Skip to content

Add finalize_model_grads_with_empty_cache #1133

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
zhuzilin merged 1 commit into from
Dec 16, 2025
Merged

Add finalize_model_grads_with_empty_cache #1133

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
12 changes: 11 additions & 1 deletion slime/backends/megatron_utils/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -500,6 +500,16 @@ def should_disable_forward_pre_hook(args: Namespace) -> bool:
return args.use_distributed_optimizer and args.overlap_param_gather


def finalize_model_grads_with_empty_cache(*args, **kwargs):
# trigger empty cache when there are less than 10% free memory before the final reduce scatter.
# TODO: this is an ad-hoc method and we should figure out why the oom happens in the first place.
device = torch.cuda.current_device()
free, total = torch.cuda.mem_get_info(device)
if free / total < 0.1:
clear_memory()
return finalize_model_grads(*args, **kwargs)


def train(
rollout_id: int,
model: Sequence[DDP],
Expand Down Expand Up @@ -550,7 +560,7 @@ def train(
config.param_sync_func = [model_chunk.start_param_sync for model_chunk in model]
if len(model) == 1:
config.param_sync_func = config.param_sync_func[0]
config.finalize_model_grads_func = finalize_model_grads
config.finalize_model_grads_func = finalize_model_grads_with_empty_cache

pre_hook_enabled = False

Expand Down