Skip to content

Slightly improve DistributedDataParallel (single-GPU binding) multi-process distributed training performance #4870

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
apaszke merged 1 commit into from
Jan 27, 2018
Merged

Slightly improve DistributedDataParallel (single-GPU binding) multi-process distributed training performance #4870

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
23 changes: 13 additions & 10 deletions torch/nn/parallel/distributed.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,11 +189,13 @@ def train(self, mode=True):
module.train(mode)

def _sync_params(self):
params = [p.data for p in self.module.parameters()]
result = broadcast_coalesced(params, self.device_ids, self.broadcast_bucket_size)
for tensors, module in zip(result[1:], self._module_copies[1:]):
for tensor, param in zip(tensors, module.parameters()):
param.data.set_(tensor)
if len(self.device_ids) > 1:
# intra-node parameter sync
params = [p.data for p in self.module.parameters()]
result = broadcast_coalesced(params, self.device_ids, self.broadcast_bucket_size)
for tensors, module in zip(result[1:], self._module_copies[1:]):
for tensor, param in zip(tensors, module.parameters()):
param.data.set_(tensor)

buffers = list(self.module._all_buffers())
if len(buffers) > 0:
Expand All @@ -203,11 +205,12 @@ def _sync_params(self):
for buf, synced in zip(buffers, _unflatten_dense_tensors(flat_buffers, buffers)):
buf.copy_(synced)

# intra-node buffer sync
result = broadcast_coalesced(buffers, self.device_ids, self.broadcast_bucket_size)
for tensors, module in zip(result[1:], self._module_copies[1:]):
for tensor, buf in zip(tensors, module._all_buffers()):
buf.set_(tensor)
if len(self.device_ids) > 1:
# intra-node buffer sync
result = broadcast_coalesced(buffers, self.device_ids, self.broadcast_bucket_size)
for tensors, module in zip(result[1:], self._module_copies[1:]):
for tensor, buf in zip(tensors, module._all_buffers()):
buf.set_(tensor)

def _register_grad_hooks(self):
self._grad_accs = [] # need to keep them in scope
Expand Down