# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
"""
Train a network across multiple GPUs.
"""
import contextlib
import logging
import sys
import time
from argparse import Namespace
from itertools import chain
from typing import Any, Dict, List
import torch
from fairseq import models, optim, utils
from fairseq.dataclass.configs import FairseqConfig
from fairseq.dataclass.utils import convert_namespace_to_omegaconf
from fairseq.distributed import utils as distributed_utils
from fairseq.file_io import PathManager
from fairseq.logging import meters, metrics
from fairseq.models.ema import build_ema
from fairseq.nan_detector import NanDetector
from fairseq.optim import lr_scheduler
from omegaconf import OmegaConf
from utils import checkpoint_utils
logger = logging.getLogger(__name__)
class Trainer(object):
"""Main class for data parallel training.
This class supports synchronous distributed data parallel training,
where multiple workers each have a full model replica and gradients
are accumulated across workers before each update. We use
:class:`~torch.nn.parallel.DistributedDataParallel` to handle
communication of the gradients across workers.
"""
def __init__(self, cfg: FairseqConfig, task, model, criterion, quantizer=None):
if isinstance(cfg, Namespace):
logger.warning(
"argparse.Namespace configuration is deprecated! Automatically converting to OmegaConf"
)
cfg = convert_namespace_to_omegaconf(cfg)
self.cfg = cfg
self.task = task
# catalog shared parameters
shared_params = _catalog_shared_params(model)
self.tpu = cfg.common.tpu
self.cuda = torch.cuda.is_available() and not cfg.common.cpu and not self.tpu
if self.cuda:
self.device = torch.device("cuda")
elif self.tpu:
self.device = utils.get_tpu_device()
else:
self.device = torch.device("cpu")
if self.is_fsdp:
import fairscale
if self.cfg.common.bf16:
raise ValueError(
"FullyShardedDataParallel is not compatible with --bf16 or "
"--memory-efficient-bf16"
)
if self.cfg.distributed_training.zero_sharding != "none":
raise ValueError(
"FullyShardedDataParallel is not compatible with --zero-sharding "
"option (it's already built in)"
)
if max(self.cfg.optimization.update_freq) > 1 and fairscale.__version__ < "0.4.0":
raise RuntimeError(
"Please update to fairscale 0.4.0 or newer when combining "
"--update-freq with FullyShardedDataParallel"
)
else:
if (
hasattr(self.cfg.distributed_training, "cpu_offload")
and self.cfg.distributed_training.cpu_offload
):
raise ValueError("--cpu-offload requires --ddp-backend=fully_sharded")
# copy model and criterion to current device/dtype
self._criterion = criterion
self._model = model
if not self.is_fsdp:
if cfg.common.fp16:
assert not cfg.common.amp, "Cannot use fp16 and AMP together"
self._criterion = self._criterion.half()
self._model = self._model.half()
elif cfg.common.bf16:
self._criterion = self._criterion.to(dtype=torch.bfloat16)
self._model = self._model.to(dtype=torch.bfloat16)
elif cfg.common.amp:
self._amp_retries = 0
if (
not cfg.distributed_training.pipeline_model_parallel
# the DistributedFairseqModel wrapper will handle moving to device,
# so only handle cases which don't use the wrapper
and not self.use_distributed_wrapper
):
self._criterion = self._criterion.to(device=self.device)
self._model = self._model.to(device=self.device)
self.pipeline_model_parallel = cfg.distributed_training.pipeline_model_parallel
self.last_device = None
if self.cuda and self.pipeline_model_parallel:
self.last_device = torch.device(
cfg.distributed_training.pipeline_devices[-1]
)
# check that shared parameters are preserved after device transfer
for shared_param in shared_params:
ref = _get_module_by_path(self._model, shared_param[0])
for path in shared_param[1:]:
logger.info(
"detected shared parameter: {} <- {}".format(shared_param[0], path)
)
_set_module_by_path(self._model, path, ref)
self._dummy_batch = None # indicates we don't have a dummy batch at first
self._lr_scheduler = None
self._num_updates = 0
self._num_xla_compiles = 0 # for TPUs
self._optim_history = None
self._optimizer = None
self._warn_once = set()
self._wrapped_criterion = None
self._wrapped_model = None
self._ema = None
# TODO(myleott): support tpu
if self.cuda and self.data_parallel_world_size > 1:
self._grad_norm_buf = torch.cuda.DoubleTensor(self.data_parallel_world_size)
else:
self._grad_norm_buf = None
self.quantizer = quantizer
if self.quantizer is not None:
self.quantizer.set_trainer(self)
# get detailed cuda environment
if self.cuda:
self.cuda_env = utils.CudaEnvironment()
if self.data_parallel_world_size > 1:
self.cuda_env_arr = distributed_utils.all_gather_list(
self.cuda_env, group=distributed_utils.get_global_group()
)
else:
self.cuda_env_arr = [self.cuda_env]
if self.data_parallel_rank == 0:
utils.CudaEnvironment.pretty_print_cuda_env_list(self.cuda_env_arr)
else:
self.cuda_env = None
self.cuda_env_arr = None
metrics.log_start_time("wall", priority=790, round=0)
self._start_time = time.time()
self._previous_training_time = 0
self._cumulative_training_time = None
def reinitialize(self):
"""Reinitialize the Trainer, typically after model params change."""
self._lr_scheduler = None
self._optimizer = None
self._wrapped_criterion = None
self._wrapped_model = None
@property
def data_parallel_world_size(self):
if self.cfg.distributed_training.distributed_world_size == 1:
return 1
return distributed_utils.get_data_parallel_world_size()
@property
def data_parallel_process_group(self):
return distributed_utils.get_data_parallel_group()
@property
def data_parallel_rank(self):
if self.cfg.distributed_training.distributed_world_size == 1:
return 0
return distributed_utils.get_data_parallel_rank()
@property
def is_data_parallel_master(self):
# NOTE: this returns true for all model parallel replicas with data
# parallel rank 0
return self.data_parallel_rank == 0
@property
def use_distributed_wrapper(self) -> bool:
return (
self.data_parallel_world_size > 1 and not self.cfg.optimization.use_bmuf
) or (
self.is_fsdp and self.cfg.distributed_training.cpu_offload
)
@property
def should_save_checkpoint_on_current_rank(self) -> bool:
"""Indicates whether to save checkpoints on the current DDP rank."""
