def test_manual_poptorch_opts_train_grad_accum(tmpdir):
"""
Ensure if the user passes manual poptorch Options
and grad accumulation differs to accumulate_grad_batches, we
"""
model = IPUModel()
inference_opts = poptorch.Options()
inference_opts.Training.gradientAccumulation(1)
training_opts = poptorch.Options()
training_opts.Training.gradientAccumulation(2)
trainer = Trainer(default_root_dir=tmpdir,
ipus=1,
fast_dev_run=True,
accumulate_grad_batches=1,
plugins=IPUPlugin(inference_opts=inference_opts,
training_opts=training_opts))
with pytest.warns(
UserWarning,
match=f"Training poptorch.Options set gradientAccumulation to {2}. "
f"This is different to accumulate_grad_batches which was set to {1}. "
f"To change gradientAccumulation, please set accumulate_grad_batches in the Trainer. "
f"Setting poptorch.Options gradientAccumulation to {1}",
):
trainer.fit(model)
assert isinstance(trainer.accelerator.training_type_plugin, IPUPlugin)
assert trainer.accelerator.training_type_plugin.inference_opts.Training.gradient_accumulation == 1
def test_replication_factor(tmpdir):
"""Ensure if the user passes manual poptorch Options with custom parameters set, we set them correctly in the
dataloaders."""
plugin = IPUPlugin()
trainer = Trainer(ipus=2, default_root_dir=tmpdir, fast_dev_run=True, plugins=plugin)
assert trainer.ipus == 2
def test_manual_poptorch_opts_inference_grad_accum(tmpdir):
"""
Ensure if the user passes manual poptorch Options
and grad accumulation is set greater than 1 for inference, we warn and set to 1.
"""
model = IPUModel()
inference_opts = poptorch.Options()
inference_opts.Training.gradientAccumulation(4)
training_opts = poptorch.Options()
training_opts.Training.gradientAccumulation(1)
trainer = Trainer(default_root_dir=tmpdir,
ipus=1,
fast_dev_run=True,
plugins=IPUPlugin(inference_opts=inference_opts,
training_opts=training_opts))
with pytest.warns(
UserWarning,
match=
"Inference poptorch.Options should set gradientAccumulation to 1. "
"Setting gradientAccumulation to 1 for inference options.",
):
trainer.fit(model)
assert isinstance(trainer.accelerator.training_type_plugin, IPUPlugin)
assert trainer.accelerator.training_type_plugin.inference_opts.Training.gradient_accumulation == 1
def test_manual_poptorch_opts_ipu_count(tmpdir):
"""
Ensure if the user passes manual poptorch Options
and the number of ipus do not match, we warn and we set it for the user.
"""
manual_ipus = 1
expected_ipus = 2
model = IPUModel()
inference_opts = poptorch.Options()
inference_opts.replicationFactor(manual_ipus)
training_opts = poptorch.Options()
training_opts.replicationFactor(manual_ipus)
trainer = Trainer(default_root_dir=tmpdir,
ipus=expected_ipus,
fast_dev_run=True,
plugins=IPUPlugin(inference_opts=inference_opts,
training_opts=training_opts))
with pytest.warns(
UserWarning,
match=
f"Manual poptorch.Options set replicationFactor to {manual_ipus} "
f"which differs to the ipus={expected_ipus} flag passed to the Trainer. "
f"Setting to {expected_ipus} in the poptorch.Options."):
trainer.fit(model)
assert isinstance(trainer.accelerator.training_type_plugin, IPUPlugin)
assert trainer.accelerator.training_type_plugin.training_opts.replication_factor == 2
assert trainer.accelerator.training_type_plugin.inference_opts.replication_factor == 2
def test_manual_poptorch_opts_custom(tmpdir):
"""
Ensure if the user passes manual poptorch Options with custom parameters set,
we respect them in our poptorch options and the dataloaders.
