Skip to content

[Dy2Stat]Add NoNeedBufferVarsInferer to reduce memory usage #34177

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 2 commits into from
Jul 16, 2021
Merged

[Dy2Stat]Add NoNeedBufferVarsInferer to reduce memory usage #34177

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d7dcb94
Add NoNeedBufferVarsInferer
Aurelius84 Jul 15, 2021
69d9ee5
fix code style
Aurelius84 Jul 15, 2021
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
24 changes: 21 additions & 3 deletions paddle/fluid/framework/executor_cache.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,6 +77,7 @@ void ParseSafeEagerDeletionSkipVars(
const std::vector<std::string> &output_var_names,
std::vector<std::string> *skip_eager_delete_vars) {
auto all_ops = program.Block(0).AllOps();
auto &op_info_map = OpInfoMap::Instance();
// NOTE: skip `shape` and `fill_constant` op created by
// fluid.backward.gradients, one forward output will generate one `shape`
// and `fill_constant`.
Expand All @@ -86,16 +87,33 @@ void ParseSafeEagerDeletionSkipVars(
// step 2: parse the necessary variable of backward op
std::unordered_set<std::string> op_outputs;
std::unordered_set<std::string> op_inputs;
std::unordered_set<std::string> no_need_buffer_ins;

for (auto i = backward_op_start_index; i < all_ops.size(); ++i) {
framework::OpDesc *op = all_ops[i];
for (const std::string &in_arg_name : op->InputArgumentNames()) {
op_inputs.emplace(in_arg_name);
// NOTE: skip NoNeedBufferVars of grad_op and GC its memory in advance.
auto &op_info = op_info_map.Get(op->Type());
auto &inferer = op_info.NoNeedBufferVarsInferer();
no_need_buffer_ins.clear();
if (inferer != nullptr) {
no_need_buffer_ins =
inferer(op->Inputs(), op->Outputs(), op->GetAttrMap());
}
for (auto &in_names : op->Inputs()) {
if (no_need_buffer_ins.count(in_names.first) == 0) {
for (auto &in_name : in_names.second) {
op_inputs.emplace(in_name);
}
} else {
VLOG(2) << op->Type() << " has no_need_buffer_in: " << in_names.first
<< " , skip it.";
}
}

for (const std::string &out_arg_name : op->OutputArgumentNames()) {
op_outputs.emplace(out_arg_name);
}
}

// For the grad op input variables, if it is not output of grad_op, it may
// be output of forward op and we should set the variables as skip_var to
// prevent it being deleted when grad op is called multiple times.
Expand Down