Skip to content

[Precision Depth Alignment] paddle.sqrt and paddle.rsqrt accuracy and torch alignment #75367

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 22, 2025
Merged

[Precision Depth Alignment] paddle.sqrt and paddle.rsqrt accuracy and torch alignment #75367

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions paddle/phi/kernels/funcs/activation_functor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4322,11 +4322,11 @@ struct CudaSqrtFunctor : public BaseActivationFunctor<T> {

template <typename T>
struct CudaSqrtGradFunctor : public BaseActivationFunctor<T> {
T one_half = static_cast<T>(0.5f);
T two = static_cast<T>(2);

// dx = dout * 0.5 / out
// dx = dout / (2 * out)
__device__ __forceinline__ T operator()(const T dout, const T out) const {
return one_half * dout / out;
return dout / (two * out);
}

static constexpr ActBwdOpFwdDeps FwdDeps() {
Expand Down Expand Up @@ -4413,7 +4413,7 @@ struct CudaRsqrtGradFunctor : public BaseActivationFunctor<T> {
const T arg_out) const {
MPType dout = static_cast<MPType>(arg_dout);
MPType out = static_cast<MPType>(arg_out);
return static_cast<T>(minus_one_half * dout * out * out * out);
return static_cast<T>(minus_one_half * dout * (out * out * out));
}

static constexpr ActBwdOpFwdDeps FwdDeps() {
Expand Down
Loading