Skip to content

[Distributed] Test if xm.all_reduce produces a single dynamo graph #4920

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
merged 2 commits into from
Apr 21, 2023
Merged

[Distributed] Test if xm.all_reduce produces a single dynamo graph #4920

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3e0f6c3
add a single graph dynamo test
alanwaketan Apr 20, 2023
5768f87
Fix linters
alanwaketan Apr 20, 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
17 changes: 15 additions & 2 deletions test/test_mp_replication.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,10 @@
import torch_xla.distributed.xla_multiprocessing as xmp


def all_reduce(tensor):
return xm.all_reduce(xm.REDUCE_SUM, tensor)


def _mp_fn(index):
device = xm.xla_device()
world_size = xm.xrt_world_size()
Expand All @@ -13,23 +17,32 @@ def _mp_fn(index):
twos = ones + 1.0
threes = ones + 2.0
fours = ones + 3.0
fives = ones + 4.0
scale = 0.5
xones = ones.to(device)
xtwos = twos.to(device)
xthrees = threes.to(device)
xfours = fours.to(device)
xfives = fives.to(device)
xm.all_reduce(xm.REDUCE_SUM, [xones, xtwos])
xthrees = xm.all_reduce(xm.REDUCE_SUM, xthrees)
xthrees = all_reduce(xthrees)
xfours = xm.all_reduce(xm.REDUCE_SUM, xfours, scale=scale)

compiled_all_reduce = torch.compile(
Copy link
Collaborator

@JackCaoG JackCaoG Apr 20, 2023

Choose a reason for hiding this comment

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

I would prefer move dynamo tests under test/dynamo, we can do that in a follow up pr.

Copy link
Collaborator Author

@alanwaketan alanwaketan Apr 20, 2023

Choose a reason for hiding this comment

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

Sure, will do that as a follow up.

all_reduce, backend='torchxla_trace_once', fullgraph=True)
xfives = compiled_all_reduce(xfives)

if (not xones.cpu().allclose(ones * float(world_size)) or
not xtwos.cpu().allclose(twos * float(world_size)) or
not xthrees.cpu().allclose(threes * float(world_size)) or
not xfours.cpu().allclose(fours * float(world_size) * scale)):
not xfours.cpu().allclose(fours * float(world_size) * scale) or
not xfives.cpu().allclose(fives * float(world_size))):
print('xm.all_reduce() produced wrong reductions', file=sys.stderr)
print(xones, file=sys.stderr)
print(xtwos, file=sys.stderr)
print(xthrees, file=sys.stderr)
print(xfours, file=sys.stderr)
print(xfives, file=sys.stderr)
sys.exit(1)
else:
print(
Expand Down