Skip to content

Run torch.special.logit in reduced precision for bf16/f16 inputs #2156

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

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
Open

Run torch.special.logit in reduced precision for bf16/f16 inputs #2156

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
9682a82
Cast intermediate tensor to bfloat16
Silv3S Oct 7, 2025
0ff60c8
syntax
Silv3S Oct 8, 2025
c788ff1
Modify files in install_xpu_headers only if they changes
Oct 8, 2025
2385a78
Merge remote-tracking branch 'origin/main' into dev/ssiwek/logit_bf16
Silv3S Oct 9, 2025
3996d71
Calculate in declared precision
Silv3S Oct 9, 2025
e3de893
Merge branch 'main' into dev/ssiwek/logit_bf16
Silv3S Oct 17, 2025
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
31 changes: 12 additions & 19 deletions src/ATen/native/xpu/sycl/UnarySpecialOpsKernels.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,31 +130,24 @@ void exp2_kernel(TensorIteratorBase& iter) {
}

template <typename scalar_t>
struct Logit0Functor {
using T_ACC = acc_type_device<scalar_t, c10::DeviceType::XPU>;
struct LogitFunctor {
scalar_t operator()(scalar_t x) const {
const T_ACC x_acc = static_cast<T_ACC>(x);
// suppress compiler optimization on data type promotion.
volatile T_ACC res = std::log(x_acc / (T_ACC(1) - x_acc));
return res;
return std::log(x / (1 - x));
}
};

template <typename scalar_t>
struct Logit1Functor {
struct LogitEpsFunctor {
using T_ACC = acc_type_device<scalar_t, c10::DeviceType::XPU>;
scalar_t operator()(scalar_t x) const {
const T_ACC x_acc = static_cast<T_ACC>(x);
T_ACC z = x_acc < lo_ ? lo_ : (x_acc > hi_ ? hi_ : x_acc);
// suppress compiler optimization on data type promotion.
volatile T_ACC res = std::log(z / (T_ACC(1) - z));
return res;
scalar_t x_clamped = x < low_ ? low_ : (x > high_ ? high_ : x);
return std::log(x_clamped / (1 - x_clamped));
}
Logit1Functor(const T_ACC lo, const T_ACC hi) : lo_(lo), hi_(hi) {}
LogitEpsFunctor(const T_ACC low, const T_ACC high) : low_(low), high_(high) {}

private:
T_ACC lo_;
T_ACC hi_;
scalar_t low_;
scalar_t high_;
};

void logit_kernel(TensorIteratorBase& iter, const Scalar& eps_scalar) {
Expand All @@ -167,11 +160,11 @@ void logit_kernel(TensorIteratorBase& iter, const Scalar& eps_scalar) {
using T_ACC = acc_type_device<scalar_t, c10::DeviceType::XPU>;
const T_ACC eps = eps_scalar.to<T_ACC>();
if (eps < T_ACC(0)) {
gpu_kernel(iter, Logit0Functor<scalar_t>());
gpu_kernel(iter, LogitFunctor<scalar_t>());
} else {
const T_ACC lo = eps;
const T_ACC hi = T_ACC(1) - eps;
gpu_kernel(iter, Logit1Functor<scalar_t>(lo, hi));
const T_ACC low = eps;
const T_ACC high = T_ACC(1) - eps;
gpu_kernel(iter, LogitEpsFunctor<scalar_t>(low, high));
}
});
}
Expand Down