch/autograd/function.py", line 307, in apply
[rank0]:     return user_fn(self, *args)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/usr/local/lib/python3.12/dist-packages/torch/autograd/function.py", line 600, in wrapper
[rank0]:     outputs = fn(ctx, *args)
[rank0]:               ^^^^^^^^^^^^^^
[rank0]:   File "/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/ops/fuser.py", line 267, in backward
[rank0]:     raise RuntimeError(
[rank0]: RuntimeError: Expected op 0 to generate 0 param grads, but got 2

I haven't fully investigated, but I suspect that FSDP is manipulating module parameters so that they are only available in the forward pass. This messes with a check the operation fuser does in the backward pass to make sure the number of params and param grads match.

This PR tweaks the operation fuser to avoid calling parameters() in the backward pass. In particular, it counts params for each op in the forward pass and caches the counts for use in the backward pass.

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

• Operation fuser avoids calling op parameters function in backward pass

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