348 lines
14 KiB
Python
348 lines
14 KiB
Python
# Copyright The PyTorch Lightning team.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
import io
|
|
import os
|
|
import time
|
|
from multiprocessing.queues import SimpleQueue
|
|
from typing import Any, Callable, Dict, List, Optional, Union
|
|
|
|
import torch
|
|
import torch.multiprocessing as mp
|
|
from torch.nn import Module
|
|
from torch.utils.data import DataLoader
|
|
|
|
import pytorch_lightning as pl
|
|
from pytorch_lightning.overrides import LightningDistributedModule
|
|
from pytorch_lightning.plugins.io.xla_plugin import XLACheckpointIO
|
|
from pytorch_lightning.plugins.precision import PrecisionPlugin
|
|
from pytorch_lightning.strategies.ddp_spawn import _FakeQueue, _SpawnOutput, DDPSpawnStrategy
|
|
from pytorch_lightning.trainer.connectors.data_connector import DataConnector
|
|
from pytorch_lightning.trainer.states import TrainerFn
|
|
from pytorch_lightning.utilities import _TPU_AVAILABLE, find_shared_parameters, set_shared_parameters
|
|
from pytorch_lightning.utilities.apply_func import move_data_to_device
|
|
from pytorch_lightning.utilities.data import has_len
|
|
from pytorch_lightning.utilities.distributed import rank_zero_debug, rank_zero_only, ReduceOp
|
|
from pytorch_lightning.utilities.exceptions import MisconfigurationException
|
|
from pytorch_lightning.utilities.model_helpers import is_overridden
|
|
from pytorch_lightning.utilities.seed import reset_seed
|
|
from pytorch_lightning.utilities.types import _PATH, STEP_OUTPUT
|
|
|
|
if _TPU_AVAILABLE:
|
|
import torch_xla.core.xla_env_vars as xenv
|
|
import torch_xla.core.xla_model as xm
|
|
import torch_xla.distributed.xla_multiprocessing as xmp
|
|
from torch_xla.core.xla_model import rendezvous
|
|
from torch_xla.distributed.parallel_loader import MpDeviceLoader
|
|
else:
|
|
xm, xmp, MpDeviceLoader, rendezvous = [None] * 4
|
|
|
|
|
|
class TPUSpawnStrategy(DDPSpawnStrategy):
|
|
"""Strategy for training multiple TPU devices using the :func:`torch.multiprocessing.spawn` method."""
|
|
|
|
def __init__(
|
|
self,
|
|
accelerator: Optional["pl.accelerators.accelerator.Accelerator"] = None,
|
|
parallel_devices: Optional[List[int]] = None,
|
|
checkpoint_io: Optional[XLACheckpointIO] = None,
|
|
precision_plugin: Optional[PrecisionPlugin] = None,
|
|
debug: bool = False,
|
|
**_: Any,
|
|
) -> None:
|
|
checkpoint_io = checkpoint_io or XLACheckpointIO()
|
|
super().__init__(
|
|
accelerator=accelerator,
|
|
parallel_devices=parallel_devices,
|
|
checkpoint_io=checkpoint_io,
|
|
precision_plugin=precision_plugin,
|
|
)
|
|
self.debug = debug
|
|
self.tpu_local_core_rank = 0
|
|
self.tpu_global_core_rank = 0
|
|
self.start_method = "fork"
|
|
|
|
@property
|
|
def global_rank(self) -> int:
|
|
return self.tpu_global_core_rank
|
|
|
|
@property
|
|
def local_rank(self) -> int:
|
|
return self.tpu_local_core_rank
|
|
|
|
@property
|
|
def world_size(self) -> int:
|
|
return xm.xrt_world_size()
|
|
|
|
@property
|
|
def root_device(self) -> torch.device:
|
|
return xm.xla_device()
|
|
|
|
@staticmethod
|
|
def _validate_dataloader(dataloaders: Union[List[DataLoader], DataLoader]) -> None:
|
|
if not isinstance(dataloaders, list):
|
|
dataloaders = [dataloaders]
|
|
|
|
for dataloader in dataloaders:
|
|
if not has_len(dataloader):
|
|
raise MisconfigurationException(
|
|
"TPUs do not currently support IterableDataset objects, the dataset must implement `__len__`."