if (
self.is_fsdp and self.cfg.distributed_training.use_sharded_state
) or getattr(self.cfg.model, "base_layers", 0) > 0:
return True
else:
return self.is_data_parallel_master
@property
def always_call_state_dict_during_save_checkpoint(self) -> bool:
if self.is_fsdp and not self.cfg.distributed_training.use_sharded_state:
# FSDP calls communication collective when consolidating checkpoints
return True
else:
return False
@property
def checkpoint_suffix(self) -> str:
"""Suffix to add to the checkpoint file name."""
if self.is_fsdp and self.cfg.distributed_training.use_sharded_state:
return self.cfg.checkpoint.checkpoint_suffix + "-shard{0}".format(
self.data_parallel_rank
)
else:
return self.cfg.checkpoint.checkpoint_suffix or ""
@property
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
@property
def model(self):
if self._wrapped_model is None:
if self.use_distributed_wrapper:
self._wrapped_model = models.DistributedFairseqModel(
self.cfg.distributed_training,
self._model,
process_group=self.data_parallel_process_group,
device=self.device,
)
else:
self._wrapped_model = self._model
return self._wrapped_model
@property
def ema(self):
if self._ema is None:
self._build_ema()
return self._ema
def _build_ema(self):
if self.cfg.ema.store_ema:
self._ema = build_ema(self._model, self.cfg.ema, self.device)
logger.info(
"Exponential Moving Average Shadow Model is initialized."
)
@property
def optimizer(self):
if self._optimizer is None:
self._build_optimizer()
return self._optimizer
@property
def lr_scheduler(self):
if self._lr_scheduler is None:
self._build_optimizer() # this will initialize self._lr_scheduler
return self._lr_scheduler
def _build_optimizer(self):
params = list(
filter(
lambda p: p.requires_grad,
chain(self.model.parameters(), self.criterion.parameters()),
)
)
if self.is_fsdp and self.cfg.common.fp16:
# FullyShardedDataParallel always uses MemoryEfficientFP16 wrapper,
# mostly for the grad scaling. But if we don't have the
# --memory-efficient-fp16 flag set, then we're effectively doing
# regular --fp16 and can allow the use of optimizers that would
# otherwise be unsupported by MemoryEfficientFP16Optimizer.
allow_unsupported = not self.cfg.common.memory_efficient_fp16
self._optimizer = optim.MemoryEfficientFP16Optimizer.build_optimizer(
self.cfg, params, allow_unsupported=allow_unsupported
)
elif self.cfg.common.fp16 or self.cfg.common.bf16 or self.cfg.common.amp:
if self.cuda and torch.cuda.get_device_capability(0)[0] < 7:
logger.info(
"NOTE: your device does NOT support faster training with --fp16 or --amp, "
"please switch to FP32 which is likely to be faster"
)
if (
self.cfg.common.memory_efficient_fp16
or self.cfg.common.memory_efficient_bf16
):
self._optimizer = optim.MemoryEfficientFP16Optimizer.build_optimizer(
self.cfg, params
)
elif self.cfg.common.amp:
self._optimizer = optim.AMPOptimizer.build_optimizer(self.cfg, params)
else:
self._optimizer = optim.FP16Optimizer.build_optimizer(self.cfg, params)
else:
if self.cuda and torch.cuda.get_device_capability(0)[0] >= 7:
logger.info("NOTE: your device may support faster training with --fp16 or --amp")
self._optimizer = optim.build_optimizer(self.cfg.optimizer, params)
if self.is_fsdp:
assert (
not self.cfg.optimization.use_bmuf
), "--ddp-backend=fully_sharded is not compatible with BMUF"
assert self._optimizer.supports_flat_params, (
"--ddp-backend=fully_sharded is only compatible with pointwise "
"optimizers (e.g., Adam, AdamW, Adadelta, Adamax, SGD, etc.). "
"However, the sharding will result in slightly different results when "
"using non-pointwise optimizers (e.g., Adagrad, Adafactor, LAMB)"
)
if self.cfg.optimization.use_bmuf:
self._optimizer = optim.FairseqBMUF(
self.cfg.bmuf,
self._optimizer,
)
if self.cfg.distributed_training.zero_sharding == "os":
if (
self.cfg.common.fp16
and not self.cfg.common.memory_efficient_fp16
and not self.cfg.common.memory_efficient_bf16
) and not self.cfg.common.fp16_no_flatten_grads:
raise ValueError(
"ZeRO is incomptabile with fp16 and flattened grads. "
"Please use --fp16-no-flatten-grads"
)
else:
optim.shard_(self._optimizer, self.data_parallel_process_group)
# We should initialize the learning rate scheduler immediately after
# building the optimizer, so that the initial learning rate is set.
self._lr_scheduler = lr_scheduler.build_lr_scheduler(
self.cfg.lr_scheduler,
self.optimizer,
)
self._lr_scheduler.step_update(0)
@property
def is_fsdp(self):
return self.cfg.distributed_training.ddp_backend == "fully_sharded"
def consolidate_optimizer(self):
"""For OSS, we need to consolidate the state dict."""
