class

SoftplusBackward

SoftplusBackward

source edit

Python wrappers

Autograd node for Softplus $y = \log(1 + e^x)$ — the smooth surrogate of ReLU whose derivative is the logistic sigmoid.

Saves the input so grad_formula can recompute $\sigma(x)$ via the storage-level sigmoid primitive (a single MLX or Accelerate call).

Math

$y = \log(1 + e^x), \qquad \frac{\partial y}{\partial x} = \sigma(x)$

Attributes

schema_v1OpSchema

"softplus", AmpPolicy::ForceFP32 — guards against

e^x

overflow in float16 for

|x| > 11

or so.

Static methods

static

dispatch

int dispatch(backend::IBackend & be, const int & a, const int & s, Dtype dt)

source edit

Forward — calls IBackend::softplus to compute $y = \log(1 + e^x)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Eager-mode backward: $\partial L/\partial x = \sigma(x) \cdot \partial L/\partial y$ , evaluated via the storage-level sigmoid kernel applied to the saved input.

class

SoftplusBackward

SoftplusBackward

source edit

Python wrappers

Autograd node for Softplus $y = \log(1 + e^x)$ — the smooth surrogate of ReLU whose derivative is the logistic sigmoid.

Saves the input so grad_formula can recompute $\sigma(x)$ via the storage-level sigmoid primitive (a single MLX or Accelerate call).

Math

$y = \log(1 + e^x), \qquad \frac{\partial y}{\partial x} = \sigma(x)$

Attributes

schema_v1OpSchema

"softplus", AmpPolicy::ForceFP32 — guards against

e^x

overflow in float16 for

|x| > 11

or so.

Static methods

static

dispatch

int dispatch(backend::IBackend & be, const int & a, const int & s, Dtype dt)

source edit

Forward — calls IBackend::softplus to compute $y = \log(1 + e^x)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Eager-mode backward: $\partial L/\partial x = \sigma(x) \cdot \partial L/\partial y$ , evaluated via the storage-level sigmoid kernel applied to the saved input.