[PyTorch] Fix AttentionParams comparison logic

[cyanguwa]

Description

DotProductAttention calls get_attention_backend() to choose an appropriate backend for a given AttentionParams. This call is CPU intensive and is only called when attention_params != _attention_backends["attention_params"]. For FP8, there are a number of entries in AttentionParams.fp8_meta, for example, fp8_meta["scaling_fwd"], but what's really important is fp8_meta["recipe"]. This PR overwrites the __eq__ function of AttentionParams to only compare the recipe entry for FP8 configs.

Before this PR, NVTE_DEBUG=1 NVTE_DEBUG_LEVEL=1 pytest -o log_cli=true --log-cli-level=INFO -s -v tests/pytorch/test_sanity.py::test_sanity_attention_extra_state[dtype0-large] produces:

INFO     DotProductAttention:attention.py:8245 Running with FusedAttention backend (sub-backend 2)
INFO     DotProductAttention:attention.py:8245 Running with FusedAttention backend (sub-backend 2)
INFO     DotProductAttention:attention.py:8245 Running with FusedAttention backend (sub-backend 2)
INFO     DotProductAttention:attention.py:8245 Running with FusedAttention backend (sub-backend 2)
INFO     DotProductAttention:attention.py:8245 Running with FusedAttention backend (sub-backend 2)
PASSED

After this PR:

INFO     DotProductAttention:attention.py:8245 Running with FusedAttention backend (sub-backend 2)
PASSED

This test_sanity_attention_extra_state test runs an FP8 config in multiple ways (i.e. with and without checkpointing), but it's the same FP8 config, so there should be only one INFO message (i.e. one get_attention_backend call), which is what's happening after the PR change.

Fixes #1348

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:

• Restrict AttentionParams.fp8_meta comparison to just fp8_meta["recipe"]

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 L0