Skip to content

[Dy2St] Fix missing Tensor name when trans to contiguous #62896

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
SigureMo merged 2 commits into from
Mar 22, 2024
Merged

[Dy2St] Fix missing Tensor name when trans to contiguous #62896

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
10aac83
[Dy2St] Fix missing Tensor name in trans to contiguous
SigureMo Mar 20, 2024
8465d02
remove vlog(0)
SigureMo Mar 20, 2024
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
2 changes: 1 addition & 1 deletion paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1153,7 +1153,7 @@ def GenerateNodeCreationCodes(self, for_backward=False, is_inplaced=False):
for name, (ttype, pos) in forward_inputs_position_map.items():
if name in need_pre_contiguous_set:
pre_contiguous_list.append(
f"{indent}const auto& {name}_tmp = (require_any_grad && {name}.is_dense_tensor() && !std::dynamic_pointer_cast<phi::DenseTensor>({name}.impl())->meta().is_contiguous()) ? paddle::Tensor(std::make_shared<phi::DenseTensor>(paddle::experimental::Trans2Contiguous(*(std::dynamic_pointer_cast<phi::DenseTensor>({name}.impl())))), {name}.mutable_autograd_meta()) : {name};"
f"{indent}const auto& {name}_tmp = (require_any_grad && {name}.is_dense_tensor() && !std::dynamic_pointer_cast<phi::DenseTensor>({name}.impl())->meta().is_contiguous()) ? paddle::Tensor(std::make_shared<phi::DenseTensor>(paddle::experimental::Trans2Contiguous(*(std::dynamic_pointer_cast<phi::DenseTensor>({name}.impl())))), {name}.mutable_autograd_meta(), {name}.name()) : {name};"
)
self.inputs_call_list_tmp[pos] = (
self.inputs_call_list_tmp[pos] + '_tmp'
Expand Down
3 changes: 2 additions & 1 deletion paddle/fluid/eager/to_static/run_program_op_func.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,7 +124,8 @@ static std::vector<paddle::Tensor> Trans2ContiguousTensors(
std::make_shared<phi::DenseTensor>(
paddle::experimental::Trans2Contiguous(
*(std::dynamic_pointer_cast<phi::DenseTensor>(t.impl())))),
t.mutable_autograd_meta());
t.mutable_autograd_meta(),
t.name());
} else {
res.emplace_back(t);
}
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/eager/to_static/run_program_op_node.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -201,8 +201,8 @@ static void ShareTensorsIntoScopeWithName(
const std::vector<std::string> &tensor_names,
paddle::framework::Scope *scope) {
for (size_t i = 0; i < tensors.size(); ++i) {
VLOG(4) << "Share Tensor Into Scope: " << i;
auto name = tensor_names[i];
VLOG(4) << "Share Tensor Into Scope: " << name;
if (name == paddle::framework::kFakeVarName ||
name == paddle::framework::kEmptyVarName) {
continue;
Expand Down
3 changes: 2 additions & 1 deletion paddle/fluid/pybind/eager_method.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1831,7 +1831,8 @@ static PyObject* tensor__setitem_dygraph(TensorObject* self,
paddle::experimental::Trans2Contiguous(
*(std::dynamic_pointer_cast<phi::DenseTensor>(
transback_sub_tensor.impl())))),
transback_sub_tensor.mutable_autograd_meta())
transback_sub_tensor.mutable_autograd_meta(),
transback_sub_tensor.name())
: transback_sub_tensor;

grad_node = std::shared_ptr<SetValueWithTensorGradNode>(
Expand Down
8 changes: 5 additions & 3 deletions paddle/phi/api/include/tensor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -142,14 +142,16 @@ class PADDLE_API Tensor final {
explicit Tensor(const std::string& name) : name_(name) {}

/**
* @brief Construct a new Tensor object by a TensorBase pointer and
* autograd_meta
* @brief Construct a new Tensor object by a TensorBase pointer, autograd meta
* and name
*
* @param tensor_impl
* @param autograd_meta
* @param name
*/
Tensor(std::shared_ptr<phi::TensorBase> tensor_impl,
std::shared_ptr<AbstractAutogradMeta> autograd_meta);
std::shared_ptr<AbstractAutogradMeta> autograd_meta,
const std::string& name);

/* Part 2: Dimension, DataType and DataLayout methods */

Expand Down
7 changes: 5 additions & 2 deletions paddle/phi/api/lib/tensor.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,8 +53,11 @@ Tensor::Tensor(std::shared_ptr<phi::TensorBase> tensor_impl)
}

Tensor::Tensor(std::shared_ptr<phi::TensorBase> tensor_impl,
std::shared_ptr<AbstractAutogradMeta> autograd_meta)
: impl_(std::move(tensor_impl)), autograd_meta_(std::move(autograd_meta)) {
std::shared_ptr<AbstractAutogradMeta> autograd_meta,
const std::string &name)
: impl_(std::move(tensor_impl)),
autograd_meta_(std::move(autograd_meta)),
name_(name) {
Copy link
Member Author

@SigureMo SigureMo Mar 21, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

请 @zyfncg review 这里，这个构造函数是在 #57352 引入的，仅用于 trans2contiguous，这个 API 是否可以修改？如果不可以修改是新增一个更合适嘛？

Copy link
Contributor

@zyfncg zyfncg Mar 21, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

可以直接改这个API，新增接口的代价会更高一些

PADDLE_ENFORCE_NOT_NULL(
impl_,
phi::errors::InvalidArgument("TensorImpl with nullptr is not supported"));
Expand Down