|
|
" HINT: You can mock the length on your dataset to bypass this MisconfigurationException."
|
|
)
|
|
|
|
@staticmethod
|
|
def _validate_patched_dataloaders(model: "pl.LightningModule") -> None:
|
|
"""Validate and fail fast if the dataloaders were passed directly to fit."""
|
|
connector: DataConnector = model.trainer._data_connector
|
|
sources = (
|
|
connector._train_dataloader_source,
|
|
connector._val_dataloader_source,
|
|
connector._test_dataloader_source,
|
|
connector._predict_dataloader_source,
|
|
)
|
|
for source in sources:
|
|
if not source.is_module():
|
|
TPUSpawnStrategy._validate_dataloader(source.instance)
|
|
|
|
def connect(self, model: "pl.LightningModule") -> None:
|
|
TPUSpawnStrategy._validate_patched_dataloaders(model)
|
|
self.wrapped_model = xmp.MpModelWrapper(LightningDistributedModule(model))
|
|
return super().connect(model)
|
|
|
|
def setup(self, trainer: "pl.Trainer") -> None:
|
|
self.start_method = "fork"
|
|
self.accelerator.setup(trainer)
|
|
self.setup_optimizers(trainer)
|
|
self.setup_precision_plugin()
|
|
self._move_optimizer_state()
|
|
|
|
if self.debug:
|
|
os.environ["PT_XLA_DEBUG"] = str(1)
|
|
|
|
shared_params = find_shared_parameters(self.model)
|
|
self.model_to_device()
|
|
if is_overridden("on_post_move_to_device", self.lightning_module):
|
|
self.model.module.on_post_move_to_device()
|
|
else:
|
|
set_shared_parameters(self.model.module, shared_params)
|
|
|
|
self.setup_optimizers(trainer)
|
|
self.precision_plugin.connect(self.model, None, None)
|
|
|
|
def _setup_model(self, model: Module) -> Module:
|
|
return model
|
|
|
|
@property
|
|
def distributed_sampler_kwargs(self) -> Dict[str, int]:
|
|
return dict(num_replicas=xm.xrt_world_size(), rank=xm.get_ordinal())
|
|
|
|
@property
|
|
def is_distributed(self) -> bool:
|
|
# HOST_WORLD_SIZE is None outside the xmp.spawn process
|
|
return os.getenv(xenv.HOST_WORLD_SIZE, None) and self.world_size != 1
|
|
|
|
def process_dataloader(self, dataloader: DataLoader) -> MpDeviceLoader:
|
|
TPUSpawnStrategy._validate_dataloader(dataloader)
|
|
dataloader = MpDeviceLoader(dataloader, self.root_device)
|
|
# Mimic interface to torch.utils.data.DataLoader
|
|
dataloader.dataset = dataloader._loader.dataset
|
|
return dataloader
|
|
|
|
def configure_ddp(self) -> None:
|
|
pass
|
|
|
|
def init_dist_connection(self, global_rank: int, world_size: int) -> None:
|
|
pass
|
|
|
|
def set_world_ranks(self, process_idx: int = 0) -> None:
|
|
pass
|
|
|
|
def model_to_device(self) -> None:
|
|
self.model = self.wrapped_model.to(self.root_device)
|
|
|
|
def barrier(self, name: Optional[str] = None) -> None:
|
|
if self.is_distributed:
|
|
rendezvous(name)
|
|
|
|
def _collect_rank_zero_results(self, trainer: "pl.Trainer", results: Any) -> Optional["_SpawnOutput"]:
|
|
rank_zero_debug("Finalizing the TPU spawn environment.")
