mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
d2078fac11
Introduce _DistWrapper class that wraps a process group and provides functional variants of collectives. It works without c10d enabled and is exception robust. Introduce tensor_narrow_n that handle narrowing over multiple dimentions. Fixes #ISSUE_NUMBER Pull Request resolved: https://github.com/pytorch/pytorch/pull/81828 Approved by: https://github.com/wanchaol
27 lines
1020 B
Python
27 lines
1020 B
Python
import torch
|
|
from torch.distributed._shard.metadata import ShardMetadata
|
|
from typing import Sequence
|
|
|
|
def narrow_tensor_by_index(tensor: torch.Tensor, offsets: Sequence[int], sizes: Sequence[int]) -> torch.Tensor:
|
|
"""
|
|
Narrow the tensor according to ``offsets`` and ``sizes``.
|
|
"""
|
|
narrowed_tensor = tensor
|
|
for idx, (offset, size) in enumerate(zip(offsets, sizes)):
|
|
if size < tensor.size(idx):
|
|
# Reshape to get shard for this rank and we don't want autograd
|
|
# recording here for the narrow op and 'local_shard' should be a
|
|
# leaf variable in the autograd graph.
|
|
narrowed_tensor = narrowed_tensor.narrow(
|
|
idx,
|
|
offset,
|
|
size
|
|
)
|
|
return narrowed_tensor
|
|
|
|
def narrow_tensor(tensor: torch.Tensor, metadata: ShardMetadata) -> torch.Tensor:
|
|
"""
|
|
Narrow the tensor according to the metadata
|
|
"""
|
|
return narrow_tensor_by_index(tensor, metadata.shard_offsets, metadata.shard_sizes)
|