Skip to content

group the tensors to serialize in linear rather than quadratic time #3555

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
copybara-service merged 1 commit into from
Aug 31, 2026
+7 −7
Merged

group the tensors to serialize in linear rather than quadratic time #3555

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
14 changes: 7 additions & 7 deletions checkpoint/orbax/checkpoint/_src/handlers/base_pytree_checkpoint_handler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,13 +151,13 @@ def _group_value(
if handler not in grouped:
grouped[handler] = BatchRequest(handler, [], [], [], [])
request = grouped[handler]
grouped[handler] = dataclasses.replace(
request,
keys=request.keys + [tuple_key],
values=request.values + [value],
infos=request.infos + [info],
args=request.args + [arg],
)
# `BatchRequest` is mutable, so the entry is appended in place: rebuilding
# it per leaf copies all four lists and makes this loop quadratic in the
# number of tensors, which dominates for a tree with many small leaves.
request.keys.append(tuple_key)
request.values.append(value)
request.infos.append(info)
request.args.append(arg)

jax.tree_util.tree_map_with_path(
_group_value,
Expand Down
Loading