Skip to content
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

fix serialization of partitioned states in optax #852

Merged
merged 1 commit into from
Jan 9, 2025
Merged

fix serialization of partitioned states in optax #852

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
8 changes: 5 additions & 3 deletions src/levanter/tensorstore_serialization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,7 +95,7 @@ def path_from_key_path(key_path):

def _sharding_from_leaf(leaf, axis_mapping, mesh) -> Optional[jax.sharding.Sharding]:
if is_named_array(leaf):
if leaf.array is None:
if not is_jax_array_like(leaf.array):
return None
return hax.partitioning.sharding_for_axis(leaf.axes, axis_mapping, mesh)
elif hasattr(leaf, "sharding") and getattr(leaf, "sharding") is not None:
Expand Down Expand Up @@ -140,11 +140,11 @@ def tree_deserialize_leaves_tensorstore(
manager = array_ser.GlobalAsyncCheckpointManager()

shardings: PyTree[Optional[Sharding]] = jtu.tree_map(
partial(_sharding_from_leaf, axis_mapping=axis_mapping, mesh=mesh), pytree, is_leaf=is_named_array
partial(_sharding_from_leaf, axis_mapping=axis_mapping, mesh=mesh), pytree, is_leaf=_is_named_or_none
)

# TODO: support ShapeDtypeStructs that are not NamedArrays
leaf_key_paths = jax_utils.leaf_key_paths(shardings, is_leaf=is_named_array)
leaf_key_paths = jax_utils.leaf_key_paths(shardings, is_leaf=_is_named_or_none)
paths = _fs_paths_from_key_paths(checkpoint_dir, leaf_key_paths)
paths = jtu.tree_leaves(paths, is_leaf=lambda x: x is None)

Expand All @@ -157,6 +157,8 @@ def tree_deserialize_leaves_tensorstore(
real_leaves = [x for x in shardings_leaves if x is not None]
real_paths = [paths[i] for i in real_indices]

assert len(real_leaves) == len(real_paths), f"{len(real_leaves)} != {len(real_paths)}"

deser_leaves = manager.deserialize_with_paths(shardings=real_leaves, paths=real_paths)
# now we need to recreate the original structure

Expand Down
Loading