"""
model = IPUModel()
training_opts = poptorch.Options()
training_opts.deviceIterations(8)
training_opts.replicationFactor(2)
training_opts.Training.gradientAccumulation(2)
inference_opts = poptorch.Options()
inference_opts.deviceIterations(16)
inference_opts.replicationFactor(1)
inference_opts.Training.gradientAccumulation(1)
class TestCallback(Callback):
def on_fit_end(self, trainer: Trainer,
pl_module: LightningModule) -> None:
# ensure dataloaders were correctly set up during training.
plugin = trainer.accelerator.training_type_plugin
assert isinstance(plugin, IPUPlugin)
assert plugin.training_opts.replication_factor == 2
assert plugin.inference_opts.replication_factor == 1
val_dataloader = trainer.val_dataloaders[0]
train_dataloader = trainer.train_dataloader
assert isinstance(train_dataloader, CombinedLoader)
train_dataloader = train_dataloader.loaders
assert isinstance(val_dataloader, poptorch.DataLoader)
assert isinstance(train_dataloader, poptorch.DataLoader)
assert train_dataloader.options.replication_factor == 2
assert val_dataloader.options.replication_factor == 1
plugin = IPUPlugin(inference_opts=inference_opts,
training_opts=training_opts)
# ensure we default to the training options replication factor
assert plugin.replication_factor == 2
trainer = Trainer(default_root_dir=tmpdir,
fast_dev_run=True,
plugins=plugin,
callbacks=TestCallback())
trainer.fit(model)
plugin = trainer.accelerator.training_type_plugin
assert isinstance(plugin, IPUPlugin)
training_opts = plugin.training_opts
assert training_opts.device_iterations == 8
assert training_opts.replication_factor == 2
assert training_opts.Training.gradient_accumulation == 2
inference_opts = plugin.inference_opts
assert inference_opts.device_iterations == 16
assert inference_opts.replication_factor == 1
assert inference_opts.Training.gradient_accumulation == 1
def test_replication_factor(tmpdir):
"""Ensure if the user passes manual poptorch Options with custom parameters set, we set them correctly in the
dataloaders."""
plugin = IPUPlugin()
trainer = Trainer(ipus=2,
default_root_dir=tmpdir,
fast_dev_run=True,
strategy=plugin)
assert trainer.ipus == 2
assert trainer.training_type_plugin.replication_factor == 2
model = BoringModel()
training_opts = poptorch.Options()
inference_opts = poptorch.Options()
training_opts.replicationFactor(8)
inference_opts.replicationFactor(7)
plugin = IPUPlugin(inference_opts=inference_opts,
training_opts=training_opts)
trainer = Trainer(default_root_dir=tmpdir, ipus=1, strategy=plugin)
trainer.optimizers = model.configure_optimizers()[0]
plugin.model = model
model.trainer = trainer
trainer.state.fn = TrainerFn.FITTING
trainer.training_type_plugin.pre_dispatch()
trainer.state.stage = RunningStage.TRAINING
assert trainer.training_type_plugin.replication_factor == 8
trainer.state.stage = RunningStage.VALIDATING
assert trainer.training_type_plugin.replication_factor == 7
for fn, stage in (
(TrainerFn.VALIDATING, RunningStage.VALIDATING),
(TrainerFn.TESTING, RunningStage.TESTING),
(TrainerFn.PREDICTING, RunningStage.PREDICTING),
):
trainer.state.fn = fn
trainer.state.stage = stage
trainer.training_type_plugin.pre_dispatch()
assert trainer.training_type_plugin.replication_factor == 7
def test_autoreport(tmpdir):
"""Ensure autoreport dumps to a file."""
model = IPUModel()
autoreport_path = os.path.join(tmpdir, 'report/')
trainer = Trainer(default_root_dir=tmpdir,
ipus=1,
fast_dev_run=True,
plugins=IPUPlugin(autoreport=True,
autoreport_dir=autoreport_path))
trainer.fit(model)
assert os.path.exists(autoreport_path)
assert os.path.isfile(autoreport_path + 'profile.pop')
def test_manual_poptorch_opts(tmpdir):
"""Ensure if the user passes manual poptorch Options, we run with the correct object."""
