Skip to content

Support sparse gradients in DistributedDataParallel #19443

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
pietern wants to merge 4 commits into from
Closed

Support sparse gradients in DistributedDataParallel #19443

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
16b28b9
V1: Initial commit
pietern Apr 18, 2019
fa4b1fd
V3: Rebase
pietern Jun 19, 2019
bbed05f
V4: Include test
pietern Jun 19, 2019
0ef3d6f
V5: Fix assert
pietern Jun 19, 2019
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
40 changes: 40 additions & 0 deletions test/test_c10d.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2646,6 +2646,46 @@ def forward(self, x):
loss = criterion(output, target)
loss.backward()

def test_sparse_gradients(self):
store = c10d.FileStore(self.file.name, self.world_size)
process_group = c10d.ProcessGroupGloo(store, self.rank, self.world_size)

class SparseGradientModule(nn.Module):
def __init__(self):
super(SparseGradientModule, self).__init__()
self.embedding = nn.EmbeddingBag(10, 10, sparse=True)

def forward(self, x):
return F.softmax(self.embedding(x), dim=1)

# Ensure initialized weights and inputs are identical across processes
torch.manual_seed(1337)

vanilla_model = SparseGradientModule()
ddp_model = DistributedDataParallel(
copy.deepcopy(vanilla_model),
process_group=process_group,
)

mult = 2
batch_size = mult * self.world_size
criterion = nn.CrossEntropyLoss()
input = torch.randint(0, 10, [batch_size, 2])
target = torch.randint(0, 10, [batch_size])

# Run with entire batch against single process version
criterion(vanilla_model(input), target).backward()

# Run with partial batch against multi process version
partial_input = input.split(mult)[self.rank]
partial_target = target.split(mult)[self.rank]
criterion(ddp_model(partial_input), partial_target).backward()

# Check that the gradients are sparse and identical
vanilla_parameter = next(vanilla_model.parameters())
ddp_parameter = next(ddp_model.parameters())
self.assertEqual(vanilla_parameter.grad, ddp_parameter.grad)


class ReducerModule(nn.Module):
def __init__(self):
Expand Down
2 changes: 2 additions & 0 deletions torch/csrc/distributed/c10d/init.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,7 @@ PyObject* c10d_init(PyObject* _unused) {
shared_ptr_class_<::c10d::Reducer>(module, "Reducer")
.def(py::init<
std::vector<std::vector<torch::autograd::Variable>>,
std::vector<std::vector<bool>>,
std::vector<std::vector<size_t>>,
std::shared_ptr<::c10d::ProcessGroup>>())
.def(
Expand Down Expand Up @@ -543,6 +544,7 @@ They are used in specifying strategies for reduction collectives, e.g.,
&::c10d::compute_bucket_assignment_by_size,
py::arg("tensors"),
py::arg("bucket_size"),
py::arg("expect_sparse_gradient"),
py::call_guard<py::gil_scoped_release>());

module.def(
Expand Down
Loading