Skip to content

Add data-type promotion to gelu_backward. #7090

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
May 22, 2024
Merged

Add data-type promotion to gelu_backward. #7090

Show file tree
Hide file tree
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
16 changes: 16 additions & 0 deletions test/test_operations.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1982,6 +1982,22 @@ def foo(x):
for dtype in test_dtypes:
test(dtype)

def test_gelu_backward_different_types(self):

def foo(grad, inp):
return torch.ops.aten.gelu_backward.default(grad, inp)

grad = torch.rand(10, 10, dtype=torch.bfloat16)
inp = torch.rand(10, 10)

Xgrad = grad.to(xm.xla_device())
Xinp = inp.to(xm.xla_device())

r = foo(grad, inp)
Xr = foo(Xgrad, Xinp)

self.assertEqual(r, Xr.cpu())


class MNISTComparator(nn.Module):

Expand Down
4 changes: 3 additions & 1 deletion torch_xla/csrc/aten_xla_type.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1504,8 +1504,10 @@ at::Tensor XLANativeFunctions::gelu_backward(const at::Tensor& grad,
const at::Tensor& self,
c10::string_view approximate) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
at::ScalarType result_type = at::result_type(grad, self);
return bridge::AtenFromXlaTensor(tensor_methods::gelu_backward(
bridge::GetXlaTensor(grad), bridge::GetXlaTensor(self), approximate));
bridge::GetXlaTensor(grad.to(result_type)),
bridge::GetXlaTensor(self.to(result_type)), approximate));
}

at::Tensor XLANativeFunctions::hardtanh(const at::Tensor& self,
Expand Down