9961d244f2
Previously, the slope tensor was set to fp16 to improve efficiency. While this worked correctly in FA, it caused precision issues in soft_max. This change applies different data types for different operators to balance both accuracy and performance.