Skip to content

fix: resolve GPU memory leak in pipeline parallel training #148

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
kilinchange merged 1 commit into from
Apr 24, 2026
Merged

fix: resolve GPU memory leak in pipeline parallel training #148

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
7 changes: 7 additions & 0 deletions infini_train/src/nn/parallel/pp/pipeline_schedule.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,13 @@ std::vector<std::shared_ptr<Tensor>> PipelineSchedule::ReceiveFromPrev(int peer_
// FIXME(jym): The data type between stages is not float32, which will cause a crash
auto tensor = std::make_shared<Tensor>(shapes[i], DataType::kFLOAT32, stage_->device());
tensor->set_requires_grad(true);

// Mark as non-leaf to prevent the autograd engine from creating an AccumulateGrad
// for this tensor. Otherwise, IRecv's next_functions_ would hold AccumulateGrad,
// which holds a shared_ptr back to this tensor, forming a reference cycle:
// tensor -> grad_fn_(IRecv) -> next_functions_ -> AccumulateGrad -> tensor
// This cycle prevents the autograd graph from being released after backward.
tensor->set_is_leaf(false);
recv_tensors.push_back(tensor);
}

Expand Down
Loading