attention_mask fill with -inf for UnfusedDotProductAttention

[Agoniii]

Description

UnfusedDotProductAttention in TE uses -10000 to fill in the attention mask, but the value is not small enough for some cases which leads to large diff between TE and HF.

The logits of model forward with FP32.
HF baseline:

logits: tensor([[[-0.6598, -4.5184, -4.0881,  ..., -5.0321, -5.0322, -5.0315],
         [ 2.3831,  2.2049, -0.3057,  ..., -6.5912, -6.5907, -6.5909],
         [-2.3334, -2.7213, -4.7815,  ..., -9.8526, -9.8525, -9.8523],
         ...,
         [ 3.3538,  1.4944,  1.0958,  ..., -6.4574, -6.4574, -6.4577],
         [ 5.6523,  0.4127,  1.0138,  ..., -8.7066, -8.7066, -8.7068],
         [ 0.3297, -0.7106, -0.8580,  ..., -7.7929, -7.7928, -7.7930]]],
       device='cuda:0', grad_fn=<UnsafeViewBackward0>)

TE before:

logits: tensor([[[-0.6598, -4.5184, -4.0881,  ..., -5.0321, -5.0322, -5.0316],
         [ 2.1408,  2.3935, -0.1826,  ..., -6.7005, -6.7000, -6.7002],
         [-2.5307, -3.0045, -4.9231,  ..., -9.4020, -9.4019, -9.4017],
         ...,
         [ 3.3538,  1.4944,  1.0958,  ..., -6.4574, -6.4574, -6.4578],
         [ 5.6524,  0.4127,  1.0138,  ..., -8.7066, -8.7066, -8.7068],
         [ 0.3297, -0.7106, -0.8580,  ..., -7.7929, -7.7928, -7.7930]]],
       device='cuda:0', grad_fn=<TransposeBackward0>)

TE after:

logits: tensor([[[-0.6598, -4.5184, -4.0881,  ..., -5.0321, -5.0322, -5.0316],
         [ 2.3831,  2.2049, -0.3057,  ..., -6.5912, -6.5907, -6.5909],
         [-2.3334, -2.7213, -4.7815,  ..., -9.8526, -9.8525, -9.8523],
         ...,
         [ 3.3538,  1.4944,  1.0958,  ..., -6.4574, -6.4574, -6.4578],
         [ 5.6524,  0.4127,  1.0138,  ..., -8.7066, -8.7066, -8.7068],
         [ 0.3297, -0.7106, -0.8580,  ..., -7.7929, -7.7928, -7.7930]]],
       device='cuda:0', grad_fn=<TransposeBackward0>)

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:

• Change A
• Change B

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

[xrennvidia]

/te-ci pytorch

[xrennvidia]

/te-ci pytorch