Skip to content
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

[prim] temp solve of paddlescience ldc，cylinder OOM #55116

Merged
merged 9 commits into from
Jul 6, 2023
Merged

[prim] temp solve of paddlescience ldc，cylinder OOM #55116

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
789b85e
modify eular_beam
xiaoguoguo626807 Jun 19, 2023
ee00174
modify matmul infermeta
xiaoguoguo626807 Jun 20, 2023
8d52bc7
add test
xiaoguoguo626807 Jun 20, 2023
54abc6a
modify timeout
xiaoguoguo626807 Jun 20, 2023
6715bbb
fix conflict
xiaoguoguo626807 Jun 26, 2023
05d8d26
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
xiaoguoguo626807 Jun 27, 2023
13be2c5
modify mutiply_double nullptr
xiaoguoguo626807 Jun 27, 2023
8caf474
modify tanh_triple_gradnode create
xiaoguoguo626807 Jun 28, 2023
6f14b78
tanh_triple
xiaoguoguo626807 Jul 4, 2023
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
23 changes: 12 additions & 11 deletions paddle/fluid/prim/api/composite_backward/composite_double_backward_api.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -412,7 +412,8 @@ void multiply_double_grad(const Tensor& x,
}

} else {
x_grad = nullptr;
auto dx = full<T>(phi::vectorize(x.dims()), 0.0, x.dtype());
set_output<T>(dx, x_grad);
}
}
if (y_grad) {
Expand All @@ -433,22 +434,22 @@ void multiply_double_grad(const Tensor& x,
set_output<T>(dy, y_grad);
}
} else {
y_grad = nullptr;
auto dy = full<T>(phi::vectorize(y.dims()), 0.0, y.dtype());
set_output<T>(dy, y_grad);
}
}
if (grad_out_grad) {
Tensor ddout;
if (grad_x_grad && grad_y_grad) {
auto ddout = grad_x_grad.get() * y + grad_y_grad.get() * x;
set_output<T>(ddout, grad_out_grad);
ddout = grad_x_grad.get() * y + grad_y_grad.get() * x;
} else if (grad_x_grad) {
auto ddout = grad_x_grad.get() * y;
set_output<T>(ddout, grad_out_grad);
ddout = grad_x_grad.get() * y;
} else if (grad_y_grad) {
auto ddout = grad_y_grad.get() * x;
set_output<T>(ddout, grad_out_grad);
ddout = grad_y_grad.get() * x;
} else {
grad_out_grad = nullptr;
ddout = full<T>(phi::vectorize(grad_out.dims()), 0.0, grad_out.dtype());
}
set_output<T>(ddout, grad_out_grad);
}
}

Expand All @@ -461,10 +462,10 @@ void add_double_grad(const Tensor& y,
Tensor* grad_out_grad) {
if (grad_out_grad) {
// ddout = ddx + ddy
Tensor ddout = full<T>(phi::vectorize(grad_out.dims()), 0.0, y.dtype());
if (!grad_x_grad && !grad_y_grad) {
grad_out_grad = nullptr;
set_output<T>(ddout, grad_out_grad);
} else {
Tensor ddout = full<T>(phi::vectorize(grad_out.dims()), 0.0, y.dtype());
if (grad_x_grad) {
ddout = ddout + grad_x_grad.get();
}
Expand Down
8 changes: 5 additions & 3 deletions python/paddle/fluid/backward.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2489,9 +2489,11 @@ def calc_gradient_helper(
block, targets, inputs, block_no_grad_set, op_path_dict
)

# only for composite to add grad_op input,
# tmp_targets includes targets and other outputs
# of the same forward op who create targets
# only for composite to add grad_var of the last forward op
# who has more than one output, but targets only has one,
# so targets_gradients only add one grad_var,
# eg: op1 -> op2 -> var1 / var2 targets = var1,
# targets_gradients = var1_grad, need to add var2_grad here.
tmp_targets = targets

if core._is_bwd_prim_enabled():
Expand Down