Distributed Optimizers¶
-
class
torch.distributed.optim.
ZeroRedundancyOptimizer
(params, optimizer_class, group=None, parameters_as_bucket_view=False, **default)[source]¶ This class wraps an arbitrary
optim.Optimizer
and shards its states across ranks in the group as described by ZeRO. The optimizer instance in each rank is only responsible for updating1 / world_size
parameters and hence only needs to keep1 / world_size
optimizer states. After parameters are updated locally, each rank will broadcast its parameters to all other peers to keep all model replicas in the same state.ZeroRedundancyOptimizer
can be used in conjunction withtorch.nn.parallel.DistributedDataparallel
to reduce per-rank peak memory consumption.ZeroRedundancyOptimizer
use a greedy algorithm to pack a number of parameters at each rank. Each parameter belongs to a single rank and is not divided among ranks. The partition is arbitrary and might not match the the parameter registration or usage order.- Parameters
params (
Iterable
) – anIterable
oftorch.Tensor
s- Keyword Arguments
optimizer_class (
torch.nn.Optimizer
) – the class of the local optimizer.group (
ProcessGroup
, optional) –torch.distributed
ProcessGroup
(default:group.WORLD
initialized bytorch.distributed.init_process_group()
).parameters_as_bucket_views (bool) – when enabled, parameters will be packed into larger buckets to speed up communication and
param.data
fields will point to bucket views at different offsets. When disabled, each individual parameter will be communicated separately, butparams.data
will stay intact.**default – all trailing arguments will be forwarded to the given optimizer.
Example:
>>> import torch.nn as nn >>> from torch.distributed.optim import ZeroRedundancyOptimizer >>> from torch.nn.parallel import DistributedDataParallel as DDP >>> model = nn.Sequential(*[nn.Linear(2000, 2000).to(rank) for _ in range(20)]) >>> ddp = DDP(model, device_ids=[rank]) >>> opt = ZeroRedundancyOptimizer( >>> ddp.parameters(), >>> optimizer_class=torch.optim.Adam, >>> lr=0.01 >>> ) >>> ddp(inputs).sum().backward() >>> opt.step()
-
add_param_group
(param_group)[source]¶ Add a param group to the
Optimizer
sparam_groups
.This can be useful when fine tuning a pre-trained network, as frozen layers can be made trainable and added to the
Optimizer
as training progresses.- Parameters
param_group (dict) – Specifies what Tensors should be optimized along with group specific optimization options.
-
consolidate_state_dict
(to=0)[source]¶ Update the consolidated state_dict list, one per rank.
- Parameters
to (int) – the rank that receives the global states. (default: 0)
-
load_state_dict
(state_dict)[source]¶ Restore the global parameter groups as well as the shard.
- Parameters
state_dict (dict) – optimizer state. Should be an object returned from a call to
state_dict()
-
local_state_dict
()[source]¶ Gets this rank’s
state_dict
.- Returns
The state of the optimizer as a
dict
. It contains two entries:- state - a dict holding current optimization state. Its content
differs between optimizer classes.
param_groups - a dict containing all parameter groups
-
partition_parameters
()[source]¶ Partitions parameters across distributed data parallel ranks.
- Returns
a list of
param_groups
(which is a list of dict) where each element of the list contains the param_groups for a rank. Element 0 corresponds to rank 0, etc. We need all the ranks for the broadcast insidestep()
.
-
static
rank_local_state_dict
(rank, state_dict)[source]¶ Returns the local_state_dict for a given rank.