ColossalAI/colossalai/communication/utils.py

71 lines
2.5 KiB
Python

import torch
import torch.distributed as dist
from colossalai.context.parallel_mode import ParallelMode
from colossalai.core import global_context as gpc
from colossalai.utils import get_current_device
def send_tensor_meta(tensor, need_meta=True, next_rank=None):
"""Sends tensor meta information before sending a specific tensor.
Since the recipient must know the shape of the tensor in p2p communications,
meta information of the tensor should be sent before communications. This function
synchronizes with :func:`recv_tensor_meta`.
:param tensor: Tensor to be sent
:param need_meta: If False, meta information won't be sent
:param next_rank: The rank of the next member in pipeline parallel group
:type tensor: Tensor
:type need_meta: bool, optional
:type next_rank: int
:return: False
:rtype: bool
"""
if need_meta:
if next_rank is None:
next_rank = gpc.get_next_global_rank(ParallelMode.PIPELINE)
tensor_kwargs = {'dtype': torch.long, 'device': get_current_device()}
send_shape = torch.tensor(tensor.size(), **tensor_kwargs)
send_ndims = torch.tensor(len(tensor.size()), **tensor_kwargs)
ops = [
dist.P2POp(dist.isend, send_ndims, next_rank),
dist.P2POp(dist.isend, send_shape, next_rank)
]
reqs = dist.batch_isend_irecv(ops)
for req in reqs:
req.wait()
torch.cuda.synchronize()
return False
def recv_tensor_meta(tensor_shape, prev_rank=None):
"""Recieves tensor meta information before recieving a specific tensor.
Since the recipient must know the shape of the tensor in p2p communications,
meta information of the tensor should be recieved before communications. This function
synchronizes with :func:`send_tensor_meta`.
:param tensor_shape: The shape of the tensor to be recieved
:param prev_rank: The rank of the source of the tensor
:type tensor_shape: torch.Size
:type prev_rank: int, optional
:return: The shape of the tensor to be recieved
:rtype: torch.Size
"""
if tensor_shape is None:
if prev_rank is None:
prev_rank = gpc.get_prev_global_rank(ParallelMode.PIPELINE)
tensor_kwargs = {'dtype': torch.long, 'device': get_current_device()}
recv_ndims = torch.empty((), **tensor_kwargs)
dist.recv(recv_ndims, prev_rank)
recv_shape = torch.empty(recv_ndims, **tensor_kwargs)
dist.recv(recv_shape, prev_rank)
tensor_shape = torch.Size(recv_shape)
return tensor_shape