Skip to content

Fix BFloat16 conversion error in eager checkpoint loading #3347

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
phu0ngng wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix BFloat16 conversion error in eager checkpoint loading #3347

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7bf11b5
fix bf16 type conversion
phu0ngng Mar 7, 2026
e8faa02
Merge branch 'main' of github.com:AI-Hypercomputer/maxtext into fix_c…
phu0ngng Mar 7, 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
9 changes: 7 additions & 2 deletions src/maxtext/checkpoint_conversion/to_maxtext.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -595,9 +595,14 @@ def main(
hf_config_obj = AutoConfig.from_pretrained(model_id, token=hf_token, revision=revision)
hf_model = get_hf_model(model_id, token=hf_token, revision=revision)
hf_state_dict_numpy = hf_model.state_dict()
# Convert all to numpy immediately in eager mode
# Convert all to numpy immediately in eager mode.
# torch.Tensor.numpy() does not support bfloat16, so cast to float32 first.
import torch # pylint: disable=g-import-not-at-top
Copy link
Collaborator

@shuningjin shuningjin Mar 8, 2026 •
edited
Loading

Choose a reason for hiding this comment

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

move this import to the top?

Copy link
Contributor Author

@phu0ngng phu0ngng Mar 9, 2026 •
edited
Loading

Choose a reason for hiding this comment

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

We can, but then it will require torch in the lazy loading path as well, even though torch is not needed there. I think we should go with the current implementation to avoid torch requirements for lazy load.

Copy link
Contributor Author

@phu0ngng phu0ngng Mar 11, 2026

Choose a reason for hiding this comment

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

@shuningjin what do you think?

Copy link
Contributor Author

@phu0ngng phu0ngng Mar 13, 2026

Choose a reason for hiding this comment

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

@shuningjin, friendly reminder.

for k, v in hf_state_dict_numpy.items():
hf_state_dict_numpy[k] = v.numpy()
if v.dtype == torch.bfloat16:
hf_state_dict_numpy[k] = v.float().numpy()
else:
hf_state_dict_numpy[k] = v.numpy()
del hf_model
max_logging.log("HuggingFace model loaded and converted to NumPy.")
print_ram_usage("After full HF model load")
Expand Down
Loading