model = IPUModel()
inference_opts = poptorch.Options()
training_opts = poptorch.Options()
trainer = Trainer(default_root_dir=tmpdir,
ipus=1,
fast_dev_run=True,
plugins=IPUPlugin(inference_opts=inference_opts,
training_opts=training_opts))
trainer.fit(model)
assert isinstance(trainer.accelerator.training_type_plugin, IPUPlugin)
assert trainer.accelerator.training_type_plugin.training_opts == training_opts
assert trainer.accelerator.training_type_plugin.inference_opts == inference_opts
def test_device_iterations_ipu_plugin(tmpdir):
class TestCallback(Callback):
def on_train_start(self, trainer: Trainer,
pl_module: LightningModule) -> None:
assert trainer.accelerator.training_type_plugin.device_iterations == 2
# assert device iterations has been set correctly within the poptorch options
poptorch_model = trainer.accelerator.training_type_plugin.poptorch_models[
RunningStage.TRAINING]
assert poptorch_model._options.toDict()['device_iterations'] == 2
raise SystemExit
model = IPUModel()
trainer = Trainer(default_root_dir=tmpdir,
fast_dev_run=True,
ipus=1,
plugins=IPUPlugin(device_iterations=2),
callbacks=TestCallback())
assert isinstance(trainer.accelerator.training_type_plugin, IPUPlugin)
with pytest.raises(SystemExit):
trainer.fit(model)
def test_poptorch_models_at_different_stages(tmpdir):
plugin = IPUPlugin()
trainer = Trainer(default_root_dir=tmpdir, strategy=plugin, ipus=8)
model = BoringModel()
model.trainer = trainer
plugin.model = model
trainer.optimizers = model.configure_optimizers()[0]
trainer.state.fn = TrainerFn.FITTING
trainer.training_type_plugin.pre_dispatch()
assert list(trainer.training_type_plugin.poptorch_models) == [
RunningStage.TRAINING, RunningStage.VALIDATING
]
for fn, stage in (
(TrainerFn.VALIDATING, RunningStage.VALIDATING),
(TrainerFn.TESTING, RunningStage.TESTING),
(TrainerFn.PREDICTING, RunningStage.PREDICTING),
):
trainer.state.fn = fn
trainer.state.stage = stage
trainer.training_type_plugin.pre_dispatch()
assert list(trainer.training_type_plugin.poptorch_models) == [stage]
def test_pure_half_precision(tmpdir):
class TestCallback(Callback):
def on_train_start(self, trainer: Trainer,
pl_module: LightningModule) -> None:
assert trainer.accelerator.model.precision == 16
assert trainer.accelerator.training_type_plugin.convert_model_to_half
for param in trainer.accelerator.model.parameters():
assert param.dtype == torch.float16
raise SystemExit
model = IPUModel()
trainer = Trainer(default_root_dir=tmpdir,
fast_dev_run=True,
ipus=1,
precision=16,
plugins=IPUPlugin(convert_model_to_half=True),
callbacks=TestCallback())
assert isinstance(trainer.accelerator.training_type_plugin, IPUPlugin)
assert isinstance(trainer.accelerator.precision_plugin, IPUPrecisionPlugin)
assert trainer.accelerator.precision_plugin.precision == 16
with pytest.raises(SystemExit):
trainer.fit(model)
def select_training_type_plugin(self) -> TrainingTypePlugin:
if isinstance(
self.distributed_backend, Accelerator
) and self.distributed_backend.training_type_plugin is not None:
plugin = self.distributed_backend.training_type_plugin
elif self.use_ddp2:
plugin = DDP2Plugin(
