def __init__(self, module, device_ids=None, output_device=None, dim=0):
super(torch.nn.DataParallel, self).__init__()
#super().__init__(module, device_ids, output_device, dim)
if not torch.cuda.is_available():
self.module = module
self.device_ids = []
return
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
self.dim = dim
self.module = module
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
self.src_device_obj = torch.device("cuda:{}".format(
self.device_ids[0]))
#print("module:", module, "dim:", dim, "device_ids:", self.device_ids, \
# "output_device:", self.output_device, "src_device_obj:", self.src_device_obj)
_check_balance(self.device_ids)
if len(self.device_ids) == 1:
#print("len(self.device_ids):", len(self.device_ids))
self.module.cuda(device_ids[0])
def __init__(self, module, device_ids=None, output_device=None, dim=0):
super(DataParallel, self).__init__()
if not torch.cuda.is_available():
self.module = module
self.device_ids = []
return
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
self.dim = dim
self.module = module
self.device_ids = list(map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
self.src_device_obj = torch.device("cuda:{}".format(self.device_ids[0]))
self.n_device = len(device_ids)
self.seen = 0
_check_balance(self.device_ids)
if len(self.device_ids) == 1:
self.module.cuda(device_ids[0])
def __init__(self, module, device_ids=None, output_device=None, dim=0):
super(DataParallelImbalance, self).__init__(module, device_ids,
output_device, dim)
if not torch.cuda.is_available():
self.module = module
self.device_ids = []
return
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
if not all(t.is_cuda and t.device.index == device_ids[0]
for t in chain(module.parameters(), module.buffers())):
raise RuntimeError("module must have its parameters and buffers "
"on device %d (device_ids[0])" % device_ids[0])
self.dim = dim
self.module = module
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
if len(self.device_ids) == 1:
self.module.cuda(device_ids[0])
def __init__(self,
module,
device_ids=None,
output_device=None,
dim=0,
broadcast_buffers=True,
process_group=None,
bucket_cap_mb=25,
check_reduction=False,
randk=1,
seed=2147483647):
super(RandomKSparsifiedDDP, self).__init__()
torch.manual_seed(seed)
# Use all devices by default
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if len(device_ids) > 1:
raise RuntimeError(
"This module only supports Multi-Process Single-GPU mode.")
if output_device is None:
output_device = device_ids[0]
if process_group is None:
self.process_group = _get_default_group()
else:
self.process_group = process_group
self.dim = dim
self.module = module
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
self.broadcast_buffers = broadcast_buffers
self.check_reduction = check_reduction
self.randk = randk
self.masks = {}
self.global_step = 0
MB = 1024 * 1024
# used for intra-node param sync and inter-node sync as well
self.broadcast_bucket_size = 250 * MB
# reduction bucket size
self.bucket_bytes_cap = bucket_cap_mb * MB
# Sync params and buffers
module_states = list(self.module.state_dict().values())
if len(module_states) > 0:
self._dist_broadcast_coalesced(module_states,
self.broadcast_bucket_size)
self._ddp_init_helper()
def data_parallel(
module,
inputs,
device_ids=None,
output_device=None,
dim=0,
module_kwargs=None,
non_scatter_kwargs=None,
):
r"""Evaluates module(input) in parallel across the GPUs given in device_ids.
This is the functional version of the DataParallel module.
Args:
module (Module): the module to evaluate in parallel
inputs (Tensor): inputs to the module
device_ids (list of int or torch.device): GPU ids on which to replicate module
output_device (list of int or torch.device): GPU location of the output
Use -1 to indicate the CPU.
(default: device_ids[0])
Returns:
a Tensor containing the result of module(input) located on
output_device
"""
if not isinstance(inputs, tuple):
inputs = (inputs,)
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
device_ids = list(map(lambda x: _get_device_index(x, True), device_ids))
output_device = _get_device_index(output_device, True)
src_device_obj = torch.device("cuda:{}".format(device_ids[0]))
for tensor in chain(module.parameters(), module.buffers()):
if tensor.device != src_device_obj:
raise RuntimeError(
"module must have its parameters and buffers "
"on device {} (device_ids[0]) but found one of "
"them on device: {}".format(src_device_obj, tensor.device)
)
inputs, module_kwargs = scatter_kwargs(inputs, module_kwargs, device_ids, dim)
if len(device_ids) == 1:
return module(*inputs[0], **module_kwargs[0])
used_device_ids = device_ids[: len(inputs)]
replicas = replicate(module, used_device_ids)
outputs = parallel_apply(replicas, inputs, module_kwargs, used_device_ids)
return gather(outputs, output_device, dim)
def __init__(self,
module,
device_ids=None,
output_device=None,
dim=0,
broadcast_buffers=True,
process_group=None,
bucket_cap_mb=25,
check_reduction=False):
super(DistributedDataParallel, self).__init__()
# Use all devices by default
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
if process_group is None:
self.process_group = _get_default_group()
else:
self.process_group = process_group
self.dim = dim
self.module = module
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
self.broadcast_buffers = broadcast_buffers
if check_reduction:
