[PyTorch] Minor optimizations to reduce CPU overheads in modules

[timmoon10]

Description

We have observed that TE modules experience non-trivial CPU overhead, which often becomes a performance bottleneck in the forward pass of small models. For example, measuring the CPU runtime for Megatron-core modules with BF16 compute and TP=1:

• ColumnParallelLinear: 74 us per forward pass
• TEColumnParallelLinear: 140 us per forward pass

Unfortunately this overhead is distributed throughout the forward pass. Many basic PyTorch operations, e.g. getting attributes from torch.Tensor, involve O(1 us) overhead, so even basic checks to handle all of our advanced features will eventually add up to something non-trivial.

This PR makes a few minor optimizations:

• Avoid importing from te.pytorch.cpu_offload in every forward pass
• Memoize NCCL process group properties
• Avoid custom logic in torch.nn.Module.__setattr__ when possible
• Avoid custom logic for accessing params in torch.nn.Module when possible
• Avoid accessing tensor attrs more than necessary

I see a 1.22x speedup, with 115 us per forward 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

• Avoid importing from te.pytorch.cpu_offload in every forward pass
• Memoize NCCL process group properties
• Avoid custom logic in torch.nn.Module.__setattr__ when possible
• Avoid custom logic for accessing params in torch.nn.Module when possible
• Avoid accessing tensor attrs more than necessary

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

[yaox12]

Can you propagate the CPU offloading importing fix to GroupedLinear as well?

TransformerEngine/transformer_engine/pytorch/module/grouped_linear.py

Line 783 in c0caadb

from ..cpu_offload import CPUOffloadEnabled

[ptrendx]

I wonder if we couldn't instead just do something like

te_params = self.get_te_params()  # calls _fast_getattr internally, te_params is a normal object
te_params.fp8_parameters = FP8GlobalStateManager.with_fp8_parameters()

[ptrendx]

Otherwise it will be hard to enforce everybody using only the _fast_get/setattr I think.

[timmoon10]

Even better we could store these attrs in fp8_meta or some other dict. I feel like the behavior of torch.nn.Module is a hint we shouldn't change its attrs frequently.

[timmoon10]

I tried wrapping these attrs in a property that internally calls _fast_setattr, but that added ~10 us overhead (probably from the extra indirection when getting the attrs). I think it's a good idea to refactor these frequently changed attrs so they are not held directly by the module, but I think that would be beyond the scope of this PR.

[timmoon10]

Moving this logic into the __setattr__ function makes things a little cleaner. It adds ~2 us overhead, but it's still a win of ~6 us compared to the baseline.

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[ptrendx]

Suggested change

	if tensor is None:
	return None
	if tensor.dtype == dtype:
	return tensor
	if tensor is None or tensor.dtype == dtype
	return tensor