🐛 Inappropriate Test Tolerance Threshold for BF16 Precision Validation​

[townwish4git]

🚀 ​​Issue Description​​
Currently, in our model precision testing pipeline, we apply ​​identical precision error tolerance thresholds​​ for both FP16 and BF16 data types. This approach is technically problematic because BF16's inherent precision is lower than FP16 due to its shorter mantissa (7 bits vs. 10 bits). Applying the same strict thresholds to BF16 fails to account for its fundamental numerical characteristics and may lead to unnecessary test failures for valid results