if self.cfg.checkpoint.no_save_optimizer_state:
return
self._gathered_optim_state = None
if hasattr(self.optimizer.optimizer, "consolidate_state_dict"):
self.optimizer.optimizer.consolidate_state_dict()
elif self.is_fsdp and not self.model.use_sharded_state:
st = self.model.gather_full_optim_state_dict(
self.optimizer
) # only returns on rank 0
self._gathered_optim_state = st
def state_dict(self):
state_dict = {
"args": None, # legacy
"cfg": (
OmegaConf.to_container(self.cfg, resolve=True, enum_to_str=True)
if OmegaConf.is_config(self.cfg)
else self.cfg
),
"model": self.model.state_dict(),
"criterion": (
self.criterion.state_dict()
if utils.has_parameters(self.criterion)
else None
),
"optimizer_history": (self._optim_history or [])
+ [
{
"criterion_name": self.get_criterion().__class__.__name__,
"optimizer_name": self.optimizer.__class__.__name__,
"lr_scheduler_state": self.lr_scheduler.state_dict(),
"num_updates": self.get_num_updates(),
}
],
"task_state": self.task.state_dict() if self.task is not None else {},
"extra_state": {
"metrics": metrics.state_dict(),
"previous_training_time": self.cumulative_training_time(),
},
}
if self.cfg.ema.store_ema:
# Save EMA model state as extra state
state_dict["extra_state"]["ema"] = self.ema.get_model().state_dict()
if self.cfg.ema.ema_fp32:
# Save EMA params in fp32
state_dict["extra_state"]["ema_fp32_params"] = self.ema.fp32_params
if not self.cfg.checkpoint.no_save_optimizer_state:
if self._gathered_optim_state is not None:
state_dict["last_optimizer_state"] = self._gathered_optim_state
self._gathered_optim_state = None
else:
state_dict["last_optimizer_state"] = self.optimizer.state_dict()
if self.is_fsdp:
# save meta data for recombining checkpoint upon loading
state_dict["fsdp_metadata"] = self.model.local_metadata_dict()
return state_dict
def save_checkpoint(self, filename, extra_state):
"""Save all training state in a checkpoint file."""
logger.info(f"Saving checkpoint to {filename}")
# call state_dict on all ranks in case it needs internal communication
state_dict = utils.move_to_cpu(self.state_dict())
state_dict["extra_state"].update(extra_state)
if self.should_save_checkpoint_on_current_rank:
checkpoint_utils.torch_persistent_save(
state_dict,
filename,
async_write=self.cfg.checkpoint.write_checkpoints_asynchronously,
)
logger.info(f"Finished saving checkpoint to {filename}")
def load_checkpoint(
self,
filename,
reset_optimizer=False,
reset_lr_scheduler=False,
optimizer_overrides=None,
reset_meters=False,
):
"""
Load all training state from a checkpoint file.
rank = 0 will load the checkpoint, and then broadcast it to all
other ranks.
"""
extra_state, self._optim_history, last_optim_state = None, [], None
logger.info(f"Preparing to load checkpoint {filename}")
is_distributed = self.data_parallel_world_size > 1
bexists = PathManager.isfile(filename)
if bexists:
load_on_all_ranks = (
self.cfg.checkpoint.load_checkpoint_on_all_dp_ranks
# TPUs don't support broadcast yet, so load checkpoints
# on every worker for now
or self.tpu
# FSDP requires loading checkpoint shards on all ranks
or (self.is_fsdp and self.cfg.distributed_training.use_sharded_state)
or getattr(self.cfg.model, "base_layers", 0) > 0
)
if load_on_all_ranks or self.data_parallel_rank == 0:
state = checkpoint_utils.load_checkpoint_to_cpu(
filename, load_on_all_ranks=load_on_all_ranks
)
last_optim_state = state.get("last_optimizer_state", None)
# If doing zero_sharding, do not broadcast global optimizer
# state. Later we will broadcast sharded states to each rank
# to avoid memory from exploding.
if (
not load_on_all_ranks
and self.cfg.distributed_training.zero_sharding == "os"
and "last_optimizer_state" in state
and is_distributed
):
state["last_optimizer_state"] = "SHARDED"
else:
last_optim_state = None
state = None
if is_distributed and not load_on_all_ranks:
state = distributed_utils.broadcast_object(
state,
src_rank=0,
group=self.data_parallel_process_group,
dist_device=self.device,
)
if self.data_parallel_rank > 0:
last_optim_state = state.get("last_optimizer_state", None)
# load model parameters
try:
if self.cfg.checkpoint.use_ema_weights_to_init_param and "extra_state" in state and "ema" in state["extra_state"]:
logger.info("use_ema_weights_to_init_param = True, will use EMA weights in the ckpt to init the model param...")
ema_state_dict = state["extra_state"]["ema_fp32_params"] if "ema_fp32_params" in state["extra_state"] else state["extra_state"]["ema"]
self.model.load_state_dict(
ema_state_dict, strict=True, model_cfg=self.cfg.model
)
else:
self.model.load_state_dict(
state["model"], strict=True, model_cfg=self.cfg.model
)
# save memory for later steps
if not (self.cfg.ema.store_ema and (self.cfg.checkpoint.use_latest_weights_to_init_ema or not ("extra_state" in state and "ema" in state["extra_state"]))):
del state["model"]
if utils.has_parameters(self.get_criterion()):
self.get_criterion().load_state_dict(
state["criterion"], strict=True
)
del state["criterion"]
except Exception:
raise Exception(
"Cannot load model parameters from checkpoint {}; "
"please ensure that the architectures match.".format(filename)
)
extra_state = state["extra_state"]
self._optim_history = state["optimizer_history"]
if last_optim_state is not None and not reset_optimizer:
# rebuild optimizer after loading model, since params may have changed
self._build_optimizer()
# only reload optimizer and lr_scheduler if they match
last_optim = self._optim_history[-1]
assert (
last_optim["criterion_name"] == self.get_criterion().__class__.__name__
), f"Criterion does not match; please reset the optimizer (--reset-optimizer). {last_optim['criterion_name']} vs {self.get_criterion().__class__.__name__}"
assert (
last_optim["optimizer_name"] == self.optimizer.__class__.__name__
), f"Optimizer does not match; please reset the optimizer (--reset-optimizer). {last_optim['optimizer_name']} vs {self.optimizer.__class__.__name__}"
if not reset_lr_scheduler:
self.lr_scheduler.load_state_dict(last_optim["lr_scheduler_state"])
if self.is_fsdp and not self.model.use_sharded_state:
# if use_sharded_state, the last_optim_state is already sharded, skip this
last_optim_state = self.model.get_shard_from_optim_state_dict(
last_optim_state
)
elif not load_on_all_ranks and is_distributed:
last_optim_state = self.optimizer.broadcast_global_state_dict(
last_optim_state
)
self.optimizer.load_state_dict(last_optim_state, optimizer_overrides)
self.set_num_updates(last_optim["num_updates"])
if extra_state is not None:
itr_state = extra_state["train_iterator"]
epoch = itr_state["epoch"]
if "previous_training_time" in extra_state:
self._previous_training_time = extra_state["previous_training_time"]
self._start_time = time.time()
self.lr_step(epoch)
if (
itr_state.get("version", 1) >= 2
and itr_state["iterations_in_epoch"] == 0
):
# reset meters at start of epoch
reset_meters = True
if "metrics" in extra_state and not reset_meters:
metrics.load_state_dict(extra_state["metrics"])
# reset TimeMeters, since their start times don't make sense anymore
for meter in metrics.get_meters("default"):
if isinstance(meter, meters.TimeMeter):
meter.reset()
if self.cfg.ema.store_ema:
if self.cfg.checkpoint.use_latest_weights_to_init_ema or "ema" not in extra_state:
if "ema" not in extra_state:
logger.warn(
"EMA not found in checkpoint. But store_ema is True. "
"EMA is re-initialized from checkpoint."
