Skip to content

Do not shard unused parameters #1773

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
kctezcan wants to merge 1 commit into
base: develop
Choose a base branch
from
Open

Do not shard unused parameters #1773

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
18 changes: 11 additions & 7 deletions src/weathergen/model/model_interface.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,13 +189,17 @@ def load_model(cf, model, device, run_id: str, mini_epoch=-1):
maybe_sharded_sd = {}
for param_name, full_tensor in params.items():
sharded_meta_param = meta_sharded_sd.get(param_name)
sharded_tensor = distribute_tensor(
full_tensor,
sharded_meta_param.device_mesh,
sharded_meta_param.placements,
)
# maybe_sharded_sd[param_name.replace("module.", "")] = nn.Parameter(sharded_tensor)
maybe_sharded_sd[param_name] = torch.nn.Parameter(sharded_tensor)
if sharded_meta_param is None:
logger.info(f"Sharding meta parameters is None for: {param_name}")
Copy link
Collaborator

@clessig clessig Feb 6, 2026

Choose a reason for hiding this comment

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

Is it correct that sharded_meta_param is None means that this is a parameter in the checkpoint that is not present in the current model?

maybe_sharded_sd[param_name] = torch.nn.Parameter(full_tensor)
else:
sharded_tensor = distribute_tensor(
full_tensor,
sharded_meta_param.device_mesh,
sharded_meta_param.placements,
)
# maybe_sharded_sd[param_name.replace("module.", "")] = nn.Parameter(sharded_tensor)
maybe_sharded_sd[param_name] = torch.nn.Parameter(sharded_tensor)
Copy link
Collaborator

@clessig clessig Feb 6, 2026

Choose a reason for hiding this comment

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

Can we please remove the line below.

# choose `assign=True` for sharded model since we cannot call `copy_` on meta tensor
mkeys, ukeys = model.load_state_dict(maybe_sharded_sd, strict=False, assign=True)

Expand Down
Loading