DP Enhancement (#8280)
This commit is contained in:
@@ -545,6 +545,15 @@ class GroupCoordinator:
|
||||
else:
|
||||
torch.distributed.all_reduce(input_, group=self.device_group)
|
||||
|
||||
def reduce_scatter_tensor(
|
||||
self,
|
||||
output: torch.Tensor,
|
||||
input: torch.Tensor,
|
||||
) -> None:
|
||||
# TODO(ch-wan): support other backends
|
||||
torch.distributed.reduce_scatter_tensor(output, input, group=self.device_group)
|
||||
return output
|
||||
|
||||
def reduce_scatter(
|
||||
self,
|
||||
output: torch.Tensor,
|
||||
|
||||
Reference in New Issue
Block a user