|
|
checkpoint_callback = trainer.checkpoint_callback
|
|
best_model_path = checkpoint_callback.best_model_path if checkpoint_callback else None
|
|
|
|
# requires to compute the state_dict on all processes in case Metrics are present
|
|
state_dict = self.lightning_module.state_dict()
|
|
|
|
# save the last weights
|
|
weights_path = None
|
|
if trainer.state.fn == TrainerFn.FITTING:
|
|
weights_path = os.path.join(trainer.default_root_dir, ".temp.ckpt")
|
|
self.checkpoint_io.save_checkpoint(state_dict, weights_path)
|
|
|
|
# We use `local_rank` here as separate filesystems are used for each VM for TPU Pod Training
|
|
if self.local_rank != 0:
|
|
return
|
|
|
|
# adds the `callback_metrics` to the queue
|
|
extra = _FakeQueue()
|
|
if is_overridden("add_to_queue", self.lightning_module):
|
|
# TODO: Remove the if in v1.7
|
|
self.lightning_module.add_to_queue(extra)
|
|
self.add_to_queue(trainer, extra)
|
|
|
|
return _SpawnOutput(best_model_path, weights_path, trainer.state, results, extra)
|
|
|
|
def broadcast(self, obj: object, src: int = 0) -> object:
|
|
if not self.is_distributed:
|
|
return obj
|
|
buffer = io.BytesIO()
|
|
torch.save(obj, buffer)
|
|
data = bytearray(buffer.getbuffer())
|
|
data_tensor = torch.tensor(data, device=self.root_device, dtype=torch.float)
|
|
data = xm.all_gather(data_tensor)
|
|
buffer = io.BytesIO(data.cpu().byte().numpy())
|
|
obj = torch.load(buffer)
|
|
return obj
|
|
|
|
def reduce_boolean_decision(self, decision: bool) -> bool:
|
|
decision = torch.tensor(int(decision), device=self.lightning_module.device)
|
|
decision = self.reduce(decision, reduce_op="sum")
|
|
decision = bool(decision == self.world_size)
|
|
return decision
|
|
|
|
def reduce(self, output, group: Optional[Any] = None, reduce_op: Optional[Union[ReduceOp, str]] = None):
|
|
if not isinstance(output, torch.Tensor):
|
|
output = torch.tensor(output, device=self.lightning_module.device)
|
|
|
|
_invalid_reduce_op = isinstance(reduce_op, ReduceOp) and reduce_op != ReduceOp.SUM
|
|
_invalid_reduce_op_str = isinstance(reduce_op, str) and reduce_op.lower() not in ("sum", "mean", "avg")
|
|
if _invalid_reduce_op or _invalid_reduce_op_str:
|
|
raise MisconfigurationException(
|
|
"Currently, TPUSpawn Strategy only support `sum`, `mean`, `avg` reduce operation."
|
|
)
|
|
|
|
output = xm.mesh_reduce("reduce", output, sum)
|
|
|
|
if isinstance(reduce_op, str) and reduce_op.lower() in ("avg", "mean"):
|
|
output = output / self.world_size
|
|
|
|
return output
|
|
|
|
def get_mp_spawn_kwargs(self, trainer: Optional["pl.Trainer"] = None) -> Dict[str, Any]:
|
|
return {
|
|
"nprocs": len(self.parallel_devices),
|
|
"start_method": self.start_method,
|
|
}
|
|
|
|
def spawn(self, function: Callable, *args: Any, **kwargs: Any) -> Optional[Union[Any, "_SpawnOutput"]]:
|
|
context = mp.get_context(self.start_method or "fork")
|
|
return_queue = context.SimpleQueue()
|
|
xmp.spawn(self._wrapped_function, args=(function, args, kwargs, return_queue), **self.get_mp_spawn_kwargs())
|
|
return return_queue.get()
|
|
|
|
def _wrapped_function(
|
|
self, process_idx: int, function: Callable, args: Any, kwargs: Any, return_queue: SimpleQueue
|
|
) -> None:
|
|
self._worker_setup(process_idx)
|
|
result = function(*args, **kwargs)
|
|
if self.local_rank == 0:
|
|
return_queue.put(move_data_to_device(result, "cpu"))
|
|
|
|
