Skip to content

chg: passing through trust_remote_code #778

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
ChenhanYu merged 2 commits into from
Jan 14, 2026
Merged

chg: passing through trust_remote_code #778

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f17dc2f
chg: passing through trust_remote_code
ChenhanYu Jan 14, 2026
4de7932
chg: update
ChenhanYu Jan 14, 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
2 changes: 1 addition & 1 deletion modelopt/torch/export/plugins/megatron_importer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,7 +75,7 @@ def __init__(
workspace_dir: str | None = None,
dtype=torch.bfloat16,
dequantize: bool = True,
trust_remote_code: bool = False,
trust_remote_code: bool = True,
Copy link
Collaborator

@kevalmorabia97 kevalmorabia97 Jan 14, 2026

Choose a reason for hiding this comment

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

Why do we set default to True?

Copy link
Contributor

@meenchen meenchen Jan 14, 2026 •
edited
Loading

Choose a reason for hiding this comment

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

I have the same question here. Is this an intended change?

Copy link
Collaborator Author

@ChenhanYu ChenhanYu Jan 14, 2026

Choose a reason for hiding this comment

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

It was True in the past. I set it to False in #769 but found that Nemotron needs this to be True. So I decide to pass through and continue using the previous default value.

Copy link
Collaborator

@kevalmorabia97 kevalmorabia97 Jan 15, 2026

Choose a reason for hiding this comment

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

Why dont we pass explicitly whenever we need to set to True? I'm afraid Security team might have concerns if we set default to True

verbose: bool = False,
moe_router_dtype: torch.dtype | None = None,
):
Expand Down
9 changes: 8 additions & 1 deletion modelopt/torch/export/unified_export_megatron.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -137,7 +137,7 @@ def __init__(
pretrained_model_name_or_path: str | os.PathLike | None = None,
export_extra_modules: bool = False,
dtype=torch.bfloat16,
trust_remote_code: bool = True,
trust_remote_code: bool = False,
moe_router_dtype: torch.dtype | None = None,
):
"""Create a GPTModel exporter instance."""
Expand Down Expand Up @@ -1205,6 +1205,7 @@ def export_mcore_gpt_to_hf(
export_extra_modules: bool = False,
dtype: torch.dtype = torch.bfloat16,
export_dir: Path | str = tempfile.gettempdir(),
trust_remote_code: bool = False,
moe_router_dtype: torch.dtype | None = None,
):
"""Export Megatron Core GPTModel to unified checkpoint and save to export_dir.
Expand All @@ -1225,6 +1226,7 @@ def export_mcore_gpt_to_hf(
pretrained_model_name_or_path,
export_extra_modules=export_extra_modules,
dtype=dtype,
trust_remote_code=trust_remote_code,
moe_router_dtype=moe_router_dtype,
)
exporter.save_pretrained(export_dir, pretrained_model_name_or_path)
Expand All @@ -1235,6 +1237,7 @@ def import_mcore_gpt_from_hf(
pretrained_model_path: str,
workspace_dir: str | None = None,
dtype: torch.dtype = torch.bfloat16,
trust_remote_code: bool = False,
moe_router_dtype: torch.dtype | None = None,
):
"""Import GPTModel state_dict from supported HuggingFace pretrained model path.
Expand All @@ -1243,13 +1246,17 @@ def import_mcore_gpt_from_hf(
model: The Megatron Core GPTModel instance.
pretrained_model_path: A path to a *directory* containing model weights saved using
[`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
workspace_dir: The directory to save the workspace.
dtype: The weights data type to import.
trust_remote_code: If True, this allows importing from a wider range of sources.
moe_router_dtype: The data type to import the moe router weights.
"""
importer = GPTModelImporter(
model,
pretrained_model_path,
workspace_dir=workspace_dir,
dtype=dtype,
trust_remote_code=trust_remote_code,
moe_router_dtype=moe_router_dtype,
)
importer._import_state_dict()