def tpu_train(self, tpu_core_idx, model):
# put model on tpu
model.to(xm.xla_device())
# get the appropriate tpu ranks
self.tpu_local_core_rank = xm.get_local_ordinal()
self.tpu_global_core_rank = xm.get_ordinal()
# avoid duplicating progress bar
self.progress_bar_refresh_rate = self.progress_bar_refresh_rate if self.tpu_global_core_rank == 0 else 0
# track current tpu
self.current_tpu_idx = tpu_core_idx
self.proc_rank = self.tpu_local_core_rank
set_proc_rank(self.proc_rank)
# CHOOSE OPTIMIZER
# allow for lr schedulers as well
self.optimizers, self.lr_schedulers, self.optimizer_frequencies = self.init_optimizers(model)
# init 16 bit for TPU
if self.precision == 16:
os.environ['XLA_USE_BF16'] = str(1)
log.info(f'INIT TPU local core: {self.tpu_local_core_rank},'
f' global rank: {self.tpu_global_core_rank}')
# continue training routine
self.run_pretrain_routine(model)
self.save_spawn_weights(model)
def ddp_train(self, process_idx, model):
"""
Entry point into a DP thread
:param gpu_idx:
:param model:
:param cluster_obj:
:return:
"""
# node rank using relative slurm id
# otherwise default to node rank 0
try:
node_id = os.environ['SLURM_NODEID']
self.node_rank = int(node_id)
except Exception:
self.node_rank = 0
# show progressbar only on progress_rank 0
self.progress_bar_refresh_rate = (self.progress_bar_refresh_rate
if self.node_rank == 0
and process_idx == 0 else 0)
# determine which process we are and world size
if self.use_ddp:
self.proc_rank = self.node_rank * self.num_processes + process_idx
self.world_size = self.num_nodes * self.num_processes
elif self.use_ddp2:
self.proc_rank = self.node_rank
self.world_size = self.num_nodes
# set warning rank
set_proc_rank(self.proc_rank)
# let the exp know the rank to avoid overwriting logs
if self.logger is not None:
self.logger.rank = self.proc_rank
# set up server using proc 0's ip address
# try to init for 20 times at max in case ports are taken
# where to store ip_table
model.trainer = self
model.init_ddp_connection(self.proc_rank, self.world_size)
# CHOOSE OPTIMIZER
# allow for lr schedulers as well
self.optimizers, self.lr_schedulers, self.optimizer_frequencies = self.init_optimizers(
model)
# MODEL
# copy model to each gpu
if self.on_gpu:
self.root_gpu = self.data_parallel_device_ids[process_idx]
torch.cuda.set_device(self.root_gpu)
model.cuda(self.root_gpu)
# set model properties before going into wrapper
self.copy_trainer_model_properties(model)
# AMP
# run through amp wrapper before going to distributed DP
if self.use_amp:
# An example
model, optimizers = model.configure_apex(amp, model,
self.optimizers,
self.amp_level)
self.optimizers = optimizers
# DDP2 uses all GPUs on the machine
if self.distributed_backend == 'ddp':
device_ids = [self.root_gpu]
elif self.use_ddp2:
device_ids = self.data_parallel_device_ids
else: # includes ddp_cpu
device_ids = None
# allow user to configure ddp
model = model.configure_ddp(model, device_ids)
# continue training routine
self.run_pretrain_routine(model)
# when ddp ends, we save the model
self.save_spawn_weights(model)
def ddp_train(self, process_idx, model):
"""
Entry point into a DP thread
:param gpu_idx:
:param model:
:param cluster_obj:
:return:
"""
# node rank using relative slurm id if under slurm management
# otherwise use given node rank or default to node rank 0
try:
node_id = os.environ[
'SLURM_NODEID'] if self.is_slurm_managing_tasks else os.environ[
'NODE_RANK']
self.node_rank = int(node_id)
except KeyError:
log.warning(
"SLURM_NODEID or NODE_RANK environment variable is not defined. Set as 0."
