A high-performance inference engine for AI models
remove clamp
eugene37c4a74fcea2
diff --git a/crates/uzu/src/backends/metal/kernel/mlp/activation_mul.metal b/crates/uzu/src/backends/metal/kernel/mlp/activation_mul.metalindex 0f8b2a15a..9a8c3a95c 100644--- a/crates/uzu/src/backends/metal/kernel/mlp/activation_mul.metal+++ b/crates/uzu/src/backends/metal/kernel/mlp/activation_mul.metal@@ -20,8 +20,6 @@ inline T gelu_approx(T x) {constexpr float k1 = 0.7978845608f; // sqrt(2/pi)float xf = float(x);float t = k1 * (xf + k0 * xf * xf * xf);- // Avoid NaNs from fast `tanh` implementations that may overflow internally.- t = clamp(t, -20.0f, 20.0f);float yf = 0.5f * xf * (1.0f + metal::precise::tanh(t));return T(yf);}