# This argument is no longer used since the reducer
# will ensure reduction completes even if some parameters
# do not receive gradients.
pass
MB = 1024 * 1024
# used for intra-node param sync and inter-node sync as well
self.broadcast_bucket_size = int(250 * MB)
# reduction bucket size
self.bucket_bytes_cap = int(bucket_cap_mb * MB)
# Sync params and buffers
module_states = list(self.module.state_dict().values())
if len(module_states) > 0:
self._dist_broadcast_coalesced(module_states,
self.broadcast_bucket_size)
self._ddp_init_helper()
def __init__(
self,
module,
device_ids=None,
output_device=None,
dim=0,
broadcast_buffers=True,
process_group=None,
bucket_cap_mb=25,
check_reduction=False,
find_unused_parameters=False,
):
super(DistributedDataParallelPythonBuckets, self).__init__()
# Use all devices by default
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
if process_group is None:
self.process_group = _get_default_group()
else:
self.process_group = process_group
self.dim = dim
self.module = module
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
self.broadcast_buffers = broadcast_buffers
self.check_reduction = check_reduction
MB = 1024 * 1024
# used for intra-node param sync and inter-node sync as well
self.broadcast_bucket_size = 250 * MB
# reduction bucket size
self.bucket_bytes_cap = bucket_cap_mb * MB
# Sync params and buffers
module_states = list(self.module.state_dict().values())
if len(module_states) > 0:
self._dist_broadcast_coalesced(module_states,
self.broadcast_bucket_size)
self._ddp_init_helper()
def __init__(self,
module,
device_id=None,
output_device=None,
dim=0,
broadcast_buffers=True,
process_group=None,
bucket_cap_mb=25,
check_reduction=False,
sparse_ratio=0.1,
sparse_threshold=1024,
mem_decay=1.0):
super(DistributedDataParallel, self).__init__()
self.module = module
if device_id is None:
raise RuntimeError("device_id cannot be None")
if output_device is None:
output_device = device_id
if process_group is None:
self.process_group = _get_default_group()
else:
self.process_group = process_group
self.dim = dim
self.module = module
self.device_id = _get_device_index(device_id, True)
self.output_device = _get_device_index(output_device, True)
self.broadcast_buffers = broadcast_buffers
self.check_reduction = check_reduction
self.sparse_ratio = sparse_ratio
self.sparse_threshold = sparse_threshold
self.mem_decay = mem_decay
MB = 1024 * 1024
# used for intra-node param sync and inter-node sync as well
self.broadcast_bucket_size = 250 * MB
module_states = list(self.module.state_dict().values())
if len(module_states) > 0:
self._dist_broadcast_coalesced(module_states,
self.broadcast_bucket_size)
self._ddp_init_helper()
def _to_device_index(devices):
if not devices:
raise RuntimeError("Cannot replicate using an empty device list.")
if isinstance(devices, list) and isinstance(devices[0], list):
device_ids = []
seen = set()
for i, replica_devs in enumerate(devices):
assert len(replica_devs) == len(devices[0]), (
"Cannot replicate to unidentical number of devices, but got "
"device list {} and {} for replica {} and {}.").format(
devices[0], devices[i], 0, i)
assert len(seen.intersection(replica_devs)) == 0, (
"Devices {} are shared by multiple replicas.").format(
seen.intersection(replica_devs))
seen.update(replica_devs)
device_ids.append(_to_device_index(replica_devs))
return device_ids
else:
assert len(devices) == len(
set(devices)), ("Duplicated device ids {}.").format(devices)
return list(map(lambda x: _get_device_index(x, True), devices))
def _check_balance(device_ids):
imbalance_warn = """
There is an imbalance between your GPUs. You may want to exclude GPU {} which
has less than 75% of the memory or cores of GPU {}. You can do so by setting
the device_ids argument to DataParallel, or by setting the CUDA_VISIBLE_DEVICES
environment variable."""
device_ids = list(map(lambda x: _get_device_index(x, True), device_ids))
#print("_check_balance device_ids:", device_ids)
dev_props = [torch.cuda.get_device_properties(i) for i in device_ids]
#print("_check_balance dev_props:", dev_props)
def warn_imbalance(get_prop):
values = [get_prop(props) for props in dev_props]
min_pos, min_val = min(enumerate(values), key=operator.itemgetter(1))
max_pos, max_val = max(enumerate(values), key=operator.itemgetter(1))
if min_val / max_val < 0.75:
warnings.warn(
imbalance_warn.format(device_ids[min_pos],
device_ids[max_pos]))
return True
return False
#print("warn_imbalance total_memory:", warn_imbalance(lambda props: props.total_memory), \
# "warn_imbalance multi_processor_count:", warn_imbalance(lambda props: props.multi_processor_count), \
# "multi_processor_count:", props.multi_processor_count)
if warn_imbalance(lambda props: props.total_memory):
return
if warn_imbalance(lambda props: props.multi_processor_count):
return
def replicate_module(self, module: torch.nn.Module,
devices: List[int]) -> List[torch.nn.Module]:
assert self.n_mask_samples % len(devices) == 0
copies = replicate(module, devices)
def walk(module: torch.nn.Module, copy: torch.nn.Module):
module_map = {id(module): copy}
for name, ref in module._modules.items():
module_map.update(walk(ref, getattr(copy, name)))
return module_map
devices = [_get_device_index(d) for d in devices]
# Copy the custom parameters
all_params = [p.get() for p in self.pointer_values]
if (not self.masking_enabled) or (not self.training):
scattered = _broadcast_coalesced_reshape(all_params, devices)
else:
# Here is more complicated, because there might be non-masked parameters which has to be handled in the
# usual way
masked_indices = [
i for i, n in enumerate(self.param_names) if self.is_masked(n)
]
simple_indices = [
i for i, n in enumerate(self.param_names)
if not self.is_masked(n)
]
masked_params = scatter([all_params[i] for i in masked_indices],
devices)
simple_params = _broadcast_coalesced_reshape(
[all_params[i] for i in simple_indices], devices)
scattered = [[None] * len(all_params) for _ in devices]
for d in range(len(devices)):
for mi, mp in zip(masked_indices, masked_params[d]):
scattered[d][mi] = mp
for si, sp in zip(simple_indices, simple_params[d]):
scattered[d][si] = sp
for i, c in enumerate(copies):
device_map = walk(module, c)
for j, p in enumerate(self.pointer_values):
setattr(device_map[id(p.parent)], p.name, scattered[i][j])
self.update_rnn_params(c)
return copies
def forward(ctx, target_device, dim, *inputs):
assert all(map(lambda i: i.is_cuda, inputs))
target_device = _get_device_index(target_device, True)
ctx.target_device = target_device
ctx.dim = dim
ctx.input_gpus = tuple(map(lambda i: i.get_device(), inputs))
if all(t.dim() == 0 for t in inputs) and dim == 0:
inputs = tuple(t.view(1) for t in inputs)
warnings.warn('Was asked to gather along dimension 0, but all '
'input tensors were scalars; will instead unsqueeze '
'and return a vector.')
