mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 00:21:07 +01:00
d5236f8517
Summary: While working on https://github.com/pytorch/pytorch/issues/38911, I realized that `nccl.reduce` only needs a single output tensor, while our current implementation requires a list of output tensors. This, along with a TODO I fixed in reduce_add, should have some speed up for data parallel. Pull Request resolved: https://github.com/pytorch/pytorch/pull/39688 Differential Revision: D22034547 Pulled By: mrshenli fbshipit-source-id: e74d54d673ebbb062474b1bb5cc93a095a3a5f6c
86 lines
2.7 KiB
Python
86 lines
2.7 KiB
Python
import warnings
|
|
|
|
import torch._six
|
|
import torch.cuda
|
|
|
|
|
|
__all__ = ['all_reduce', 'reduce', 'broadcast', 'all_gather', 'reduce_scatter']
|
|
|
|
SUM = 0 # ncclRedOp_t
|
|
|
|
|
|
def is_available(tensors):
|
|
if not hasattr(torch._C, '_nccl_all_reduce'):
|
|
warnings.warn('PyTorch is not compiled with NCCL support')
|
|
return False
|
|
|
|
devices = set()
|
|
for tensor in tensors:
|
|
if tensor.is_sparse:
|
|
return False
|
|
if not tensor.is_contiguous():
|
|
return False
|
|
if not tensor.is_cuda:
|
|
return False
|
|
device = tensor.get_device()
|
|
if device in devices:
|
|
return False
|
|
devices.add(device)
|
|
|
|
return True
|
|
|
|
|
|
def version():
|
|
return torch._C._nccl_version()
|
|
|
|
|
|
def unique_id():
|
|
return torch._C._nccl_unique_id()
|
|
|
|
|
|
def init_rank(num_ranks, uid, rank):
|
|
return torch._C._nccl_init_rank(num_ranks, uid, rank)
|
|
|
|
|
|
def all_reduce(inputs, outputs=None, op=SUM, streams=None, comms=None):
|
|
if outputs is None:
|
|
outputs = inputs
|
|
torch._C._nccl_all_reduce(inputs, outputs, op, streams, comms)
|
|
|
|
|
|
# `output` used to be `outputs`, taking in a list of tensors. So we have two
|
|
# arguments for BC reasons.
|
|
def reduce(inputs, output=None, root=0, op=SUM, streams=None, comms=None, *, outputs=None):
|
|
if outputs is not None:
|
|
if output is not None:
|
|
raise ValueError(
|
|
"'output' and 'outputs' can not be both specified. 'outputs' is deprecated in "
|
|
"favor of 'output', taking in a single output tensor. The signature of reduce is: "
|
|
"reduce(inputs, output=None, root=0, op=SUM, streams=None, comms=None).")
|
|
else:
|
|
warnings.warn(
|
|
"nccl.reduce with an output tensor list is deprecated. "
|
|
"Please specify a single output tensor with argument 'output' instead instead.")
|
|
output = outputs[root]
|
|
elif not isinstance(output, torch.Tensor) and isinstance(output, torch._six.container_abcs.Sequence):
|
|
# User called old API with positional arguments of list of output tensors.
|
|
warnings.warn(
|
|
"nccl.reduce with an output tensor list is deprecated. "
|
|
"Please specify a single output tensor.")
|
|
output = output[root]
|
|
elif output is None:
|
|
output = inputs[root]
|
|
torch._C._nccl_reduce(inputs, output, root, op, streams, comms)
|
|
|
|
|
|
def broadcast(inputs, root=0, streams=None, comms=None):
|
|
torch._C._nccl_broadcast(inputs, root, streams, comms)
|
|
|
|
|
|
def all_gather(inputs, outputs, streams=None, comms=None):
|
|
torch._C._nccl_all_gather(inputs, outputs, streams, comms)
|
|
|
|
|
|
def reduce_scatter(inputs, outputs, op=SUM, streams=None, comms=None):
|
|
torch._C._nccl_reduce_scatter(inputs, outputs, op, streams, comms)
|