[PyTorch] Fix L3 FA tests

[cyanguwa]

Description

This PR fixes the L3 tests for FP8 current scaling in L3_pytorch_FA_versions_test. The fix is only related to the selection logic in the test and not the backend support itself.

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 refactoring

Changes

Please list the changes introduced in this PR:

• L3 CI test fix.

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

[greptile-apps]

Greptile Summary

Fixed backend availability checking logic in FP8 attention tests (test_mha_fp8_vs_f16 and test_dpa_fp8_vs_f16). Previously, F16 reference backend availability was only checked when fp8_dpa_bwd=False, which could cause test failures when fp8_dpa_bwd=True if the F16 backend wasn't available.

Key changes:

• Both tests now unconditionally check availability of FP8 and F16 backends using two separate calls to get_available_attention_backends
• Improved variable naming: fused_attn_supported split into fused_attn_supported_fp8 and fused_attn_supported_f16 for clarity
• Test execution and comparisons properly gated by checking which backends are actually available
• More descriptive skip message: "No reference backend available" when F16 backend is missing

Confidence Score: 5/5

• This PR is safe to merge with no risk
• Test-only changes that fix incorrect skip logic without modifying any production code or backend implementation. The changes ensure proper backend availability checking and correct test gating
• No files require special attention

Important Files Changed

Filename	Overview
tests/pytorch/attention/test_attention.py	Fixed backend availability checks to always verify both FP8 and F16 backends regardless of fp8_dpa_bwd flag, ensuring proper test execution and skipping

_{Last reviewed commit: 229b1ef}

[greptile-apps]

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[cyanguwa]

/te-ci pytorch L3

[KshitijLakhani]

So if the value for fused_attn_supported_fp8=1/True and fp8_dpa_bwd=0/False then fused_attn_supported_f16 be 0/False or 1/True right ?
If fused_attn_supported_f16 is 0/False, there's no fp16 to compare the fp8 and hence we skip, but
,if fused_attn_supported_f16 is 1/True, then compare fp8 and fp16 fwd only.

Now, if fp8_dpa_bwd=1/True, then fused_attn_supported_f16 will be False/0 always
In that case what does fp8 get compared to ? Maybe I missed it, but I see no logic for this comparison - could you point me to it ?

[cyanguwa]

Makes sense. I removed the if not fp8_bwd logic. Could you please take a look at 229b1ef?

[cyanguwa]

/te-ci pytorch L3

[KshitijLakhani]

LGTM !
Thanks !

[cyanguwa]

Pipeline 44991844.