Fixes #186278
fast::tanh uses expression (exp(2x)-1)/(exp(2x)+1) as the approximation. For gelu the expression for x scales cubically and will overflow the internal fp32 when abs(input) > 10.0. At this point the tanh approximation is already very much saturated already so we'll just clamp it to [-10, 10]. Perfwise clamping is cheap.
Pull...