Skip to content

Delete output tensor early #4742

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
ericharper merged 1 commit into from
May 29, 2026
Merged

Delete output tensor early #4742

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
8 changes: 8 additions & 0 deletions megatron/core/pipeline_parallel/schedules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -708,6 +708,13 @@ def forward_backward_no_pipelining(
total_num_tokens += num_tokens
if not forward_only:
backward_step(input_tensor, output_tensor, output_tensor_grad, config)
# Release the autograd graph head before the next forward_step.
# Without this, the previous microbatch's output_tensor stays
# live until the next iteration rebinds the variable, deferring
# autograd-node teardown onto the next forward's dispatch path
# and triggering PyTorch's "AccumulateGrad node's stream does
# not match" warning. See issue #4124.
del output_tensor
# Run computation for last microbatch out of context handler (want to
# synchronize gradients).
output_tensor, num_tokens = forward_step(
Expand All @@ -730,6 +737,7 @@ def forward_backward_no_pipelining(

if not forward_only:
backward_step(input_tensor, output_tensor, output_tensor_grad, config)
del output_tensor

if config.finalize_model_grads_func is not None and not forward_only:
# Finalize model grads (perform full grad all-reduce / reduce-scatter for
Expand Down
Loading