)
self.node_rank = 0
# show progressbar only on progress_rank 0
if (self.node_rank != 0 or
process_idx != 0) and self.progress_bar_callback is not None:
self.progress_bar_callback.disable()
# determine which process we are and world size
if self.use_ddp:
self.proc_rank = self.node_rank * self.num_processes + process_idx
self.world_size = self.num_nodes * self.num_processes
elif self.use_ddp2:
self.proc_rank = self.node_rank
self.world_size = self.num_nodes
# set warning rank
set_proc_rank(self.proc_rank)
# let the exp know the rank to avoid overwriting logs
if self.logger is not None:
self.logger.rank = self.proc_rank
# set up server using proc 0's ip address
# try to init for 20 times at max in case ports are taken
# where to store ip_table
model.trainer = self
model.init_ddp_connection(self.proc_rank, self.world_size,
self.is_slurm_managing_tasks)
# CHOOSE OPTIMIZER
# allow for lr schedulers as well
self.optimizers, self.lr_schedulers, self.optimizer_frequencies = self.init_optimizers(
model)
# MODEL
# copy model to each gpu
if self.on_gpu:
self.root_gpu = process_idx
torch.cuda.set_device(self.root_gpu)
model.cuda(self.root_gpu)
# set model properties before going into wrapper
self.copy_trainer_model_properties(model)
# AMP
# run through amp wrapper before going to distributed DP
# TODO: remove in v0.8.0
if self.use_amp and not self.use_native_amp:
model, optimizers = model.configure_apex(amp, model,
self.optimizers,
self.amp_level)
self.optimizers = optimizers
# DDP2 uses all GPUs on the machine
if self.distributed_backend == 'ddp':
device_ids = [self.root_gpu]
elif self.use_ddp2:
device_ids = self.data_parallel_device_ids
else: # includes ddp_cpu
device_ids = None
# allow user to configure ddp
model = model.configure_ddp(model, device_ids)
# continue training routine
self.run_pretrain_routine(model)
# when ddp ends, we save the model
self.save_spawn_weights(model)
def horovod_train(self, model):
# Horovod: initialize library
hvd.init()
if torch.cuda.is_available() and self.on_gpu:
# Horovod: pin GPU to local rank
torch.cuda.set_device(hvd.local_rank())
model.cuda(hvd.local_rank())
# Only show progress bar from the first worker
self.progress_bar_refresh_rate = self.progress_bar_refresh_rate if hvd.rank(
) == 0 else 0
# CHOOSE OPTIMIZER
# allow for lr schedulers as well
self.optimizers, self.lr_schedulers, self.optimizer_frequencies = self.init_optimizers(
model)
# Horovod: scale the learning rate by the number of workers to account for
# increased total batch size
for optimizer in self.optimizers:
for param_group in optimizer.param_groups:
param_group['lr'] *= hvd.size()
if self.use_amp:
# An example
model, optimizers = model.configure_apex(amp, model,
self.optimizers,
self.amp_level)
self.optimizers = optimizers
# Horovod: broadcast parameters & optimizer state to ensure consistent initialization
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
for optimizer in self.optimizers:
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
def filter_named_parameters(model, optimizer):
opt_params = set([
p for group in optimizer.param_groups
for p in group.get('params', [])
])
return [(name, p) for name, p in model.named_parameters()
if p in opt_params]
# Horovod: wrap optimizers to perform gradient aggregation via allreduce
self.optimizers = [
hvd.DistributedOptimizer(optimizer,
named_parameters=filter_named_parameters(
model, optimizer))
for optimizer in self.optimizers
]
# Update logger rank info from Horovod to avoid race conditions from different ranks
# creating directories / writing files in the same locations.
self.proc_rank = hvd.rank()
set_proc_rank(self.proc_rank)
if self.logger:
self.logger.rank = self.proc_rank
if model.logger:
model.logger.rank = self.proc_rank
with ExitStack() as stack:
for optimizer in self.optimizers:
# Synchronization will be performed explicitly following backward()
stack.enter_context(optimizer.skip_synchronize())
self.run_pretrain_routine(model)
