Skip to content

Fix split_state_dict function for MoE models #3667

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
yaox12 merged 3 commits into from
Mar 10, 2026
Merged

Fix split_state_dict function for MoE models #3667

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
910fa29
fix split_state_dict function
eternally-z Mar 3, 2026
c65fd5c
minor fix
eternally-z Mar 3, 2026
e867137
Fix lint error
kunlunl Mar 3, 2026
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
21 changes: 13 additions & 8 deletions megatron/core/optimizer/optimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1161,20 +1161,25 @@ def _split_state_dict(self, state_dict):
state_dicts = [None] * len(self.chained_optimizers)
if state_dict is not None:
if len(self.model_chunks) == 1:
state_dicts[0] = state_dict
# When there is only one global model chunk, all sub-optimizers
# (e.g., dense and MoE parts) use the same model state dict.
state_dicts = [state_dict] * len(self.chained_optimizers)
Comment thread
kunlunl marked this conversation as resolved.
Resolved
else:
# Split state_dict if needed
# Split state_dict by chunk identity
prefix = "model" if "model0" in state_dict.keys() else "model_"
offset = 0
chunk_id_to_global_idx = {id(chunk): idx for idx, chunk in enumerate(self.model_chunks)}
Comment thread
kunlunl marked this conversation as resolved.
Resolved
for optimizer_idx, optimizer in enumerate(self.chained_optimizers):
if hasattr(optimizer, "model_chunks"):
d = {}
for chunk_idx in range(len(optimizer.model_chunks)):
for chunk_idx, model_chunk in enumerate(optimizer.model_chunks):
assert id(model_chunk) in chunk_id_to_global_idx, (
"Sub-optimizer model chunk was not found in chained optimizer model chunks"
)
global_idx = chunk_id_to_global_idx[id(model_chunk)]
assert (
f"{prefix}{offset}" in state_dict
), f"Wrong state_dict format, cannot find '{prefix}{offset}'"
d[f"{prefix}{chunk_idx}"] = state_dict[f"{prefix}{offset}"]
offset += 1
f"{prefix}{global_idx}" in state_dict
), f"Wrong state_dict format, cannot find '{prefix}{global_idx}'"
d[f"{prefix}{chunk_idx}"] = state_dict[f"{prefix}{global_idx}"]
if len(d) > 0:
state_dicts[optimizer_idx] = d
return state_dicts
Expand Down