Skip to content

[Megatron-LM] Fix torch_dist ckpt format saving #83

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
wants to merge 5 commits into
base: rocm_dev
Choose a base branch
from
Open

[Megatron-LM] Fix torch_dist ckpt format saving #83

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
cf8c55e
Update filesystem_async.py to add serialization format argument
vidushi8 Jun 12, 2025
5b2915e
Update to conditionally check on serialization_format param
vidushi8 Jun 19, 2025
a97c560
Update filesystem_async.py
vidushi8 Jun 19, 2025
2dd44e3
minor update
vidushi8 Jun 19, 2025
ad2741e
address comment
vidushi8 Jul 7, 2025
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
2 changes: 1 addition & 1 deletion examples/deepseek_v2/train_deepseekv2.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
#
# See LICENSE for license information.
#################################################################################
set -e
# set -e
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@vidushi8 I think there is a better way to do this. You mentioned that there is a command to detect for multi-node that would fail on single node. You can change it to the following:

<command to detect multi-node> || true

This way, it won't fail and break the run.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Moved the multi-node NCCL_IB_HCA logic under the if condition NNODES>1


TOKENIZER_MODEL="deepseek-ai/DeepSeek-V2-Lite"

Expand Down
2 changes: 1 addition & 1 deletion examples/deepseek_v3/train_deepseekv3.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@
# See LICENSE for license information.
#################################################################################

set -e
# set -e

# exp
EXPERIMENT="deepseek_v3"
Expand Down
7 changes: 6 additions & 1 deletion megatron/core/dist_checkpointing/strategies/filesystem_async.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,12 @@
# in torch_dist format. This is a workaround to fix the issue.
from torch.distributed.checkpoint.filesystem import _StorageWriterTransforms
from functools import partial
_write_item = partial(_write_item, _StorageWriterTransforms())
import inspect
if "serialization_format" in inspect.signature(_write_item).parameters:
from torch.distributed.checkpoint.filesystem import SerializationFormat
_write_item = partial(_write_item, _StorageWriterTransforms(), serialization_format=SerializationFormat.TORCH_SAVE)
else:
_write_item = partial(_write_item, _StorageWriterTransforms())
except ImportError:
pass

Expand Down