Skip to content

TrainingStateAverager does not average extra_tensors if offload_optimizer == True #457

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

Draft
wants to merge 2 commits into
base: master
Choose a base branch
from
Draft

TrainingStateAverager does not average extra_tensors if offload_optimizer == True #457

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5244d8a
add a unit test that fails (but shouldn't)
artek0chumak Feb 17, 2022
53cf4b6
black
justheuristic Feb 17, 2022
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
22 changes: 19 additions & 3 deletions tests/test_optimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -147,34 +147,50 @@ def test_state_averager(offload_optimizer: bool, reuse_tensors: bool, sync_epoch


@pytest.mark.forked
def test_load_state_from_peers():
@pytest.mark.parametrize("dpu", [True, False])
def test_load_state_from_peers(dpu: bool):
dht1 = hivemind.DHT(start=True)
dht2 = hivemind.DHT(initial_peers=dht1.get_visible_maddrs(), start=True)

model1 = nn.Linear(2, 3)
model2 = nn.Linear(2, 3)

extras1 = (torch.randn(2, 2), -torch.rand(1))
extras2 = (-torch.randn(2, 2), torch.rand(1))

common_kwargs = dict(
optimizer=partial(torch.optim.SGD, lr=0.1),
scheduler=partial(torch.optim.lr_scheduler.LambdaLR, lr_lambda=lambda t: 1.0 / max(1, t)),
offload_optimizer=dpu,
reuse_tensors=dpu,
target_group_size=2,
prefix="my_exp",
)

avgr1 = TrainingStateAverager(
dht=dht1, params=model1.parameters(), allow_state_sharing=False, start=True, **common_kwargs
dht=dht1,
params=model1.parameters(),
allow_state_sharing=False,
start=True,
extra_tensors=extras1,
**common_kwargs,
)

avgr2 = TrainingStateAverager(dht=dht2, params=model2.parameters(), start=True, **common_kwargs)
avgr2 = TrainingStateAverager(
dht=dht2, params=model2.parameters(), start=True, extra_tensors=extras2, **common_kwargs
)

avgr2.local_epoch = 1337
model2.weight.data[...] = 42
extras2[0][:] = 9999
time.sleep(0.1)

avgr1.load_state_from_peers()
assert avgr1.local_epoch == 1337
assert torch.all(model1.weight == 42).item()
assert np.allclose(avgr1.optimizer.param_groups[0]["lr"], 0.1 / 1337)
assert torch.all(extras1[0] == extras2[0]).item() and torch.all(extras1[0] == extras2[0]).item()
assert torch.all(extras1[0] == 9999).item()


@pytest.mark.forked
Expand Down