ctx.unsqueezed_scalar = True
else:
ctx.unsqueezed_scalar = False
ctx.input_sizes = tuple(map(lambda i: i.size(ctx.dim), inputs))
return comm.gather(inputs, ctx.dim, ctx.target_device)
def forward(ctx, target_gpus, chunk_sizes, dim, input):
target_gpus = list(map(lambda x: _get_device_index(x, True), target_gpus))
ctx.dim = dim
ctx.input_device = input.get_device() if input.is_cuda else -1
streams = None
if ctx.input_device == -1:
# Perform CPU to GPU copies in a background stream
streams = [_get_stream(device) for device in target_gpus]
outputs = comm.scatter(input, target_gpus, chunk_sizes, ctx.dim, streams)
# Synchronize with the copy stream
if streams is not None:
for i, output in enumerate(outputs):
with torch.cuda.device(target_gpus[i]):
main_stream = torch.cuda.current_stream()
main_stream.wait_stream(streams[i])
output.record_stream(main_stream)
return outputs
def forward(ctx, target_gpus, *inputs):
if not all(input.is_cuda for input in inputs):
raise TypeError('Broadcast function not implemented for CPU tensors')
target_gpus = list(map(lambda x: _get_device_index(x, True), target_gpus))
ctx.target_gpus = target_gpus
if len(inputs) == 0:
return tuple()
ctx.num_inputs = len(inputs)
ctx.input_device = inputs[0].get_device()
outputs = comm.broadcast_coalesced(inputs, ctx.target_gpus)
non_differentiables = []
for idx, input_requires_grad in enumerate(ctx.needs_input_grad[1:]):
if not input_requires_grad:
for output in outputs:
non_differentiables.append(output[idx])
ctx.mark_non_differentiable(*non_differentiables)
return tuple([t for tensors in outputs for t in tensors])
def __init__(self,
module,
device_ids,
dim=0,
broadcast_buffers=False,
find_unused_parameters=False,
**kwargs):
super().__init__()
assert len(device_ids) == 1, (
'Currently, DistributedDataParallelWrapper only supports one'
'single CUDA device for each process.'
f'The length of device_ids must be 1, but got {len(device_ids)}.')
self.module = module
self.dim = dim
self.to_ddp(device_ids=device_ids,
dim=dim,
broadcast_buffers=broadcast_buffers,
find_unused_parameters=find_unused_parameters,
**kwargs)
self.output_device = _get_device_index(device_ids[0], True)
def __init__(self,
module,
split_size,
device_ids=None,
output_device=None):
"""
This method instantiates a ModelParallel Module from a module instance passed by the user. The
model must have a single input (forward(x) type of signature for the forward method) otherwise an error is
returned.
An example is here:
.. code-block:: python
from eisen.utils import ModelParallel
from eisen.models.segmentation import UNet
model = ModelParallel(
module=UNet(input_channels=1, output_channels=1),
split_size=2,
device_ids=[0, 1, 2, 3],
output_device=0
)
:param module: an instance of the model that should be parallelized
:type module: torch.nn.Module
:param split_size: split size for pipelined execution
:type split_size: int
:param device_ids: list of int or torch devices indicating GPUs to use
:type device_ids: list
:param output_device: int or torch device indicating output devices
:type output_device: int or torch device
"""
super(ModelParallel, self).__init__()
module_argument_list = inspect.getfullargspec(module.forward)[0]
if len(module_argument_list) > 2:
raise NotImplementedError(
'Support for modules with more than one input is not yet implemented.'
)
self.first_run = True
self.split_size = split_size
if not torch.cuda.is_available():
self.module = module
self.device_ids = []
self.first_run = False
return
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
self.module = module
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
if len(self.device_ids) == 1:
self.first_run = False
self.module.cuda(device_ids[0])
def replicate(network, devices, detach=False):
if not _replicatable_module(network):
raise RuntimeError("Cannot replicate network where python modules are "
"childrens of ScriptModule")
devices = list(map(lambda x: _get_device_index(x, True), devices))
num_replicas = len(devices)
params = list(network.parameters())
param_indices = {param: idx for idx, param in enumerate(params)}
param_copies = _broadcast_coalesced_reshape(params, devices, detach)
buffers = list(network.buffers())
buffers_rg = []
buffers_not_rg = []
for buf in buffers:
if buf.requires_grad and not detach:
buffers_rg.append(buf)
else:
buffers_not_rg.append(buf)
buffer_indices_rg = {buf: idx for idx, buf in enumerate(buffers_rg)}
buffer_indices_not_rg = {
buf: idx
for idx, buf in enumerate(buffers_not_rg)
}
buffer_copies_rg = _broadcast_coalesced_reshape(buffers_rg,
devices,
detach=detach)
buffer_copies_not_rg = _broadcast_coalesced_reshape(buffers_not_rg,
devices,
detach=True)
modules = list(network.modules())
module_copies = [[] for device in devices]
module_indices = {}
scriptmodule_skip_attr = {
"_parameters", "_buffers", "_modules", "forward", "_c"
}
for i, module in enumerate(modules):
module_indices[module] = i
for j in range(num_replicas):
if _is_script_module(module):
# we have to initialize ScriptModule properly so that
# it works with pybind11
def init_fn(script_module):
# Don't do anything here, we'll initialize the ScriptModule below
return
replica = torch.jit.RecursiveScriptModule._construct(
module._c._replicate_for_data_parallel(), init_fn)
else:
replica = module._replicate_for_data_parallel()
module_copies[j].append(replica)
for i, module in enumerate(modules):
for key, child in module._modules.items():
if child is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = None
else:
module_idx = module_indices[child]
for j in range(num_replicas):
replica = module_copies[j][i]
setattr(replica, key, module_copies[j][module_idx])
for key, param in module._parameters.items():
if param is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = None
else:
param_idx = param_indices[param]
for j in range(num_replicas):
replica = module_copies[j][i]
param = param_copies[j][param_idx]
setattr(
replica, key,
Parameter(param, requires_grad=param.requires_grad))