)
elif self.cfg.checkpoint.use_latest_weights_to_init_ema:
logger.info(
"use_latest_weights_to_init_ema = True. EMA is re-initialized from checkpoint."
)
self.ema.restore(state["model"], build_fp32_params=self.cfg.ema.ema_fp32)
del state["model"]
else:
logger.info(
"Loading EMA from checkpoint"
)
self.ema.restore(extra_state["ema"], build_fp32_params=False)
if self.cfg.ema.ema_fp32:
if "ema_fp32_params" in extra_state:
logger.info(
"Loading EMA fp32 params from checkpoint"
)
self.ema.build_fp32_params(extra_state["ema_fp32_params"])
else:
logger.info(
"Building EMA fp32 params from EMA model in checkpoint"
)
self.ema.build_fp32_params()
logger.info(
"Loaded checkpoint {} (epoch {} @ {} updates)".format(
filename, epoch, self.get_num_updates()
)
)
else:
logger.info("No existing checkpoint found {}".format(filename))
return extra_state
def get_train_iterator(
self,
epoch,
combine=True,
load_dataset=True,
data_selector=None,
shard_batch_itr=True,
disable_iterator_cache=False,
):
"""Return an EpochBatchIterator over the training set for a given epoch."""
if load_dataset:
logger.info("loading train data for epoch {}".format(epoch))
self.task.load_dataset(
self.cfg.dataset.train_subset,
epoch=epoch,
combine=combine,
data_selector=data_selector,
tpu=self.tpu,
)
batch_iterator = self.task.get_batch_iterator(
dataset=self.task.dataset(self.cfg.dataset.train_subset),
max_tokens=self.cfg.dataset.max_tokens,
max_sentences=self.cfg.dataset.batch_size,
max_positions=utils.resolve_max_positions(
self.task.max_positions(),
self.model.max_positions(),
self.cfg.dataset.max_tokens,
),
ignore_invalid_inputs=True,
required_batch_size_multiple=self.cfg.dataset.required_batch_size_multiple,
seed=self.cfg.common.seed,
num_shards=self.data_parallel_world_size if shard_batch_itr else 1,
shard_id=self.data_parallel_rank if shard_batch_itr else 0,
num_workers=self.cfg.dataset.num_workers,
epoch=epoch,
data_buffer_size=self.cfg.dataset.data_buffer_size,
disable_iterator_cache=disable_iterator_cache,
)
self.reset_dummy_batch(batch_iterator.first_batch)
batch_iterator.dataset.dataset._seek()
return batch_iterator
def get_valid_iterator(
self,
subset,
disable_iterator_cache=False,
):
"""Return an EpochBatchIterator over given validation subset for a given epoch."""
self.task.dataset(subset).dataset._seek()
batch_iterator = self.task.get_batch_iterator(
dataset=self.task.dataset(subset),
max_tokens=self.cfg.dataset.max_tokens_valid,
max_sentences=self.cfg.dataset.batch_size_valid,
max_positions=utils.resolve_max_positions(
self.task.max_positions(),
self.model.max_positions(),
),
ignore_invalid_inputs=self.cfg.dataset.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=self.cfg.dataset.required_batch_size_multiple,
seed=self.cfg.common.seed,
num_shards=self.data_parallel_world_size,
shard_id=self.data_parallel_rank,
num_workers=self.cfg.dataset.num_workers,
# always pass a fixed "epoch" to keep validation data consistent
# across training epochs
epoch=1,
data_buffer_size=self.cfg.dataset.data_buffer_size,
disable_iterator_cache=disable_iterator_cache,
)
self.reset_dummy_batch(batch_iterator.first_batch)
batch_iterator.dataset.dataset._seek()
return batch_iterator
def begin_epoch(self, epoch):
"""Called at the beginning of each epoch."""
logger.info("begin training epoch {}".format(epoch))
self.lr_step_begin_epoch(epoch)
if self.quantizer is not None:
self.quantizer.begin_epoch(epoch)
# task specific setup per epoch
self.task.begin_epoch(epoch, self.get_model())
if self.tpu:
import torch_xla.core.xla_model as xm
xm.rendezvous("begin_epoch") # wait for all workers
xm.mark_step()
def begin_valid_epoch(self, epoch):
"""Called at the beginning of each validation epoch."""
# task specific setup per validation epoch
self.task.begin_valid_epoch(epoch, self.get_model())
def reset_dummy_batch(self, batch):
self._dummy_batch = batch
@metrics.aggregate("train")
def train_step(self, samples, raise_oom=False):
"""Do forward, backward and parameter update."""
self._set_seed()
self.model.train()
self.criterion.train()
self.zero_grad()
metrics.log_start_time("train_wall", priority=800, round=0)
# If EMA is enabled through store_ema=True
# and task.uses_ema is True, pass the EMA model as a keyword
# argument to the task.
extra_kwargs = {}
if self.cfg.ema.store_ema and getattr(self.task, "uses_ema", False):
extra_kwargs["ema_model"] = self.ema.get_model()
# forward and backward pass
logging_outputs, sample_size, ooms = [], 0, 0
for i, sample in enumerate(samples): # delayed update loop
sample, is_dummy_batch = self._prepare_sample(sample)
def maybe_no_sync():
"""
Whenever *samples* contains more than one mini-batch, we
want to accumulate gradients locally and only call
all-reduce in the last backwards pass.
