[PyTorch] Fix get_swa_mask() for padding masks

[cyanguwa]

Description

This PR fixes the mask generation for sliding window in UnfusedDotProductAttention. It fixes the logic for padding and arbitrary masks in get_swa_mask(), adds more docstring, refactors the call site, and adds more testing in the unit tests.

Fixes #1271

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

Please list the changes introduced in this PR:

• Improve the logic in get_swa_mask() and its call site

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

[cyanguwa]

/te-ci pytorch

[Marks101]

Hi @cyanguwa,
great, I like the idea to have all the masking logic at one place 👍
I just tested this and found a problem with cross attention:

        if "padding" in attn_mask_type:
            if max_seqlen_q == max_seqlen_kv:
                attention_mask = torch.logical_or(
>                   attention_mask.squeeze(1).unsqueeze(3), attention_mask
                )
E               AttributeError: 'tuple' object has no attribute 'squeeze'

The code in UnfusedDotProductAttention made these lines dependent on the attention_type.

[cyanguwa]

Hi @cyanguwa, great, I like the idea to have all the masking logic at one place 👍 I just tested this and found a problem with cross attention:

        if "padding" in attn_mask_type:
            if max_seqlen_q == max_seqlen_kv:
                attention_mask = torch.logical_or(
>                   attention_mask.squeeze(1).unsqueeze(3), attention_mask
                )
E               AttributeError: 'tuple' object has no attribute 'squeeze'

The code in UnfusedDotProductAttention made these lines dependent on the attention_type.

Yes, I think I should use if attention_type == "self" here because there could be cross-attention cases where max_seqlen_q == max_seqlen_kv and actual_seqlen_q != actual_seqlen_kv. I'll go through attention.py and see if there're other places I should use attention_type instead.

Let me know if you observe any other issues too! :) Thanks!

[cyanguwa]

/te-ci pytorch L0

[xrennvidia]

LGTM. Thanks.