Skip to content

[DTensor] create default process group when absent #91756

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

Closed
XilunWu wants to merge 7 commits into from
Closed

[DTensor] create default process group when absent #91756

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
c6e3d76
[DTensor] create default process group when absent
XilunWu Jan 5, 2023
61e5171
Update on "[DTensor] create default process group when absent"
XilunWu Jan 5, 2023
a4a9fb9
Update on "[DTensor] create default process group when absent"
XilunWu Jan 5, 2023
d156f17
Update on "[DTensor] create default process group when absent"
XilunWu Jan 5, 2023
16bf1ba
Update on "[DTensor] create default process group when absent"
XilunWu Jan 6, 2023
7cb3154
Update on "[DTensor] create default process group when absent"
XilunWu Jan 6, 2023
61ec047
Update on "[DTensor] create default process group when absent"
XilunWu Jan 12, 2023
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
20 changes: 20 additions & 0 deletions test/distributed/_tensor/test_device_mesh.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Copyright (c) Meta Platforms, Inc. and affiliates
# Owner(s): ["oncall: distributed"]
import os
import sys

import torch
from torch.distributed._tensor.device_mesh import DeviceMesh
Expand All @@ -8,6 +10,7 @@
from torch.distributed.distributed_c10d import (
get_global_rank,
get_world_size,
is_initialized,
new_group,
ProcessGroup,
)
Expand All @@ -16,13 +19,30 @@
DTensorTestBase,
with_comms,
)
from torch.testing._internal.common_distributed import TEST_SKIPS


class DeviceMeshTest(DTensorTestBase):
@property
def world_size(self):
return 8

def test_init_process_group(self):
device_type = "cuda" if torch.cuda.is_available() else "cpu"
backend = "nccl" if device_type == "cuda" else "gloo"
# skip the test if not enough GPUs
if backend == "nccl" and torch.cuda.device_count() < self.world_size:
sys.exit(TEST_SKIPS[f"multi-gpu-{self.world_size}"].exit_code)
mesh_tensor = torch.arange(4).reshape(2, 2)
self.assertTrue(not is_initialized())
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "25364"
os.environ["WORLD_SIZE"] = f"{self.world_size}"
os.environ["RANK"] = f"{self.rank}"
mesh = DeviceMesh(device_type, mesh_tensor)
self.assertTrue(is_initialized())
self.destroy_pg()

@with_comms
def test_device_mesh_2d(self):
mesh_tensor = torch.arange(4).reshape(2, 2)
Expand Down
10 changes: 9 additions & 1 deletion torch/distributed/_tensor/device_mesh.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,8 @@
get_rank,
get_world_size,
GroupMember,
init_process_group,
is_initialized,
new_group,
ProcessGroup,
reduce_scatter,
Expand Down Expand Up @@ -109,7 +111,7 @@ def __init__(
if isinstance(mesh, torch.Tensor)
else torch.tensor(mesh, dtype=torch.int)
)
default_pg = _get_default_group()
default_pg = self._get_or_create_default_group()
self._backend = default_pg._get_backend_name()
# TODO: if user want to pass pg_options, offer a way to do it
# check default pg backend, should support device_type
Expand Down Expand Up @@ -215,6 +217,12 @@ def __init__(
)
self._dim_groups.append(new_subgroup)

def _get_or_create_default_group(self):
if not is_initialized():
_backend = "gloo" if self.device_type == "cpu" else "nccl"
init_process_group(backend=_backend)
return _get_default_group()

def __enter__(self) -> "DeviceMesh":
# set global device_mesh to this instance
set_global_device_mesh(self)
Expand Down