"""
if (
self.data_parallel_world_size > 1
and hasattr(self.model, "no_sync")
and i < len(samples) - 1
# The no_sync context manager results in increased memory
# usage with FSDP, since full-size gradients will be
# accumulated on each GPU. It's typically a better tradeoff
# to do the extra communication with FSDP.
and not self.is_fsdp
):
return self.model.no_sync()
else:
return contextlib.ExitStack() # dummy contextmanager
try:
with maybe_no_sync():
# forward and backward
loss, sample_size_i, logging_output = self.task.train_step(
sample=sample,
model=self.model,
criterion=self.criterion,
optimizer=self.optimizer,
update_num=self.get_num_updates(),
ignore_grad=is_dummy_batch,
**extra_kwargs,
)
del loss
logging_outputs.append(logging_output)
sample_size += sample_size_i
# emptying the CUDA cache after the first step can
# reduce the chance of OOM
if self.cuda and self.get_num_updates() == 0:
torch.cuda.empty_cache()
except RuntimeError as e:
if "out of memory" in str(e):
self._log_oom(e)
if raise_oom:
raise e
logger.warning(
"attempting to recover from OOM in forward/backward pass"
)
ooms += 1
self.zero_grad()
if self.cuda:
torch.cuda.empty_cache()
if self.cfg.distributed_training.distributed_world_size == 1:
return None
else:
raise e
if self.tpu and i < len(samples) - 1:
# tpu-comment: every XLA operation before marking step is
# appended to the IR graph, and processing too many batches
# before marking step can lead to OOM errors.
# To handle gradient accumulation use case, we explicitly
# mark step here for every forward pass without a backward pass
self._xla_markstep_and_send_to_cpu()
if is_dummy_batch:
if torch.is_tensor(sample_size):
sample_size.zero_()
else:
sample_size *= 0.0
if torch.is_tensor(sample_size):
sample_size = sample_size.float()
else:
sample_size = float(sample_size)
# gather logging outputs from all replicas
if self._sync_stats():
train_time = self._local_cumulative_training_time()
logging_outputs, (
sample_size,
ooms,
total_train_time,
) = self._aggregate_logging_outputs(
logging_outputs, sample_size, ooms, train_time, ignore=is_dummy_batch
)
self._cumulative_training_time = (
total_train_time / self.data_parallel_world_size
)
overflow = False
try:
with torch.autograd.profiler.record_function("reduce-grads"):
# reduce gradients across workers
self.optimizer.all_reduce_grads(self.model)
if utils.has_parameters(self.criterion):
self.optimizer.all_reduce_grads(self.criterion)
with torch.autograd.profiler.record_function("multiply-grads"):
# multiply gradients by (data_parallel_size / sample_size) since
# DDP normalizes by the number of data parallel workers for
# improved fp16 precision.
# Thus we get (sum_of_gradients / sample_size) at the end.
# In case of fp16, this step also undoes loss scaling.
# (Debugging note: Some optimizers perform this scaling on the
# fly, so inspecting model.parameters() or optimizer.params may
# still show the original, unscaled gradients.)
numer = (
self.data_parallel_world_size
if not self.cfg.optimization.use_bmuf or self._sync_stats()
else 1
)
self.optimizer.multiply_grads(numer / (sample_size or 1.0))
# Note: (sample_size or 1.0) handles the case of a zero gradient, in a
# way that avoids CPU/device transfers in case sample_size is a GPU or
# TPU object. The assumption is that the gradient itself is also 0.
with torch.autograd.profiler.record_function("clip-grads"):
# clip grads
grad_norm = self.clip_grad_norm(self.cfg.optimization.clip_norm)
# check that grad norms are consistent across workers
# on tpu check tensor is slow
if not self.tpu:
if (
not self.cfg.optimization.use_bmuf
and self.cfg.distributed_training.ddp_backend != "slow_mo"
):
self._check_grad_norms(grad_norm)
if not torch.isfinite(grad_norm).all():
# in case of AMP, if gradients are Nan/Inf then
# optimizer step is still required
if self.cfg.common.amp:
overflow = True
else:
# check local gradnorm single GPU case, trigger NanDetector
raise FloatingPointError("gradients are Nan/Inf")
with torch.autograd.profiler.record_function("optimizer"):
# take an optimization step
self.task.optimizer_step(
self.optimizer, model=self.model, update_num=self.get_num_updates()
)
if self.cfg.common.amp and overflow:
if self._amp_retries == self.cfg.common.amp_batch_retries:
logger.info("AMP: skipping this batch.")
self._amp_retries = 0
else:
self._amp_retries += 1
return self.train_step(samples, raise_oom) # recursion to feed in same batch
except FloatingPointError:
# re-run the forward and backward pass with hooks attached to print
# out where it fails
self.zero_grad()
with NanDetector(self.get_model()):
for _, sample in enumerate(samples):
sample, _ = self._prepare_sample(sample)
self.task.train_step(
sample,
self.model,
self.criterion,
self.optimizer,
self.get_num_updates(),
ignore_grad=False,
**extra_kwargs,
)
raise
except OverflowError as e:
overflow = True
logger.info(
f"NOTE: gradient overflow detected, ignoring gradient, {str(e)}"
)
grad_norm = torch.tensor(0.0).cuda()
self.zero_grad()
except RuntimeError as e:
if "out of memory" in str(e):
self._log_oom(e)
logger.error("OOM during optimization, irrecoverable")
raise e
# Some distributed wrappers (e.g., SlowMo) need access to the optimizer
# after the step
if hasattr(self.model, "perform_additional_optimizer_actions"):
if hasattr(self.optimizer, "fp32_params"):
self.model.perform_additional_optimizer_actions(
self.optimizer.optimizer, self.optimizer.fp32_params
)
else:
self.model.perform_additional_optimizer_actions(
self.optimizer.optimizer
)
logging_output = None
if not overflow or self.cfg.distributed_training.ddp_backend == "slow_mo":
self.set_num_updates(self.get_num_updates() + 1)
if self.cfg.ema.store_ema:
