tensorplay.distributed.device_mesh API

Functions 1

init_device_mesh

tensorplay.distributed.device_mesh.init_device_mesh(device_type: str, mesh_shape: tuple[int, ...], *, mesh_dim_names: tuple[str, ...] | None = None, backend_override=None) → DeviceMesh[source]

This creates a DeviceMesh with an n-dimensional array layout, where n is the length of mesh_shape. If mesh_dim_names is provided, each dimension is labeled as mesh_dim_names[i].

Note

Follows SPMD: ensure mesh_shape is identical across all ranks.

Example:

>>> mesh_1d = init_device_mesh("cuda", mesh_shape=(8,))
>>> mesh_2d = init_device_mesh("cuda", mesh_shape=(2, 8),
...                             mesh_dim_names=("dp", "tp"))

Classes 1

DeviceMesh

classFull reference ↗

class tensorplay.distributed.device_mesh.DeviceMesh(device_type: str, mesh=None, *, mesh_dim_names=None, _dim_group_names=None, _rank_map=None, _sizes=None, _strides=None, _root_mesh=None, _backend_override=None, _axis_root_dims=None)[source]

The mesh is an n-d array whose values are global ranks. Process groups are created per mesh dimension so collectives can run on each dimension independently.

Example:

>>> from tensorplay.distributed.device_mesh import init_device_mesh
>>> mesh = init_device_mesh("cuda", mesh_shape=(2, 4),
...                         mesh_dim_names=("dp", "tp"))

classmethod from_group(group, device_type=None, mesh=None, mesh_dim_names=None) → DeviceMesh[source]: Construct a DeviceMesh from one or more existing process groups.

get_coordinate() → tuple[int, ...] | None[source]: Returns this rank’s coordinate in the mesh, or None if absent.

Functions 1

init_device_mesh

Classes 1

DeviceMesh

Help improve this page

On this page