# TODO: We need to manually set _parameters with a bare
# non-parameter Tensor, otherwise gradients don't
# accumulate in the original parameters when you call
# backwards() on the DataParallel module.
replica._parameters[key] = param
for key, buf in module._buffers.items():
if buf is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = None
else:
if buf.requires_grad and not detach:
buffer_copies = buffer_copies_rg
buffer_idx = buffer_indices_rg[buf]
else:
buffer_copies = buffer_copies_not_rg
buffer_idx = buffer_indices_not_rg[buf]
for j in range(num_replicas):
replica = module_copies[j][i]
setattr(replica, key, buffer_copies[j][buffer_idx])
return [module_copies[j][0] for j in range(num_replicas)]
def replicate(network, devices, detach=False):
if not _replicatable_module(network):
raise RuntimeError("Cannot replicate network where python modules are "
"childrens of ScriptModule")
devices = list(map(lambda x: _get_device_index(x, True), devices))
num_replicas = len(devices)
params = list(network.parameters())
param_indices = {param: idx for idx, param in enumerate(params)}
param_copies = _broadcast_coalesced_reshape(params, devices, detach)
buffers = list(network.buffers())
buffers_rg = []
buffers_not_rg = []
for buf in buffers:
if buf.requires_grad and not detach:
buffers_rg.append(buf)
else:
buffers_not_rg.append(buf)
buffer_indices_rg = {buf: idx for idx, buf in enumerate(buffers_rg)}
buffer_indices_not_rg = {buf: idx for idx, buf in enumerate(buffers_not_rg)}
buffer_copies_rg = _broadcast_coalesced_reshape(buffers_rg, devices, detach=detach)
buffer_copies_not_rg = _broadcast_coalesced_reshape(buffers_not_rg, devices, detach=True)
modules = list(network.modules())
module_copies = [[] for device in devices]
module_indices = {}
scriptmodule_skip_attr = {"_parameters", "_buffers", "_modules", "forward", "_c"}
for i, module in enumerate(modules):
module_indices[module] = i
for j in range(num_replicas):
replica = module._replicate_for_data_parallel()
# This is a temporary fix for DDP. DDP needs to access the
# replicated model parameters. It used to do so through
# `mode.parameters()`. The fix added in #33907 for DP stops the
# `parameters()` API from exposing the replicated parameters.
# Hence, we add a `_former_parameters` dict here to support DDP.
replica._former_parameters = OrderedDict()
module_copies[j].append(replica)
for i, module in enumerate(modules):
for key, child in module._modules.items():
if child is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = None
else:
module_idx = module_indices[child]
for j in range(num_replicas):
replica = module_copies[j][i]
setattr(replica, key, module_copies[j][module_idx])
for key, param in module._parameters.items():
if param is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = None
else:
param_idx = param_indices[param]
for j in range(num_replicas):
replica = module_copies[j][i]
param = param_copies[j][param_idx]
# parameters in replicas are no longer leaves,
# so setattr them as non-parameter attributes
setattr(replica, key, param)
# expose the parameter for DDP
replica._former_parameters[key] = param
for key, buf in module._buffers.items():
if buf is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = None
else:
if buf.requires_grad and not detach:
buffer_copies = buffer_copies_rg
buffer_idx = buffer_indices_rg[buf]
else:
buffer_copies = buffer_copies_not_rg
buffer_idx = buffer_indices_not_rg[buf]
for j in range(num_replicas):
replica = module_copies[j][i]
setattr(replica, key, buffer_copies[j][buffer_idx])
return [module_copies[j][0] for j in range(num_replicas)]
def replicate(network, devices, detach=False):
if not _replicatable_module(network):
raise RuntimeError("Cannot replicate network where python modules are "
"childrens of ScriptModule")
devices = list(map(lambda x: _get_device_index(x, True), devices))
num_replicas = len(devices)
params = list(network.parameters())
param_indices = {param: idx for idx, param in enumerate(params)}
param_copies = _broadcast_coalesced_reshape(params, devices, detach)
buffers = list(network.buffers())
buffers_rg = []
buffers_not_rg = []
for buf in buffers:
if buf.requires_grad and not detach:
buffers_rg.append(buf)
else:
buffers_not_rg.append(buf)
buffer_indices_rg = {buf: idx for idx, buf in enumerate(buffers_rg)}
buffer_indices_not_rg = {
buf: idx
for idx, buf in enumerate(buffers_not_rg)
}
buffer_copies_rg = _broadcast_coalesced_reshape(buffers_rg,
devices,
detach=detach)
buffer_copies_not_rg = _broadcast_coalesced_reshape(buffers_not_rg,
devices,
detach=True)
modules = list(network.modules())
module_copies = [[] for device in devices]
module_indices = {}
scriptmodule_skip_attr = {
"_parameters", "_buffers", "_modules", "forward", "_c"
}
for i, module in enumerate(modules):
module_indices[module] = i
for j in range(num_replicas):
if _is_script_module(module):
# we have to initialize ScriptModule properly so that
# it works with pybind11
def init_fn(script_module):
# Don't do anything here, we'll initialize the ScriptModule below
return
replica = torch.jit.RecursiveScriptModule._construct(
module._c._replicate_for_data_parallel(), init_fn)
else:
replica = module._replicate_for_data_parallel()