# Step EMA forward with new model.
self.ema.step(
self.get_model(),
self.get_num_updates(),
)
metrics.log_scalar(
"ema_decay",
self.ema.get_decay(),
priority=10000,
round=5,
weight=0,
)
if self.tpu:
import torch_xla.core.xla_model as xm
# mark step on TPUs
self._xla_markstep_and_send_to_cpu()
# only log stats every log_interval steps
# this causes wps to be misreported when log_interval > 1
logging_output = {}
if self.get_num_updates() % self.cfg.common.log_interval == 0:
# log memory usage
mem_info = xm.get_memory_info(self.device)
gb_free = mem_info["kb_free"] / 1024 / 1024
gb_total = mem_info["kb_total"] / 1024 / 1024
metrics.log_scalar(
"gb_free", gb_free, priority=1500, round=1, weight=0
)
metrics.log_scalar(
"gb_total", gb_total, priority=1600, round=1, weight=0
)
logging_outputs = self._xla_markstep_and_send_to_cpu(
logging_outputs
)
logging_output = self._reduce_and_log_stats(
logging_outputs, sample_size, grad_norm
)
# log whenever there's an XLA compilation, since these
# slow down training and may indicate opportunities for
# optimization
self._check_xla_compilation()
else:
if self.cuda and self.cuda_env is not None:
# log minimum free memory over the iteration
gb_used = torch.cuda.max_memory_allocated() / 1024 / 1024 / 1024
torch.cuda.reset_peak_memory_stats()
gb_free = self.cuda_env.total_memory_in_GB - gb_used
metrics.log_scalar(
"gb_free", gb_free, priority=1500, round=1, weight=0
)
# log stats
logging_output = self._reduce_and_log_stats(
logging_outputs, sample_size, grad_norm
)
# clear CUDA cache to reduce memory fragmentation
if (
self.cuda
and self.cfg.common.empty_cache_freq > 0
and (
(self.get_num_updates() + self.cfg.common.empty_cache_freq - 1)
% self.cfg.common.empty_cache_freq
)
== 0
):
torch.cuda.empty_cache()
if self.cfg.common.fp16 or self.cfg.common.amp:
metrics.log_scalar(
"loss_scale",
(
self.optimizer.scaler.loss_scale
if self.cfg.common.fp16
else self.optimizer.scaler.get_scale()
),
priority=700,
round=4,
weight=0,
)
metrics.log_stop_time("train_wall")
return logging_output
@metrics.aggregate("valid")
def valid_step(self, sample, raise_oom=False):
"""Do forward pass in evaluation mode."""
if self.tpu:
import torch_xla.core.xla_model as xm
xm.rendezvous("valid_step") # wait for all workers
# If EMA is enabled through store_ema=True
# and task.uses_ema is True, pass the EMA model as a keyword
# argument to the task.
extra_kwargs = {}
if self.cfg.ema.store_ema and getattr(self.task, "uses_ema", False):
extra_kwargs["ema_model"] = self.ema.get_model()
with torch.no_grad():
self.model.eval()
self.criterion.eval()
sample, is_dummy_batch = self._prepare_sample(sample)
try:
_loss, sample_size, logging_output = self.task.valid_step(
sample, self.model, self.criterion, **extra_kwargs
)
except RuntimeError as e:
if "out of memory" in str(e):
self._log_oom(e)
if not raise_oom:
logger.warning(
"ran out of memory in validation step, retrying batch"
)
for p in self.model.parameters():
if p.grad is not None:
p.grad = None # free some memory
if self.cuda:
torch.cuda.empty_cache()
return self.valid_step(sample, raise_oom=True)
raise e
logging_outputs = [logging_output]
if is_dummy_batch:
if torch.is_tensor(sample_size):
sample_size.zero_()
else:
sample_size *= 0.0
# gather logging outputs from all replicas
if self.data_parallel_world_size > 1:
logging_outputs, (sample_size,) = self._aggregate_logging_outputs(
logging_outputs,
sample_size,
ignore=is_dummy_batch,
)
# log validation stats
if self.tpu:
logging_outputs = self._xla_markstep_and_send_to_cpu(logging_outputs)
logging_output = self._reduce_and_log_stats(logging_outputs, sample_size)
return logging_output
def zero_grad(self):
self.optimizer.zero_grad()
def lr_step_begin_epoch(self, epoch):
"""Adjust the learning rate at the beginning of the epoch."""
self.lr_scheduler.step_begin_epoch(epoch)
# prefer updating the LR based on the number of steps
return self.lr_step_update()
def lr_reinit(self, total_updates, num_updates):
self.lr_scheduler.reinit(total_updates, num_updates)
def lr_step(self, epoch, val_loss=None):
"""Adjust the learning rate at the end of the epoch."""
self.lr_scheduler.step(epoch, val_loss)
# prefer updating the LR based on the number of steps
return self.lr_step_update()
def lr_step_update(self):
"""Update the learning rate after each update."""
new_lr = self.lr_scheduler.step_update(self.get_num_updates())
if isinstance(new_lr, dict):
for k, v in new_lr.items():
metrics.log_scalar(f"lr_{k}", v, weight=0, priority=300)
new_lr = new_lr.get("default", next(iter(new_lr.values())))
else:
metrics.log_scalar("lr", new_lr, weight=0, priority=300)
return new_lr
def get_lr(self):
"""Get the current learning rate."""
return self.optimizer.get_lr()
def get_model(self):
"""Get the (non-wrapped) model instance."""
return self._model
def get_criterion(self):
"""Get the (non-wrapped) criterion instance."""
return self._criterion
def get_meter(self, name):
"""[deprecated] Get a specific meter by name."""
from fairseq import meters
if "get_meter" not in self._warn_once:
self._warn_once.add("get_meter")
utils.deprecation_warning(
"Trainer.get_meter is deprecated. Please use fairseq.metrics instead."
