Skip to content

Avoid mutating caller list in split_by_mesh_axis and validate per-array shardings in concatenate_by_mesh_axis. #367

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
copybara-service wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Avoid mutating caller list in split_by_mesh_axis and validate per-array shardings in concatenate_by_mesh_axis. #367

Show file tree
Hide file tree
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
26 changes: 25 additions & 1 deletion pathwaysutils/experimental/concatenate_by_mesh_axis.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,10 +68,21 @@ def _get_named_sharding(array: jax.Array) -> jax.sharding.NamedSharding:
if not flattened_arrays[0]:
return array_trees[0]

# Extract the shared mesh from each PyTree (from an arbitrary array in each).
# Extract the shared mesh from each PyTree and verify all leaves in a PyTree
# share that mesh.
meshes_to_concatenate = [
_get_named_sharding(array).mesh for array in input_flat_arrays[0]
]
for tree_arrays, expected_mesh in zip(
flattened_arrays, meshes_to_concatenate
):
for array in tree_arrays:
sharding = _get_named_sharding(array)
if sharding.mesh != expected_mesh:
raise ValueError(
"All arrays in each PyTree must have the same mesh. Got"
f" {sharding.mesh} vs. {expected_mesh}."
)

# Validate that the meshes are compatible.
reference_mesh = meshes_to_concatenate[0]
Expand Down Expand Up @@ -117,6 +128,19 @@ def _get_output_sharding(
) -> jax.sharding.NamedSharding:
reference_sharding = _get_named_sharding(arrays[0])
reference_spec = reference_sharding.spec
for array in arrays[1:]:
sharding = _get_named_sharding(array)
if sharding.spec != reference_spec:
raise ValueError(
"Corresponding arrays across PyTrees must have the same"
f" PartitionSpec. Got {sharding.spec} vs. {reference_spec}."
)
if sharding.memory_kind != reference_sharding.memory_kind:
raise ValueError(
"Corresponding arrays across PyTrees must have the same"
f" memory_kind. Got {sharding.memory_kind} vs."
f" {reference_sharding.memory_kind}."
)
return jax.sharding.NamedSharding(
concatenated_mesh,
reference_spec,
Expand Down
4 changes: 2 additions & 2 deletions pathwaysutils/experimental/split_by_mesh_axis.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,7 +129,7 @@ def split_by_mesh_axis(
range(axis_size, mesh.axis_sizes[mesh_axis_idx] + 1, axis_size)
)
else:
mesh_axis_sections = mesh_axis_indices_or_sections
mesh_axis_sections = list(mesh_axis_indices_or_sections)
for i, boundary in enumerate(mesh_axis_sections):
if boundary <= 0 or boundary >= mesh.axis_sizes[mesh_axis_idx]:
raise ValueError(
Expand All @@ -142,7 +142,7 @@ def split_by_mesh_axis(
"Mesh axis sections must be monotonically increasing, but got"
f" {mesh_axis_sections=}."
)
mesh_axis_sections += [mesh.axis_sizes[mesh_axis_idx]] # pyrefly: ignore[unsupported-operation]
mesh_axis_sections += [mesh.axis_sizes[mesh_axis_idx]]

submeshes = []
axis_boundary_start = 0
Expand Down
Loading