Skip to content

Add manual release of input tensors and their grads in CheckpointFunction #3267

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

Draft
Begunner wants to merge 1 commit into
base: main
Choose a base branch
from
+9 −1
Draft

Add manual release of input tensors and their grads in CheckpointFunction #3267

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
10 changes: 9 additions & 1 deletion megatron/core/tensor_parallel/random.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -564,7 +564,15 @@ def backward(ctx, *args):
*filter(lambda x: torch.is_tensor(x[0]) and x[0].requires_grad, zip(outputs, args))
)
torch.autograd.backward(outputs, args)
grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else inp for inp in detached_inputs)
grads = tuple(inp.grad.clone() if isinstance(inp, torch.Tensor) and inp.grad is not None else inp.grad if isinstance(inp, torch.Tensor) else inp for inp in detached_inputs)
cur_stream = torch.cuda.current_stream()
for t in detached_inputs:
if isinstance(t, torch.Tensor) and t.requires_grad:
t.record_stream(cur_stream)
t.untyped_storage().resize_(0)
if t.grad is not None:
t.grad.record_stream(cur_stream)
t.grad.untyped_storage().resize_(0)

_unset_checkpointing()
return (None, None) + grads
Expand Down