)
train_meters = metrics.get_meters("train")
if train_meters is None:
train_meters = {}
if name == "train_loss" and "loss" in train_meters:
return train_meters["loss"]
elif name == "train_nll_loss":
# support for legacy train.py, which assumed this meter is
# always initialized
m = train_meters.get("nll_loss", None)
return m or meters.AverageMeter()
elif name == "wall":
# support for legacy train.py, which assumed this meter is
# always initialized
m = metrics.get_meter("default", "wall")
return m or meters.TimeMeter()
elif name == "wps":
m = metrics.get_meter("train", "wps")
return m or meters.TimeMeter()
elif name in {"valid_loss", "valid_nll_loss"}:
# support for legacy train.py, which assumed these meters
# are always initialized
k = name[len("valid_") :]
m = metrics.get_meter("valid", k)
return m or meters.AverageMeter()
elif name == "oom":
return meters.AverageMeter()
elif name in train_meters:
return train_meters[name]
return None
def get_num_updates(self):
"""Get the number of parameters updates."""
return self._num_updates
def set_num_updates(self, num_updates):
"""Set the number of parameters updates."""
self._num_updates = num_updates
self.lr_step_update()
if self.quantizer:
self.quantizer.step_update(self._num_updates)
metrics.log_scalar("num_updates", self._num_updates, weight=0, priority=200)
def clip_grad_norm(self, clip_norm):
def agg_norm_fn(total_norm):
total_norm = total_norm.cuda().float() ** 2
total_norm = distributed_utils.all_reduce(
total_norm, group=self.data_parallel_process_group
)
return total_norm ** 0.5
should_agg_norm = (
self.is_fsdp
and (
self.data_parallel_process_group is not None
or torch.distributed.is_initialized()
)
)
return self.optimizer.clip_grad_norm(
clip_norm, aggregate_norm_fn=agg_norm_fn if should_agg_norm else None
)
def cumulative_training_time(self):
if self._cumulative_training_time is None:
# single GPU
return self._local_cumulative_training_time()
else:
return self._cumulative_training_time
def _local_cumulative_training_time(self):
"""Aggregate training time in seconds."""
return time.time() - self._start_time + self._previous_training_time
def _fp_convert_sample(self, sample):
def apply_half(t):
if t.dtype is torch.float32:
return t.to(dtype=torch.half)
return t
def apply_bfloat16(t):
if t.dtype is torch.float32:
return t.to(dtype=torch.bfloat16)
return t
if self.cfg.common.fp16:
sample = utils.apply_to_sample(apply_half, sample)
if self.cfg.common.bf16:
sample = utils.apply_to_sample(apply_bfloat16, sample)
return sample
def _prepare_sample(self, sample, is_dummy=False):
if sample == "DUMMY":
raise Exception(
"Trying to use an uninitialized 'dummy' batch. This usually indicates "
"that the total number of batches is smaller than the number of "
"participating GPUs. Try reducing the batch size or using fewer GPUs."
)
if sample is None or len(sample) == 0:
assert (
self._dummy_batch is not None and len(self._dummy_batch) > 0
), "Invalid dummy batch: {}".format(self._dummy_batch)
sample, _ = self._prepare_sample(self._dummy_batch, is_dummy=True)
return sample, True
# Given that PCIe/NVLink bandwidth is significantly smaller than DRAM bandwidth
# it makes sense to do the format conversion on the CPU and then transfer
# a smaller buffer to the device. This also saves GPU memory capacity.
if self.cfg.common.on_cpu_convert_precision:
sample = self._fp_convert_sample(sample)
if self.cuda:
if self.pipeline_model_parallel:
if 'target' in sample:
sample['target'] = utils.move_to_cuda(sample['target'], device=self.last_device)
else:
sample = utils.move_to_cuda(sample)
elif self.tpu and is_dummy:
# the dummy batch may not be on the appropriate device
sample = utils.move_to_cuda(sample, device=self.device)
if not self.cfg.common.on_cpu_convert_precision:
sample = self._fp_convert_sample(sample)
if self._dummy_batch == "DUMMY":
self._dummy_batch = sample
return sample, False
def _set_seed(self):
# Set seed based on args.seed and the update number so that we get
# reproducible results when resuming from checkpoints
seed = self.cfg.common.seed + self.get_num_updates()
utils.set_torch_seed(seed)
def _sync_stats(self):
# Return True if it's using multiple GPUs and DDP or multiple GPUs with
# BMUF and it's a bmuf sync with warmup iterations completed before.
if self.data_parallel_world_size == 1:
return False
elif self.cfg.optimization.use_bmuf:
return (
self.get_num_updates() + 1
) % self.cfg.bmuf.global_sync_iter == 0 and (
self.get_num_updates() + 1
) > self.cfg.bmuf.warmup_iterations
else:
return True
def _log_oom(self, exc):
msg = "OOM: Ran out of memory with exception: {}".format(exc)
logger.warning(msg)
if torch.cuda.is_available() and hasattr(torch.cuda, "memory_summary"):
for device_idx in range(torch.cuda.device_count()):
logger.warning(torch.cuda.memory_summary(device=device_idx))
sys.stderr.flush()
def _aggregate_logging_outputs(
self,
logging_outputs: List[Dict[str, Any]],
*extra_stats_to_sum,
ignore=False,
):
if self.task.__class__.logging_outputs_can_be_summed(self.get_criterion()):
return self._fast_stat_sync_sum(
logging_outputs, *extra_stats_to_sum, ignore=ignore
)
else:
return self._all_gather_list_sync(
logging_outputs, *extra_stats_to_sum, ignore=ignore
)
def _all_gather_list_sync(
self,
logging_outputs: List[Dict[str, Any]],
*extra_stats_to_sum,
ignore=False,
):
"""
Sync logging outputs across workers. all_gather_list_sync is
suitable when logging outputs are complex types.
