Skip to content

[SPMD] Fix reduce_scatter for TPUv2 #7248

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 6 commits into from
Jun 14, 2024
Merged

[SPMD] Fix reduce_scatter for TPUv2 #7248

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
4516f9d
Fix tpu v3
alanwaketan Jun 11, 2024
bb4f179
skip tpu v2
alanwaketan Jun 12, 2024
3415c6d
Fix linters
alanwaketan Jun 12, 2024
f92a4dd
Fix
alanwaketan Jun 12, 2024
658f68f
Add a debug print
alanwaketan Jun 12, 2024
6a15573
remove print
alanwaketan Jun 14, 2024
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
10 changes: 6 additions & 4 deletions test/spmd/test_xla_sharding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1213,7 +1213,8 @@ def test_manual_sharding_api_e2e(self):
self.assertEqual(xxx.shape, (8, 8))
self.assertTrue(torch.allclose(x.cpu() + 1, xxx.cpu()))

@unittest.skipIf(xr.device_type() != 'TPU', "Skip non-TPU device")
@unittest.skipIf(xr.device_type() != 'TPU' or tpu.version() < 4,
"Only runs on TPUv4")
def test_spmd_reduce_scatter(self):
xs.set_global_mesh(self._get_mesh((1, self.n_devices)))
x = torch.ones(8, 8).to(xm.xla_device())
Expand All @@ -1230,10 +1231,11 @@ def test_spmd_reduce_scatter(self):
f"reduce-scatter(f32[8,8]{{1,0}} %custom-call.2), channel_id=1, replica_groups={{{{{','.join([str(x) for x in self.device_ids])}}}}}, use_global_device_ids=true, dimensions={{0}}, to_apply=%AddComputation.3",
hlo)

expected_x = torch.ones(2, 8) * 4
expected_x = torch.ones(8 // self.n_devices, 8) * 4
self.assertTrue(torch.allclose(x.cpu(), expected_x))

@unittest.skipIf(xr.device_type() != 'TPU', "Skip non-TPU device")
@unittest.skipIf(xr.device_type() != 'TPU' or tpu.version() < 4,
"Only runs on TPUv4")
def test_spmd_reduce_scatter_canonical_index(self):
xs.set_global_mesh(self._get_mesh((1, self.n_devices)))
x = torch.ones(8, 8).to(xm.xla_device())
Expand All @@ -1250,7 +1252,7 @@ def test_spmd_reduce_scatter_canonical_index(self):
f"reduce-scatter(f32[8,8]{{1,0}} %custom-call.2), channel_id=1, replica_groups={{{{{','.join([str(x) for x in self.device_ids])}}}}}, use_global_device_ids=true, dimensions={{1}}, to_apply=%AddComputation.3",
hlo)

expected_x = torch.ones(8, 2) * 4
expected_x = torch.ones(8, 8 // self.n_devices) * 4
self.assertTrue(torch.allclose(x.cpu(), expected_x))


Expand Down