Skip to content

[BIG tensor]fix overflow problem in addmm caused by the cublas api #74289

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
Jul 30, 2025
Merged

[BIG tensor]fix overflow problem in addmm caused by the cublas api #74289

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

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 13 additions & 7 deletions paddle/phi/kernels/impl/addmm_grad_kernel_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,12 @@ void AddmmGradKernel(const Context& dev_ctx,
}
using MPType = typename phi::dtype::MPTypeTrait<T>::Type;
bool is_float16_or_bfloat16 = false;
bool is_big_tensor = false;
if (input.numel() * input.dims()[1] > std::numeric_limits<int>::max() ||
x.numel() > std::numeric_limits<int>::max() ||
y.numel() * y.dims()[1] > std::numeric_limits<int>::max()) {
is_big_tensor = true;
}
if (std::is_same<T, phi::dtype::float16>::value ||
std::is_same<T, phi::dtype::bfloat16>::value) {
is_float16_or_bfloat16 = true;
Expand Down Expand Up @@ -126,7 +132,7 @@ void AddmmGradKernel(const Context& dev_ctx,
Array2(input_grad->dims()[0], input_grad->dims()[1]);

if (row_compress && col_compress) {
if (!is_float16_or_bfloat16) {
if (!is_float16_or_bfloat16 && !is_big_tensor) {
eigen_dinput.device(place) =
eigen_dout.sum().eval().reshape(eigen_dinput_shape);
} else {
Expand All @@ -137,7 +143,7 @@ void AddmmGradKernel(const Context& dev_ctx,
.template cast<T>();
}
} else if (row_compress) {
if (!is_float16_or_bfloat16) {
if (!is_float16_or_bfloat16 && !is_big_tensor) {
eigen_dinput.device(place) =
eigen_dout.sum(Array1(0)).eval().reshape(eigen_dinput_shape);
} else {
Expand All @@ -148,7 +154,7 @@ void AddmmGradKernel(const Context& dev_ctx,
.template cast<T>();
}
} else if (col_compress) {
if (!is_float16_or_bfloat16) {
if (!is_float16_or_bfloat16 && !is_big_tensor) {
eigen_dinput.device(place) =
eigen_dout.sum(Array1(1)).eval().reshape(eigen_dinput_shape);
} else {
Expand All @@ -160,7 +166,7 @@ void AddmmGradKernel(const Context& dev_ctx,
}
} else {
// The VCOPY does not support the float16, bfloat16
if (!is_float16_or_bfloat16) {
if (!is_float16_or_bfloat16 && !is_big_tensor) {
mt_blas.VCOPY(
total_elems, out_grad.data<MPType>(), input_grad->data<MPType>());
} else {
Expand All @@ -172,7 +178,7 @@ void AddmmGradKernel(const Context& dev_ctx,
}

// The SCAL does not support the float16, bfloat16
if (!is_float16_or_bfloat16) {
if (!is_float16_or_bfloat16 && !is_big_tensor) {
mt_blas.SCAL(total_elems, beta, input_grad->data<MPType>());
} else {
phi::funcs::ForRange<Context> for_range(dev_ctx, total_elems);
Expand Down Expand Up @@ -202,7 +208,7 @@ void AddmmGradKernel(const Context& dev_ctx,
total_elems = x.dims()[0] * x.dims()[1];
// x_grad = out_grad * y'. x_grad: M x K, out_grad : M x N, y : K x N
blas.MatMul(out_grad, false, y, true, x_grad);
if (!is_float16_or_bfloat16) {
if (!is_float16_or_bfloat16 && !is_big_tensor) {
mt_blas.SCAL(total_elems, alpha, x_grad->data<MPType>());
} else {
phi::funcs::ForRange<Context> for_range(dev_ctx, total_elems);
Expand All @@ -216,7 +222,7 @@ void AddmmGradKernel(const Context& dev_ctx,
total_elems = x.dims()[1] * y.dims()[1];
// y_grad = x' * out_grad. y_grad K x N, out_grad : M x N, x : M x K
blas.MatMul(x, true, out_grad, false, y_grad);
if (!is_float16_or_bfloat16) {
if (!is_float16_or_bfloat16 && !is_big_tensor) {
mt_blas.SCAL(total_elems, alpha, y_grad->data<MPType>());
} else {
phi::funcs::ForRange<Context> for_range(dev_ctx, total_elems);
Expand Down