"""
if self.tpu:
raise NotImplementedError
if ignore:
logging_outputs = []
results = list(
zip(
*distributed_utils.all_gather_list(
[logging_outputs] + list(extra_stats_to_sum),
max_size=getattr(self.cfg.common, "all_gather_list_size", 16384),
group=self.data_parallel_process_group,
)
)
)
logging_outputs, extra_stats_to_sum = results[0], results[1:]
logging_outputs = list(chain.from_iterable(logging_outputs))
extra_stats_to_sum = [sum(s) for s in extra_stats_to_sum]
return logging_outputs, extra_stats_to_sum
def _fast_stat_sync_sum(
self,
logging_outputs: List[Dict[str, Any]],
*extra_stats_to_sum,
ignore=False,
):
"""
Sync logging outputs across workers. fast_stat_sync_sum is
faster than all_gather_list_sync, but is only suitable when
logging outputs are scalars and can be summed. Note that
*logging_outputs* cannot contain any nested dicts/lists.
"""
data = {}
for i, stat in enumerate(extra_stats_to_sum):
data["extra_stats_" + str(i)] = stat
if len(logging_outputs) > 0:
log_keys = list(logging_outputs[0].keys())
for k in log_keys:
if not ignore:
v = sum(log[k] for log in logging_outputs if k in log)
else:
v = logging_outputs[0][k]
v = torch.zeros_like(v) if torch.is_tensor(v) else 0
data["logging_outputs_" + k] = v
else:
log_keys = None
data = distributed_utils.all_reduce_dict(
data, device=self.device, group=self.data_parallel_process_group
)
extra_stats_to_sum = [
data["extra_stats_" + str(i)] for i in range(len(extra_stats_to_sum))
]
if log_keys is not None:
logging_outputs = [{k: data["logging_outputs_" + k] for k in log_keys}]
else:
logging_outputs = []
return logging_outputs, extra_stats_to_sum
def _check_grad_norms(self, grad_norm):
"""Check that grad norms are consistent across workers."""
if self._grad_norm_buf is not None:
self._grad_norm_buf.zero_()
self._grad_norm_buf[self.data_parallel_rank] = grad_norm
distributed_utils.all_reduce(
self._grad_norm_buf, group=self.data_parallel_process_group
)
def is_consistent(tensor):
max_abs_diff = torch.max(torch.abs(tensor - tensor[0]))
return (
(torch.isfinite(tensor).all()
and (max_abs_diff / (tensor[0] + 1e-6) < 1e-6).all())
or
(self.cfg.common.amp and not torch.isfinite(tensor).all())
# in case of amp non-finite grads are fine
)
if not is_consistent(self._grad_norm_buf):
pretty_detail = "\n".join(
"rank {:3d} = {:.8f}".format(r, n)
for r, n in enumerate(self._grad_norm_buf.tolist())
)
error_detail = "grad_norm across the workers:\n{}\n".format(
pretty_detail
)
# use FloatingPointError to trigger NanDetector
raise FloatingPointError(
"Fatal error: gradients are inconsistent between workers. "
"Try --ddp-backend=legacy_ddp. "
"Or are you mixing up different generation of GPUs in training?"
+ "\n"
+ "-" * 80
+ "\n{}\n".format(error_detail)
+ "-" * 80
)
def _reduce_and_log_stats(self, logging_outputs, sample_size, grad_norm=None):
if grad_norm is not None and (
not torch.is_tensor(grad_norm) or torch.isfinite(grad_norm)
):
metrics.log_speed("ups", 1.0, priority=100, round=2)
metrics.log_scalar("gnorm", grad_norm, priority=400, round=3)
if self.cfg.optimization.clip_norm > 0:
metrics.log_scalar(
"clip",
torch.where(
grad_norm > self.cfg.optimization.clip_norm,
grad_norm.new_tensor(100),
grad_norm.new_tensor(0),
),
priority=500,
round=1,
)
with metrics.aggregate() as agg:
if logging_outputs is not None:
self.task.reduce_metrics(logging_outputs, self.get_criterion())
del logging_outputs
# extra warning for criterions that don't properly log a loss value
if "loss" not in agg:
if "loss" not in self._warn_once:
self._warn_once.add("loss")
logger.warning(
"Criterion.reduce_metrics did not log a 'loss' value, "
"which may break some functionality"
)
metrics.log_scalar("loss", -1)
# support legacy interface
if self.tpu:
logging_output = {}
else:
logging_output = agg.get_smoothed_values()
logging_output["sample_size"] = sample_size
for key_to_delete in ["ppl", "wps", "wpb", "bsz"]:
if key_to_delete in logging_output:
del logging_output[key_to_delete]
return logging_output
def _check_xla_compilation(self):
import torch_xla.debug.metrics as met
compile_stats = met.metric_data("CompileTime")
if compile_stats is None:
return
num_xla_compiles = compile_stats[0]
if num_xla_compiles > self._num_xla_compiles:
logger.warning(
"XLA compilation detected on device #{}; too many of these can lead "
"to slow training, but we expect a few in the beginning".format(
self.cfg.distributed_training.distributed_rank
)
)
self._num_xla_compiles = num_xla_compiles
def _xla_markstep_and_send_to_cpu(self, data=None):
import torch_xla.core.xla_model as xm
xm.mark_step()
if data is not None:
from fairseq.utils import xla_device_to_cpu
return xla_device_to_cpu(data)
def _catalog_shared_params(module, memo=None, prefix=""):
if memo is None:
first_call = True
memo = {}
else:
first_call = False
for name, param in module._parameters.items():
param_prefix = prefix + ("." if prefix else "") + name
if param not in memo:
memo[param] = []
memo[param].append(param_prefix)
for name, m in module._modules.items():
if m is None:
continue
submodule_prefix = prefix + ("." if prefix else "") + name
_catalog_shared_params(m, memo, submodule_prefix)
if first_call:
return [x for x in memo.values() if len(x) > 1]
def _get_module_by_path(module, path):
path = path.split(".")
for name in path:
module = getattr(module, name)
return module
def _set_module_by_path(module, path, value):
path = path.split(".")
for name in path[:-1]:
module = getattr(module, name)
setattr(module, path[-1], value)