[Bugfix] Fix bias for 0-dim tensors in gemm

[yaox12]

Description

For te_gemm and te_grouped_gemm,

• If grad == false, we should do nothing to bias.
• If grad == true, we should calculate grad_bias here.

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

[timmoon10]

We should handle the case where B has no data:

Suggested change

	// torch.sum is able to handle 0-dim tensors
	if (bias.data_ptr() != nullptr && grad) bias.copy_(B.sum(0));
	if (bias.data_ptr() != nullptr && grad) {
	if (B.data_ptr() == nullptr) {
	bias.zero_();
	} else {
	bias.copy_(B.sum(0))
	}
	}

It seems we are checking data_ptr() == nullptr in order to check for empty tensors. As an alternative, it may be better to make our intention explicit:

  if (A.numel() == 0 || B.numel() == 0) {
    if (D.numel() != 0 && !accumulate) D.zero_();
    // torch.sum is able to handle 0-dim tensors
    if (bias.numel() != 0 && grad) bias.copy_(B.sum(0));
    if (pre_gelu_out.numel() != 0) pre_gelu_out.zero_();
    return;
  }

[yaox12]

Updated.

[yaox12]

/te-ci pytorch

[ksivaman]

LGTM, thanks!

[yaox12]

/te-ci pytorch

[yaox12]

/te-ci pytorch

[yaox12]

I'll merge this PR since the CI passes except for a doc test failing with irrelevant error.