Problem

Two scaled dot product attention (SDPA) CPU kernel tests fail on RHEL due to minor numerical precision differences.

Root Cause

The CPU-based fused attention kernel produces slightly different numerical results due to floating-point arithmetic differences and GCC 11 compiler optimizations.

Impact

• Tests failing: 2
• Severity: Very Low - Numerical precision issue, not functional
• Production impact: None - Attention mechanism works correctly
• Mismatched elements: 1 / 1,632 (0.1%)

Proposed Solutions

Option 1: Relax test tolerance for RHEL builds
Option 2: Exclude specific test variants
Option 3: Report to PyTorch upstream for investigation

Acceptance Criteria

• [ ] Root cause fully understood
• [ ] Fix implemented (tolerance relaxation or exclusion)
• [ ] Tests pass consistently on RHEL
• [ ] No regression in actual SDPA functionality

References

• Analysis: TEST_FAILURE_REPORT.md
• Test file: test/test_transformers.py