# This is a temporary fix for DDP. DDP needs to access the
# replicated model parameters. It used to do so through
# `mode.parameters()`. The fix added in #33907 for DP stops the
# `parameters()` API from exposing the replicated parameters.
# Hence, we add a `_former_parameters` dict here to support DDP.
replica._former_parameters = OrderedDict()
module_copies[j].append(replica)
for i, module in enumerate(modules):
for key, child in module._modules.items():
if child is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = None
else:
module_idx = module_indices[child]
for j in range(num_replicas):
replica = module_copies[j][i]
setattr(replica, key, module_copies[j][module_idx])
for key, param in module._parameters.items():
if param is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = None
else:
param_idx = param_indices[param]
for j in range(num_replicas):
replica = module_copies[j][i]
param = param_copies[j][param_idx]
# parameters in replicas are no longer leaves, so remove them from _parameters
# and setattr them as non-parameter attributes
# scripted modules don't allow deleting parameters, but also don't complain
# on assigning non-Parameter type
if (not _is_script_module(replica)):
del replica._parameters[key]
setattr(replica, key, param)
# expose the parameter for DDP
replica._former_parameters[key] = param
for key, buf in module._buffers.items():
if buf is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = None
else:
if buf.requires_grad and not detach:
buffer_copies = buffer_copies_rg
buffer_idx = buffer_indices_rg[buf]
else:
buffer_copies = buffer_copies_not_rg
buffer_idx = buffer_indices_not_rg[buf]
for j in range(num_replicas):
replica = module_copies[j][i]
setattr(replica, key, buffer_copies[j][buffer_idx])
return [module_copies[j][0] for j in range(num_replicas)]
def replicate(network, devices, detach=False):
if not _replicatable_module(network):
raise RuntimeError("Cannot replicate network where python modules are "
"childrens of ScriptModule")
devices = list(map(lambda x: _get_device_index(x, True), devices))
num_replicas = len(devices)
params = list(network.parameters())
param_indices = {param: idx for idx, param in enumerate(params)}
param_copies = _broadcast_coalesced_reshape(params, devices, detach)
buffers = list(network.buffers())
buffers_rg = []
buffers_not_rg = []
for buf in buffers:
if buf.requires_grad and not detach:
buffers_rg.append(buf)
else:
buffers_not_rg.append(buf)
buffer_indices_rg = {buf: idx for idx, buf in enumerate(buffers_rg)}
buffer_indices_not_rg = {buf: idx for idx, buf in enumerate(buffers_not_rg)}
buffer_copies_rg = _broadcast_coalesced_reshape(buffers_rg, devices, detach=detach)
buffer_copies_not_rg = _broadcast_coalesced_reshape(buffers_not_rg, devices, detach=True)
modules = list(network.modules())
module_copies = [[] for device in devices]
module_indices = {}
scriptmodule_skip_attr = {"_parameters", "_buffers", "_modules"}
for i, module in enumerate(modules):
module_indices[module] = i
for j in range(num_replicas):
if _is_script_module(module):
# we have to initialize ScriptModule properly so that
# it works with pybind11
replica = _init_script_module()
keys = set(module.__dict__.keys()) - scriptmodule_skip_attr
for key in keys:
replica.__dict__[key] = module.__dict__[key]
else:
replica = module.__new__(type(module))
replica.__dict__ = module.__dict__.copy()
replica._parameters = replica._parameters.copy()
replica._buffers = replica._buffers.copy()
replica._modules = replica._modules.copy()
module_copies[j].append(replica)
for i, module in enumerate(modules):
for key, child in module._modules.items():
if child is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = None
else:
module_idx = module_indices[child]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = module_copies[j][module_idx]
for key, param in module._parameters.items():
if param is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = None
else:
param_idx = param_indices[param]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = param_copies[j][param_idx]
for key, buf in module._buffers.items():
if buf is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = None
else:
if buf.requires_grad and not detach:
buffer_copies = buffer_copies_rg
buffer_idx = buffer_indices_rg[buf]
else:
buffer_copies = buffer_copies_not_rg
buffer_idx = buffer_indices_not_rg[buf]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = buffer_copies[j][buffer_idx]
for j in range(num_replicas):
_copy_scriptmodule_methods(modules, module_copies[j], module_indices)
return [module_copies[j][0] for j in range(num_replicas)]
def criterion_parallel_apply(modules,
inputs,
targets,
kwargs_tup=None,
devices=None):
assert len(modules) == len(inputs)
assert len(targets) == len(inputs)
if kwargs_tup is not None:
assert len(modules) == len(kwargs_tup)
else:
kwargs_tup = ({}, ) * len(modules)
if devices is not None:
assert len(modules) == len(devices)
else:
devices = [None] * len(modules)
devices = list(map(lambda x: _get_device_index(x, True), devices))
lock = threading.Lock()
results = {}
grad_enabled = torch.is_grad_enabled()
def _worker(i, module, input, target, kwargs, device=None):
torch.set_grad_enabled(grad_enabled)
if device is None:
device = get_a_var(input).get_device()
try:
with torch.cuda.device(device):
if not isinstance(input, (list, tuple)):
input = (input, )
if not isinstance(target, (list, tuple)):
target = (target, )
output = module(*input, *target, **kwargs)
with lock:
results[i] = output
except Exception:
with lock:
results[i] = ExceptionWrapper(
where="in replica {} on device {}".format(i, device))
if len(modules) > 1:
threads = [
threading.Thread(target=_worker,
args=(i, module, input, target, kwargs, device))
for i, (
module, input, target, kwargs,
device) in enumerate(zip(modules, inputs, kwargs_tup, devices))
]
for thread in threads:
thread.start()
for thread in threads:
thread.join()
else:
_worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0])
outputs = []
for i in range(len(inputs)):
output = results[i]
if isinstance(output, ExceptionWrapper):
output.reraise()
outputs.append(output)
return outputs
def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):
r"""Applies each `module` in :attr:`modules` in parallel on arguments
contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword)
on each of :attr:`devices`.
Args:
modules (Module): modules to be parallelized
inputs (tensor): inputs to the modules
devices (list of int or torch.device): CUDA devices
:attr:`modules`, :attr:`inputs`, :attr:`kwargs_tup` (if given), and
:attr:`devices` (if given) should all have same length. Moreover, each
element of :attr:`inputs` can either be a single object as the only argument
to a module, or a collection of positional arguments.
"""
assert len(modules) == len(inputs)
if kwargs_tup is not None:
assert len(modules) == len(kwargs_tup)
else:
kwargs_tup = ({}, ) * len(modules)
if devices is not None:
assert len(modules) == len(devices)
else:
devices = [None] * len(modules)
devices = list(map(lambda x: _get_device_index(x, True), devices))
lock = threading.Lock()
results = {}
grad_enabled = torch.is_grad_enabled()
def _worker(i, module, input, kwargs, device=None):
torch.set_grad_enabled(grad_enabled)
if device is None:
device = get_a_var(input).get_device()
try:
with torch.cuda.device(device):
