class

ExpBackward

ExpBackward

source edit

Python wrappers

pylucid.ops.expfunction

Autograd node for element-wise exponential $y = e^x$ .

Saves the output $y$ rather than the input because the backward formula is $\partial y / \partial x = e^x = y$ and reusing the saved output avoids a second vvexpf pass.

Math

\frac{\partial y}{\partial x} = e^x = y, \qquad \frac{\partial \mathcal{L}}{\partial x} = y\,\frac{\partial \mathcal{L}}{\partial y}.$$

Attributes

schema_v1OpSchema

Registered as "exp" with AmpPolicy::ForceFP32 — half- precision inputs are upcast to F32 prior to dispatch to avoid premature overflow / underflow.

kSavesInputbool

false.

kSavesOutputbool

true — saved tensor is

y = e^x

Notes

Dispatch: Accelerate vvexpf/vvexp (CPU) / MLX exp (GPU). Large-magnitude positive inputs overflow to +inf; large-magnitude negative inputs flush to 0.

Static methods

static

dispatch

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

source edit

Forward — calls IBackend::exp to compute $y = e^x$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Backward — $\partial y/\partial x = e^x = y$ , scaled by grad_out.

grad_formula_impl

int grad_formula_impl(const int & g, const int &, const int & out)

source edit

Graph-mode backward: $\partial x = \mathrm{out} \cdot g$ using the saved output $e^x$ , kept inside the autograd graph.

class

ExpBackward

ExpBackward

source edit

Python wrappers

pylucid.ops.expfunction

Autograd node for element-wise exponential $y = e^x$ .

Saves the output $y$ rather than the input because the backward formula is $\partial y / \partial x = e^x = y$ and reusing the saved output avoids a second vvexpf pass.

Math

\frac{\partial y}{\partial x} = e^x = y, \qquad \frac{\partial \mathcal{L}}{\partial x} = y\,\frac{\partial \mathcal{L}}{\partial y}.$$

Attributes

schema_v1OpSchema

Registered as "exp" with AmpPolicy::ForceFP32 — half- precision inputs are upcast to F32 prior to dispatch to avoid premature overflow / underflow.

kSavesInputbool

false.

kSavesOutputbool

true — saved tensor is

y = e^x

Notes

Dispatch: Accelerate vvexpf/vvexp (CPU) / MLX exp (GPU). Large-magnitude positive inputs overflow to +inf; large-magnitude negative inputs flush to 0.

Static methods

static

dispatch

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

source edit

Forward — calls IBackend::exp to compute $y = e^x$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Backward — $\partial y/\partial x = e^x = y$ , scaled by grad_out.

grad_formula_impl

int grad_formula_impl(const int & g, const int &, const int & out)

source edit

Graph-mode backward: $\partial x = \mathrm{out} \cdot g$ using the saved output $e^x$ , kept inside the autograd graph.