# https://github.com/pytorch/xla/issues/1801#issuecomment-602799542
|
|
self.barrier("end-process")
|
|
|
|
# Ensure that the rank 0 process is the one exiting last
|
|
# https://github.com/pytorch/xla/issues/2190#issuecomment-641665358
|
|
if self.local_rank == 0:
|
|
time.sleep(2)
|
|
|
|
def _worker_setup(self, process_idx: int):
|
|
reset_seed()
|
|
self.tpu_local_core_rank = xm.get_local_ordinal()
|
|
self.tpu_global_core_rank = xm.get_ordinal()
|
|
rank_zero_only.rank = self.global_rank
|
|
|
|
def validation_step(self, *args, **kwargs) -> Optional[STEP_OUTPUT]:
|
|
with self.precision_plugin.val_step_context():
|
|
return self.model(*args, **kwargs)
|
|
|
|
def test_step(self, *args, **kwargs) -> Optional[STEP_OUTPUT]:
|
|
with self.precision_plugin.test_step_context():
|
|
return self.model(*args, **kwargs)
|
|
|
|
def predict_step(self, *args, **kwargs) -> STEP_OUTPUT:
|
|
with self.precision_plugin.predict_step_context():
|
|
return self.model(*args, **kwargs)
|
|
|
|
def training_step_end(self, output: STEP_OUTPUT) -> STEP_OUTPUT:
|
|
self._pod_progress_bar_force_stdout()
|
|
return output
|
|
|
|
def validation_step_end(self, output: STEP_OUTPUT) -> STEP_OUTPUT:
|
|
self._pod_progress_bar_force_stdout()
|
|
return output
|
|
|
|
def test_step_end(self, output: STEP_OUTPUT) -> STEP_OUTPUT:
|
|
self._pod_progress_bar_force_stdout()
|
|
return output
|
|
|
|
def _pod_progress_bar_force_stdout(self) -> None:
|
|
# Why is it required? The way `pytorch_xla.distributed` streams logs
|
|
# from different vms to the main worker doesn't work well with tqdm
|
|
# Ref: https://github.com/pytorch/xla/blob/master/torch_xla/distributed/xla_dist.py#L140
|
|
# The print statement seems to force tqdm to flush stdout.
|
|
if self.tpu_global_core_rank == 0 and int(os.getenv(xenv.TPUVM_MODE, 0)) == 1:
|
|
print()
|
|
|
|
def save_checkpoint(self, checkpoint: Dict[str, Any], filepath: _PATH) -> None:
|
|
"""Save model/training states as a checkpoint file through state-dump and file-write.
|
|
|
|
Args:
|
|
checkpoint: dict containing model and trainer state
|
|
filepath: write-target file's path
|
|
"""
|
|
# `xla_model.save` needs to be called on all ranks. It internally checks if the local rank is 0
|
|
self.checkpoint_io.save_checkpoint(checkpoint, filepath)
|
|
|
|
def remove_checkpoint(self, filepath: _PATH) -> None:
|
|
"""Remove checkpoint filepath from the filesystem.
|
|
|
|
Args:
|
|
filepath: Path to checkpoint
|
|
"""
|
|
if self.local_rank == 0:
|
|
self.checkpoint_io.remove_checkpoint(filepath)
|
|
|
|
def all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None, sync_grads: bool = False) -> torch.Tensor:
|
|
"""
|
|
Function to gather a tensor from several distributed processes
|
|
Args:
|
|
tensor: tensor of shape (batch, ...)
|
|
group: not available with TPUs
|
|
sync_grads: not available with TPUs
|
|
Return:
|
|
A tensor of shape (world_size, batch, ...)
|
|
"""
|
|
if isinstance(tensor, torch.Tensor) and tensor.dim() == 0:
|
|
tensor = tensor.unsqueeze(0)
|
|
return xm.all_gather(tensor)
|
|
|
|
def teardown(self) -> None:
|
|
super().teardown()
|
|
os.environ.pop("PT_XLA_DEBUG", None)
|
|
|
|
@classmethod
|
|
def register_strategies(cls, strategy_registry: Dict) -> None:
|
|
strategy_registry.register(
|
|
"tpu_spawn_debug", cls, description="TPUSpawn Strategy with `debug` as True", debug=True
|
|
)
|