# this also avoids accidental slicing of `input` if it is a Tensor
#if not isinstance(input, (list, tuple)):
# input = (input,)
output = module(input)
with lock:
results[i] = output
except Exception:
with lock:
results[i] = ExceptionWrapper(
where="in replica {} on device {}".format(i, device))
if len(modules) > 1:
threads = [
threading.Thread(target=_worker,
args=(i, module, input, kwargs, device))
for i, (
module, input, kwargs,
device) in enumerate(zip(modules, inputs, kwargs_tup, devices))
]
for thread in threads:
thread.start()
for thread in threads:
thread.join()
else:
_worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0])
outputs = []
for i in range(len(inputs)):
output = results[i]
if isinstance(output, ExceptionWrapper):
output.reraise()
outputs.append(output)
return outputs
def replicate(network, devices, detach=False):
from ._functions import Broadcast
if not _replicatable_module(network):
raise RuntimeError("Cannot replicate network where python modules are "
"childrens of ScriptModule")
devices = list(map(lambda x: _get_device_index(x, True), devices))
num_replicas = len(devices)
params = list(network.parameters())
param_indices = {param: idx for idx, param in enumerate(params)}
param_copies = Broadcast.apply(devices, *params)
if len(params) > 0:
param_copies = [
param_copies[i:i + len(params)]
for i in range(0, len(param_copies), len(params))
]
buffers = list(network.buffers())
buffer_indices = {buf: idx for idx, buf in enumerate(buffers)}
buffer_copies = comm.broadcast_coalesced(buffers, devices)
modules = list(network.modules())
module_copies = [[] for device in devices]
module_indices = {}
scriptmodule_skip_attr = {"_parameters", "_buffers", "_modules"}
for i, module in enumerate(modules):
module_indices[module] = i
for j in range(num_replicas):
if _is_script_module(module):
# we have to initialize ScriptModule properly so that
# it works with pybind11
replica = _init_script_module()
keys = set(module.__dict__.keys()) - scriptmodule_skip_attr
for key in keys:
replica.__dict__[key] = module.__dict__[key]
else:
replica = module.__new__(type(module))
replica.__dict__ = module.__dict__.copy()
replica._parameters = replica._parameters.copy()
replica._buffers = replica._buffers.copy()
replica._modules = replica._modules.copy()
module_copies[j].append(replica)
for i, module in enumerate(modules):
for key, child in module._modules.items():
if child is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = None
else:
module_idx = module_indices[child]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = module_copies[j][module_idx]
for key, param in module._parameters.items():
if param is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = None
else:
param_idx = param_indices[param]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = param_copies[j][param_idx].detach() \
if detach else param_copies[j][param_idx]
for key, buf in module._buffers.items():
if buf is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = None
else:
buffer_idx = buffer_indices[buf]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = buffer_copies[j][buffer_idx]
for j in range(num_replicas):
_copy_scriptmodule_methods(modules, module_copies[j], module_indices)
return [module_copies[j][0] for j in range(num_replicas)]
def replicate(network, devices, detach=False):
if not _replicatable_module(network):
raise RuntimeError(
"Cannot replicate network where python modules are "
"childrens of ScriptModule"
)
devices = list(map(lambda x: _get_device_index(x, True), devices))
num_replicas = len(devices)
params = list(network.parameters())
param_indices = {param: idx for idx, param in enumerate(params)}
param_copies = _broadcast_coalesced_reshape(params, devices, detach)
buffers = list(network.buffers())
buffers_rg = []
buffers_not_rg = []
for buf in buffers:
if buf.requires_grad and not detach:
buffers_rg.append(buf)
else:
buffers_not_rg.append(buf)
buffer_indices_rg = {buf: idx for idx, buf in enumerate(buffers_rg)}
buffer_indices_not_rg = {buf: idx for idx, buf in enumerate(buffers_not_rg)}
buffer_copies_rg = _broadcast_coalesced_reshape(
buffers_rg, devices, detach=detach
)
buffer_copies_not_rg = _broadcast_coalesced_reshape(
buffers_not_rg, devices, detach=True
)
modules = list(network.modules())
module_copies = [[] for _ in devices]
module_indices = {}
scriptmodule_skip_attr = {
"_parameters",
"_buffers",
"_modules",
"forward",
"_c",
}
for i, module in enumerate(modules):
module_indices[module] = i
for j in range(num_replicas):
if _is_script_module(module):
# we have to initialize ScriptModule properly so that
# it works with pybind11
replica = _init_script_module()
attribute_names = set(
entry[0] for entry in module._c._get_attributes()
)
keys = (
set(module.__dict__.keys())
- scriptmodule_skip_attr
- attribute_names
)
for key in keys:
if not _is_script_method(module.__dict__[key]):
replica.__dict__[key] = module.__dict__[key]
for name, the_type, value in module._c._get_attributes():
if name in module._buffers.keys():
continue
replica._c._register_attribute(name, the_type, value)
else:
replica = module.__new__(type(module))
replica.__dict__ = module.__dict__.copy()
replica._parameters = replica._parameters.copy()
replica._buffers = replica._buffers.copy()
replica._modules = replica._modules.copy()
module_copies[j].append(replica)
for i, module in enumerate(modules):
for key, child in module._modules.items():
if child is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = None
else:
module_idx = module_indices[child]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = module_copies[j][module_idx]
for key, param in module._parameters.items():
if param is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = None
else:
param_idx = param_indices.get(param, None)
if param_idx is None:
continue
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = param_copies[j][param_idx]
for key, buf in module._buffers.items():
if buf is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = None
else:
if buf.requires_grad and not detach:
buffer_copies = buffer_copies_rg
buffer_idx = buffer_indices_rg.get(buf, None)
else:
buffer_copies = buffer_copies_not_rg
buffer_idx = buffer_indices_not_rg.get(buf, None)
if buffer_idx is None:
continue
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = buffer_copies[j][buffer_idx]
for j in range(num_replicas):
_copy_scriptmodule_methods(modules, module_copies[j], module_indices)
replicas = [module_copies[j][0] for j in range(num_replicas)]
for model_replica in replicas:
for _, submodule in model_replica.named_modules():
if hasattr(submodule, "on_replicate") and callable(
submodule.on_replicate
):
submodule.on_replicate()
return replicas
def __init__(self, module, device_ids=None,
broadcast_buffers=True,
compression=Compression.none
):
super(DistributedDataParallel, self).__init__()
assert device_ids and len(device_ids) == 1, (
"DistributedDataParallel device_ids contain exactlyone entry,"
" but got {}.").format(device_ids)
self.device_ids = list(map(lambda x: _get_device_index(x, True), device_ids))
self.module = module
self.broadcast_buffers = broadcast_buffers
self.require_forward_param_sync = broadcast_buffers
self._handles = {}
self._grad_accs = []
self._requires_update = set()
self._num_grads = 1
self.modules_buffers = [list(self.module.buffers())]
self._compression = compression
self._enable_async = False
self._require_backward_grad_sync = True