parallel_devices=self.parallel_devices,
cluster_environment=self.cluster_environment,
)
elif self.use_ddp and self.use_deepspeed:
plugin = DeepSpeedPlugin(
cluster_environment=self.select_cluster_environment(), parallel_devices=self.parallel_devices
)
elif self.use_ddp:
use_slurm_ddp = self.use_ddp and self.is_slurm_managing_tasks
use_torchelastic_ddp = self.use_ddp and TorchElasticEnvironment.is_using_torchelastic()
use_kubeflow_ddp = self.use_ddp and KubeflowEnvironment.is_using_kubeflow()
use_ddp_spawn = self._distrib_type == DistributedType.DDP_SPAWN
use_ddp_cpu_spawn = self.use_ddp and self.on_cpu
use_tpu_spawn = self.on_tpu and self._distrib_type == DistributedType.TPU_SPAWN
use_ddp_cpu_torch_elastic = use_ddp_cpu_spawn and TorchElasticEnvironment.is_using_torchelastic()
use_ddp_cpu_kubeflow = use_ddp_cpu_spawn and KubeflowEnvironment.is_using_kubeflow()
use_ddp_cpu_slurm = use_ddp_cpu_spawn and self.is_slurm_managing_tasks
use_ddp_sharded = self._distrib_type == DistributedType.DDP_SHARDED
use_ddp_sharded_spawn = self._distrib_type == DistributedType.DDP_SHARDED_SPAWN
use_ddp_fully_sharded = self._distrib_type == DistributedType.DDP_FULLY_SHARDED
# TODO: decouple from TE
# ddp script mode uses the same flags as TE
if os.environ.get("PL_IN_DDP_SUBPROCESS", False):
use_torchelastic_ddp = False
if use_tpu_spawn:
ddp_plugin_cls = TPUSpawnPlugin
elif use_ddp_sharded:
ddp_plugin_cls = DDPShardedPlugin
elif use_ddp_sharded_spawn:
ddp_plugin_cls = DDPSpawnShardedPlugin
elif (
use_ddp_cpu_slurm or use_slurm_ddp or use_ddp_cpu_torch_elastic or use_torchelastic_ddp
or use_kubeflow_ddp or use_ddp_cpu_kubeflow
):
ddp_plugin_cls = DDPPlugin
elif use_ddp_spawn or use_ddp_cpu_spawn:
ddp_plugin_cls = DDPSpawnPlugin
elif use_ddp_fully_sharded:
ddp_plugin_cls = DDPFullyShardedPlugin
else:
ddp_plugin_cls = DDPPlugin
plugin = ddp_plugin_cls(
parallel_devices=self.parallel_devices,
cluster_environment=self.cluster_environment,
)
elif self.use_dp:
plugin = DataParallelPlugin(parallel_devices=self.parallel_devices)
elif self.use_horovod:
plugin = HorovodPlugin(parallel_devices=self.parallel_devices)
elif self.on_tpu and isinstance(self.tpu_cores, list):
plugin = SingleTPUPlugin(self.tpu_id)
elif self.on_ipu:
plugin = IPUPlugin(parallel_devices=self.parallel_devices)
else:
single_gpu_ordinal = device_parser.determine_root_gpu_device(self.parallel_device_ids)
plugin = SingleDevicePlugin(device=torch.device(f"cuda:{single_gpu_ordinal}" if self.on_gpu else "cpu"))
return plugin
def test_no_warning_plugin(tmpdir):
with pytest.warns(None) as record:
Trainer(default_root_dir=tmpdir,
plugins=IPUPlugin(training_opts=poptorch.Options()))
assert len(record) == 0
def test_device_type_when_training_plugin_ipu_passed(tmpdir):
trainer = Trainer(strategy=IPUPlugin(), ipus=8)
assert isinstance(trainer.training_type_plugin, IPUPlugin)
assert trainer._device_type == DeviceType.IPU
assert isinstance(trainer.accelerator, IPUAccelerator)
def test_strategy_choice_ipu_plugin(tmpdir):
trainer = Trainer(strategy=IPUPlugin(), accelerator="ipu", devices=8)
assert isinstance(trainer.training_type_plugin, IPUPlugin)