Skip to content

Fix how SimpleFSDP get the nD mesh #1959

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
fegin merged 7 commits into from
Oct 30, 2025
Merged

Fix how SimpleFSDP get the nD mesh #1959

Changes from 2 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
4fb86e4
Update (base update)
fegin Oct 29, 2025
07a9fc4
Update
fegin Oct 29, 2025
df5260f
Update (base update)
fegin Oct 29, 2025
87f1665
Update
fegin Oct 29, 2025
06f3ab9
Update (base update)
fegin Oct 29, 2025
ddb033f
Update
fegin Oct 29, 2025
3a296fa
Update
fegin Oct 30, 2025
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
16 changes: 2 additions & 14 deletions torchtitan/experiments/simple_fsdp/simple_fsdp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,7 @@
Replicate,
Shard,
)
from torch.distributed.device_mesh import _mesh_resources, DeviceMesh
from torch.distributed.device_mesh import DeviceMesh
from torch.distributed.tensor._dtensor_spec import DTensorSpec
from torch.distributed.tensor._redistribute import redistribute_local_tensor
from torch.distributed.tensor.placement_types import _StridedShard, Placement
Expand Down Expand Up @@ -95,19 +95,7 @@ def _distribute_dtensor(
"""
inner_spec = tensor._spec
outer_mesh, inner_mesh = device_mesh, inner_spec.mesh
outer_global_mesh = _mesh_resources.get_root_mesh(outer_mesh)
inner_global_mesh = _mesh_resources.get_root_mesh(inner_mesh)
if outer_global_mesh != inner_global_mesh or (
outer_global_mesh is None or inner_global_mesh is None
):
raise AssertionError(
"Cannot distribute tensor across two meshes without the same root mesh: \n"
f"outer global mesh: {outer_global_mesh}\ninner global mesh: {inner_global_mesh}"
)
assert outer_mesh.mesh_dim_names is not None
assert inner_mesh.mesh_dim_names is not None
submesh_names = outer_mesh.mesh_dim_names + inner_mesh.mesh_dim_names
spanned_mesh = outer_global_mesh[submesh_names]
spanned_mesh = DeviceMesh._concatenate((outer_mesh, inner_mesh))
Copy link

@lw lw Oct 30, 2025

Choose a reason for hiding this comment

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

Nice!

One nit: isn't the argument supposed to be a list, and not a tuple?

If so, how come there is no type checking or other linting to catch this?

Note that we've already observed TorchTitan being somewhat incorrect with types, e.g., it passes lists to init_device_mesh instead of tuples.

Copy link
Contributor Author

@fegin fegin Oct 30, 2025

Choose a reason for hiding this comment

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

Good point, I believe we have not enabled type checking for TorchTitan, which we should.


if len(dp_placements) == 1:
assert dp_placements[0].is_replicate() or dp_placements[0].is_shard()
Expand Down
Loading