named_parameters = self.module.named_parameters()
named_parameters = list(named_parameters)
if len(named_parameters) > 0:
if isinstance(named_parameters[0][1], torch.Tensor):
if any([not isinstance(p, torch.Tensor) for name, p in named_parameters]):
raise ValueError('named_parameters should consistently be a sequence of '
'tuples (name, torch.Tensor)')
self._is_tensor_instance = True
# there is an issue when using torch.Tensor as key, so use its hash instead
# https://github.com/pytorch/pytorch/issues/7733
self._parameter_names = {v.__hash__(): k for k, v
in sorted(named_parameters)}
self._tensor_list = [tensor for name, tensor in named_parameters]
else:
self._is_tensor_instance = False
self._parameter_names = {v: k for k, v
in sorted(named_parameters)}
else:
self._is_tensor_instance = False
self._parameter_names = {v: 'push_pull.noname.%s' % i
for param_group in self.param_groups
for i, v in enumerate(param_group['params'])}
if size() > 1:
self._register_hooks()
named_params = self.module.named_parameters()
self._num_grads = sum(p.requires_grad for _, p in named_params)
byteps_torch_set_num_grads(self._num_grads)
# declare tensors
for name in sorted(self._parameter_names.values()):
declare("Gradient."+name)
# We use two loops for load-balancing
for name in sorted(self._parameter_names.values()):
declare("Parameter."+name)
# broadcast model state
module_states = list(self.module.state_dict().values())
if len(module_states) > 0:
bps.torch.broadcast_parameters(self.module.state_dict(), root_rank=0)
def replicate(network, devices, detach=False):
from ._functions import Broadcast
devices = list(map(lambda x: _get_device_index(x, True), devices))
num_replicas = len(devices)
params = list(network.parameters())
param_indices = {param: idx for idx, param in enumerate(params)}
param_copies = Broadcast.apply(devices, *params)
if len(params) > 0:
param_copies = [
param_copies[i:i + len(params)]
for i in range(0, len(param_copies), len(params))
]
buffers = list(network.buffers())
buffer_indices = {buf: idx for idx, buf in enumerate(buffers)}
buffer_copies = comm.broadcast_coalesced(buffers, devices)
modules = list(network.modules())
module_copies = [[] for device in devices]
module_indices = {}
for i, module in enumerate(modules):
module_indices[module] = i
for j in range(num_replicas):
replica = module.__new__(type(module))
replica.__dict__ = module.__dict__.copy()
replica._parameters = replica._parameters.copy()
replica._buffers = replica._buffers.copy()
replica._modules = replica._modules.copy()
module_copies[j].append(replica)
for i, module in enumerate(modules):
for key, child in module._modules.items():
if child is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = None
else:
module_idx = module_indices[child]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._modules[key] = module_copies[j][module_idx]
for key, param in module._parameters.items():
if param is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = None
else:
param_idx = param_indices[param]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._parameters[key] = param_copies[j][param_idx].detach() \
if detach else param_copies[j][param_idx]
for key, buf in module._buffers.items():
if buf is None:
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = None
else:
buffer_idx = buffer_indices[buf]
for j in range(num_replicas):
replica = module_copies[j][i]
replica._buffers[key] = buffer_copies[j][buffer_idx]
return [module_copies[j][0] for j in range(num_replicas)]
def parallel_apply_predict(modules, inputs, kwargs_tup=None, devices=None):
"""Applies each `module` predict method in `modules` in parallel on arguments
contained in `inputs` (positional) and `kwargs_tup` (keyword)
on each of `devices`.
Args:
modules: modules to be parallelized.
inputs: inputs to the modules.
devices: CUDA devices.
"""
assert len(modules) == len(inputs)
if kwargs_tup is not None:
assert len(modules) == len(kwargs_tup)
else:
kwargs_tup = ({},) * len(modules)
if devices is not None:
assert len(modules) == len(devices)
else:
devices = [None] * len(modules)
devices = list(map(lambda x: _get_device_index(x, True), devices))
lock = threading.Lock()
results = {}
grad_enabled = torch.is_grad_enabled()
def _worker(i, module, input, kwargs, device=None):
torch.set_grad_enabled(grad_enabled)
if device is None:
device = get_a_var(input).get_device()
try:
with torch.cuda.device(device):
# this also avoids accidental slicing of `input` if it is a Tensor
if not isinstance(input, (list, tuple)):
input = (input,)
output = module.predict(*input, **kwargs)
with lock:
results[i] = output
except Exception:
with lock:
results[i] = ExceptionWrapper(where="in replica {} on device {}".format(i, device))
if len(modules) > 1:
threads = [
threading.Thread(target=_worker, args=(i, module, input, kwargs, device))
for i, (module, input, kwargs, device) in enumerate(zip(modules, inputs, kwargs_tup, devices))
]
for thread in threads:
thread.start()
for thread in threads:
thread.join()
else:
_worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0])
outputs = []
for i in range(len(inputs)):
output = results[i]
if isinstance(output, ExceptionWrapper):
output.reraise()
outputs.append(output)
return outputs
def __init__(self,
module,
device_ids=None,
output_device=None,
dim=0,
broadcast_buffers=True,
process_group=None,
bucket_cap_mb=25):
super(DistributedDataParallel, self).__init__()
# Use all devices by default
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
if output_device is None:
output_device = device_ids[0]
if process_group is None:
self.process_group = dist.get_default_group()
else:
self.process_group = process_group
self.dim = dim
self.module = module
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
self.output_device = _get_device_index(output_device, True)
self.broadcast_buffers = broadcast_buffers
MB = 1024 * 1024
# used for intra-node param sync and inter-node sync as well
self.broadcast_bucket_size = 250 * MB
# Sync params and buffers
module_states = list(self.module.state_dict().values())
if len(module_states) > 0:
self._dist_broadcast_coalesced(module_states,
self.broadcast_bucket_size)
if len(device_ids) > 1:
# TODO: we don't need to replicate params in here. they're always going to
# be broadcasted using larger blocks in broadcast_coalesced, so it might be
# better to not pollute the caches with these small blocks
self._module_copies = replicate(self.module,
self.device_ids,
detach=True)
self._module_copies[0] = self.module
for module_copy in self._module_copies[1:]:
for param, copy_param in zip(self.module.parameters(),
module_copy.parameters()):
copy_param.requires_grad = param.requires_grad
else:
self._module_copies = [self.module]
self.modules_params_data = [[] for _ in range(len(self.device_ids))]
self.modules_buffers_data = [[] for _ in range(len(self.device_ids))]
for dev_idx, module in enumerate(self._module_copies):
self.modules_params_data[dev_idx] = [
p.data for p in module.parameters()
]
self.modules_buffers_data[dev_idx] = [
b.data for b in module.buffers()
]
bucket_bytes_cap = bucket_cap_mb * MB
# This is a triply-nested list where the "dimensions" are: devices, buckets, bucket_elems
param_buckets = []
# Split the parameters into buckets and by types as well
param_buckets = [
list(_take_tensors(m.parameters(), bucket_bytes_cap))
for m in self._module_copies
]
self.bucket_sizes = []
