Skip to content

[optim] run plan/metadata coordination over a gloo group #62

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

Open
yueming-yuan wants to merge 1 commit into
base: miles-main
Choose a base branch
from
Open

[optim] run plan/metadata coordination over a gloo group #62

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
17 changes: 16 additions & 1 deletion megatron/core/dist_checkpointing/strategies/torch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,6 +109,20 @@ def register_default_torch_strategies():

logger = getLogger(__name__)

_COORDINATION_PROCESS_GROUP = None


def _get_coordination_process_group():
"""Gloo group for checkpoint plan/metadata object collectives. On an
NCCL-only default group these would be staged through GPU and leave
permanent per-peer NCCL channel buffers on the coordinator rank."""
global _COORDINATION_PROCESS_GROUP
if not torch.distributed.is_initialized():
return None
if _COORDINATION_PROCESS_GROUP is None:
_COORDINATION_PROCESS_GROUP = torch.distributed.new_group(backend="gloo")
return _COORDINATION_PROCESS_GROUP


def flatten_state_dict(
state_dict: ShardedStateDict,
Expand Down Expand Up @@ -701,7 +715,7 @@ def async_save(
) = save_state_dict_async_plan(
pyt_state_dict,
writer,
None,
_get_coordination_process_group(),
coordinator,
planner=MCoreSavePlanner(
dedup_replicated_tensors=not self.keep_only_main_replica, flatten_state_dict=False
Expand Down Expand Up @@ -805,6 +819,7 @@ def load(self, sharded_state_dict: ShardedStateDict, checkpoint_dir: Path) -> St
checkpoint.load_state_dict(
pyt_state_dict,
fsr,
process_group=_get_coordination_process_group(),
planner=MCoreLoadPlanner(
shapes_validation_sharded_tensors=flexible_shape_sharded_tensors,
allow_shape_mismatch_sharded_tensors=allow_shape_mismatch_sharded_tensors,
Expand Down