def model(self):
if self._wrapped_model is None:
if (
self.data_parallel_world_size > 1
and not self.args.use_bmuf
and not self.tpu
):
self._wrapped_model = models.DistributedFairseqModel(
self.args, self._model,
process_group=self.data_parallel_process_group
)
else:
self._wrapped_model = self._model
return self._wrapped_model
def criterion(self):
if self._wrapped_criterion is None:
if (
utils.has_parameters(self._criterion)
and self.data_parallel_world_size > 1
and not self.args.use_bmuf
):
self._wrapped_criterion = models.DistributedFairseqModel(
self.args, self._criterion,
process_group=self.data_parallel_process_group
)
else:
self._wrapped_criterion = self._criterion
return self._wrapped_criterion
def model(self):
if self._wrapped_model is None:
if self.args.distributed_world_size > 1:
self._wrapped_model = models.DistributedFairseqModel(
self.args,
self._model,
)
else:
self._wrapped_model = self._model
if self.args.kd:
for value in self._wrapped_model.distill_models.values():
value.to(
next(self._wrapped_model.encoder.parameters()).device)
return self._wrapped_model
def model(self):
if self._wrapped_model is None:
if (
self.data_parallel_world_size > 1
and not self.cfg.optimization.use_bmuf
and not self.tpu
):
self._wrapped_model = models.DistributedFairseqModel(
self.cfg.distributed_training,
self._model,
process_group=self.data_parallel_process_group,
)
else:
self._wrapped_model = self._model
return self._wrapped_model
def criterion(self):
if self._wrapped_criterion is None:
if (
utils.has_parameters(self._criterion)
and self.use_distributed_wrapper
):
self._wrapped_criterion = models.DistributedFairseqModel(
self.cfg.distributed_training,
self._criterion,
process_group=self.data_parallel_process_group,
device=self.device,
)
else:
self._wrapped_criterion = self._criterion
return self._wrapped_criterion
def criterion(self):
if self._wrapped_criterion is None:
if (
utils.has_parameters(self._criterion)
and self.data_parallel_world_size > 1
and not self.cfg.optimization.use_bmuf
and not self.tpu
):
self._wrapped_criterion = models.DistributedFairseqModel(
self.cfg.distributed_training,
self._criterion,
process_group=self.data_parallel_process_group,
)
else:
self._wrapped_criterion = self._criterion
return self._wrapped_criterion