self.bucket_map = {}
# We transpose param_buckets, so the loop is over buckets.
# param_buckets_tuple is a doubly-nested list with "dims": devices, bucket_elems
for bucket_idx, param_buckets_tuple in enumerate(zip(*param_buckets)):
self.bucket_sizes.append(0)
# Now, we transpose again, so we iterate over bucket_elems, but getting tuples
# of params from each device.
for param_tuple in zip(*param_buckets_tuple):
if not param_tuple[0].requires_grad:
continue
for p in param_tuple:
self.bucket_map[p] = (bucket_idx,
self.bucket_sizes[bucket_idx])
self.bucket_sizes[bucket_idx] += 1
self.buckets = [[[None for _ in range(self.bucket_sizes[i])]
for _ in range(len(self.device_ids))]
for i in range(len(self.bucket_sizes))]
# The number of params ready in each bucket
self.buckets_ready_size = [[0 for _ in range(len(self.device_ids))]
for i in range(len(self.bucket_sizes))]
# coalesced bucket for only device 0
self.buckets_coalesced = [[] for _ in range(len(self.bucket_sizes))]
# We will always reduce the bucket following the reverse order
# that is, alway reduces following the order of: n - 1, n - 2, ..., 0
self.next_bucket = len(self.bucket_sizes) - 1
self.ready_buckets_not_reduced = set()
self.reduction_works = [None for _ in range(len(self.bucket_sizes))]
self.devs_ready = [0 for _ in range(len(self.bucket_sizes))]
self._register_grad_hooks()
def parallel_apply(modules,
inputs,
kwargs_tup=None,
devices=None): # pragma: no-cover
r"""Applies each `module` in :attr:`modules` in parallel on arguments
contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword)
on each of :attr:`devices`.
Args:
modules (Module): modules to be parallelized
inputs (tensor): inputs to the modules
devices (list of int or torch.device): CUDA devices
:attr:`modules`, :attr:`inputs`, :attr:`kwargs_tup` (if given), and
:attr:`devices` (if given) should all have same length. Moreover, each
element of :attr:`inputs` can either be a single object as the only argument
to a module, or a collection of positional arguments.
"""
assert len(modules) == len(inputs)
if kwargs_tup is not None:
assert len(modules) == len(kwargs_tup)
else:
kwargs_tup = ({}, ) * len(modules)
if devices is not None:
assert len(modules) == len(devices)
else:
devices = [None] * len(modules)
devices = list(map(lambda x: _get_device_index(x, True), devices))
lock = threading.Lock()
results = {}
grad_enabled = torch.is_grad_enabled()
def _worker(i, module, input, kwargs, device=None):
torch.set_grad_enabled(grad_enabled)
if device is None:
device = get_a_var(input).get_device()
try:
with torch.cuda.device(device):
# this also avoids accidental slicing of `input` if it is a Tensor
if not isinstance(input, (list, tuple)):
input = (input, )
# ---------------
# CHANGE
if module.training:
output = module.training_step(*input, **kwargs)
elif module.testing:
output = module.test_step(*input, **kwargs)
else:
output = module.validation_step(*input, **kwargs)
if module.use_dp or module.use_ddp2:
auto_squeeze_dim_zeros(output)
# ---------------
with lock:
results[i] = output
except Exception as ex:
with lock:
results[i] = ex
# TODO: fix hack (maybe not a hack)
# make sure each module knows what training state it's in...
# fixes weird bug where copies are out of sync
root_m = modules[0]
for m in modules[1:]:
m.training = root_m.training
m.testing = root_m.testing
if len(modules) > 1:
threads = [
threading.Thread(target=_worker,
args=(i, module, input, kwargs, device))
for i, (
module, input, kwargs,
device) in enumerate(zip(modules, inputs, kwargs_tup, devices))
]
for thread in threads:
thread.start()
for thread in threads:
thread.join()
else:
_worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0])
outputs = []
for i in range(len(inputs)):
output = results[i]
if isinstance(output, Exception):
raise output
outputs.append(output)
return outputs
def __init__(self,
module,
device_ids=None,
output_device=None,
dim=0,
broadcast_buffers=True,
process_group=None,
bucket_cap_mb=25,
find_unused_parameters=False,
check_reduction=False):
super(DistributedDataParallel, self).__init__()
self.is_multi_device_module = len(
{p.device
for p in module.parameters()}) > 1
self.is_cuda = all(
[p.device.type == 'cuda' for p in module.parameters()])
if not self.is_cuda or self.is_multi_device_module:
assert not device_ids and not output_device, (
"DistributedDataParallel device_ids and output_device arguments "
"only work with single-device CUDA modules, but got "
"device_ids {}, output_device {}, and module parameters {}."
).format(device_ids, output_device,
{p.device
for p in module.parameters()})
self.device_ids = None
self.output_device = None
else:
# Use all devices by default for single-device CUDA modules
if device_ids is None:
device_ids = list(range(torch.cuda.device_count()))
self.device_ids = list(
map(lambda x: _get_device_index(x, True), device_ids))
# if output_device is None:
# output_device = device_ids[0]
self.output_device = _get_device_index(output_device, True)
if self.is_multi_device_module:
assert self.is_cuda, (
"DistributedDataParallel with multi-device module only works "
"with CUDA devices, but module parameters locate in {}."
).format({p.device
for p in module.parameters()})
if process_group is None:
self.process_group = _get_default_group()
else:
self.process_group = process_group
self.dim = dim
self.module = module
self.broadcast_buffers = broadcast_buffers
self.find_unused_parameters = find_unused_parameters
self.require_backward_grad_sync = True
self.require_forward_param_sync = True
if check_reduction:
# This argument is no longer used since the reducer
# will ensure reduction completes even if some parameters
# do not receive gradients.
pass
MB = 1024 * 1024
# used for intra-node param sync and inter-node sync as well
self.broadcast_bucket_size = int(250 * MB)
# reduction bucket size
self.bucket_bytes_cap = int(bucket_cap_mb * MB)
# Sync params and buffers
module_states = list(self.module.state_dict().values())
if len(module_states) > 0:
self._distributed_broadcast_coalesced(module_states,
self.broadcast_bucket_size)
self._ddp_init_helper()
