# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0
"""Core code for sequence length warmup."""
import logging
import textwrap
from typing import Dict, Mapping, Optional
import torch
import torch.utils.data
from composer.core import Algorithm, Batch, Event, State, TimeUnit, get_precision_context
from composer.loggers import Logger
from composer.models import HuggingFaceModel
from composer.utils import dist, ensure_tuple
log = logging.getLogger(__name__)
__all__ = ['SeqLengthWarmup', 'set_batch_sequence_length']
[docs]def set_batch_sequence_length(
batch: Dict[str, torch.Tensor],
curr_seq_len: int,
truncate: bool = True,
preserve_end_of_sequence: bool = False,
) -> Batch:
"""Set the sequence length of a batch.
Changes the sequence length of all tensors in the provided dictionary
to ``curr_seq_len`` by either truncating the tensors (``truncate=True``)
or reshaping the tensors to create new examples from the extra tokens
(``truncate=False``).
.. note::
The schedule for ``curr_seq_len`` over training time should be managed
outside of this function.
.. note::
Variable input lengths can create CUDA OOM errors. To avoid this,
we follow the `PyTorch notes <https://pytorch.org/tutorials/recipes/recipes/tuning_guide.html#pre-allocate-memory-in-case-of-variable-input-length>`_
and pre-allocate the memory with a blank forward and backward pass.
Args:
batch (Dict[str, Tensor]): The input batch to the model, must be a dictionary.
curr_seq_length (int): The desired sequence length to apply.
truncate (bool, optional): Truncate sequences early, or reshape tensors to create
new examples out of the extra tokens. Default: ``True``.
preserve_end_of_sequence (bool, optional): Preserve the end-of-sequence of the batch when
truncating. Useful when input formats include a unique end-of-sequence token.
Ignored if ``truncate=False``. Default: ``False``.
E.g., if ``batch["input_ids"]`` is ``[[10, 11, 12, 13, 14, 15]]``
and ``curr_seq_length=3``, ``"input_ids"`` in the returned batch would be
``[[10, 11, 12]]`` with ``preserve_end_of_sequence=False`` and would be
``[[10, 11, 15]]`` with ``preserve_end_of_sequence=True``. This behavior applies to any
batch tensor with 2 or more dimensions.
Returns:
Dict[str, Tensor]: a Mapping of input tensors to the model,
where all tensors have curr_seq_len in the second dimension.
Example:
.. code-block::
import composer.functional as cf
for epoch in range(num_epochs):
for X, y in train_loader:
X = cf.set_batch_sequence_length(X, sequence_length)
y_hat = model(X)
loss = loss_fn(y_hat, y)
"""
assert isinstance(batch, Mapping)
# This should act like a no-op if curr_seq_len isn't shorter than the batch's sequence length
batch_seq_len = 0
# The batch sequence length is assumed to be the shape[1] dimension of any non-1D batch tensor
for batch_tensor in batch.values():
tensor_shape = batch_tensor.shape
if len(tensor_shape) > 1:
batch_seq_len = tensor_shape[1]
break
if curr_seq_len >= batch_seq_len:
return batch
if truncate:
# Truncate, but preserve end-of-sequence tokens
if preserve_end_of_sequence:
if 'attention_mask' not in batch:
raise ValueError(
'Sequence Length Warmup requires that the batch has "attention_mask" when using ``preserve_end_of_sequence=True``.',
)
r_idx = torch.arange(batch['attention_mask'].shape[0])
# eos_idx should point to the final token index for each batch sample
eos_idx = batch['attention_mask'].sum(dim=1).long() - 1
# eos_idx_truncated is the same thing, after truncation is applied
eos_idx_truncated = eos_idx.clamp(max=curr_seq_len - 1)
for k in batch.keys():
if batch[k].ndim < 2:
raise ValueError(
f'Sequence Length Warmup requires that all tensors are sequence-shaped when ``truncate=True``. '
f'Tensor "{k}" has shape {batch[k].shape}.',
)
eos_value = batch[k][r_idx, eos_idx]
batch[k] = batch[k][:, :curr_seq_len].contiguous()
batch[k][r_idx, eos_idx_truncated] = eos_value
else:
for k in batch.keys():
if batch[k].ndim < 2:
raise ValueError(
f'Sequence Length Warmup requires that all tensors are sequence-shaped when ``truncate=True``. '
f'Tensor "{k}" has shape {batch[k].shape}.',
)
batch[k] = batch[k][:, :curr_seq_len].contiguous()
else:
if 'input_ids' not in batch:
raise ValueError(
'Sequence Length Warmup requires that the batch has "input_ids" when using ``truncate=False``.',
)
input_ids_shape = batch['input_ids'].shape
# ensure new tensor shape is divisible by curr_seq_len
input_ids = batch['input_ids'].view(-1)
tensor_len = (input_ids.shape[0] // curr_seq_len) * curr_seq_len
input_ids = input_ids[:tensor_len]
input_ids = input_ids.view(-1, curr_seq_len)
batch['input_ids'] = input_ids
for k, v in batch.items():
if k == 'input_ids':
continue
if v.shape != input_ids_shape:
raise ValueError(
f'When using ``truncate=False``, Sequence Length Warmup only supports batches where all tensors have the same shape. '
f'Tensor "{k}" has shape {v.shape} but should have shape {input_ids_shape}.',
)
v = v.view(-1)
v = v[:tensor_len]
batch[k] = v.view(-1, curr_seq_len)
return batch
[docs]class SeqLengthWarmup(Algorithm):
"""Progressively increases the sequence length during training.
Changes the sequence length of all tensors in the input batch. The
sequence length increases from ``min_seq_length`` to ``max_seq_length``
in steps of ``step_size`` during the first ``duration`` fraction of
training.
The sequence length is then kept at ``max_seq_length``
for the rest of training.
Tensors are either truncated (``truncate=True``) or reshaped to
create new examples from the extra tokens (``truncate=False``).
This algorithm runs on :attr:`.Event.AFTER_DATALOADER` to modify
the sequence length of a batch of data after the model and data have been moved to
accelerators.
.. note::
``step_size`` should be a `multiple of eight <https://developer.nvidia.com/blog/optimizing-gpu-performance-tensor-cores/>`_ for
optimal throughput on NVIDIA GPUs.
.. note::
Variable input lengths can create CUDA OOM errors. To avoid this,
we follow the `PyTorch notes <https://pytorch.org/tutorials/recipes/recipes/tuning_guide.html#pre-allocate-memory-in-case-of-variable-input-length>`_
and pre-allocate the memory with a blank forward and backward pass.
See the :doc:`Method Card </method_cards/seq_length_warmup>` for more details.
Example:
.. code-block::
from composer.algorithms import SeqLengthWarmup
from composer import Trainer
seq_length_warmup = SeqLengthWarmup(duration=0.5,
min_seq_length=8,
max_seq_length=1024,
step_size=8,
truncate=True,
preserve_end_of_sequence=False)
trainer = Trainer(model=model,
train_dataloader=train_dataloader,
max_duration="1ep",
algorithms=[seq_length_warmup])
Args:
duration (float, optional): Fraction of total training for sequential length
learning. Default = ``0.3``.
min_seq_length (int, optional): Minimum sequence length to start the warmup.
Default = ``8``.
max_seq_length (int, optional): Maximum sequence length to stop the warmup.
Default = ``1024``.
step_size (int, optional): Step size of sequence length. Default = ``8``.
truncate (bool, optional): Truncate sequences early, or reshape tensors to create
new examples out of the extra tokens. Default: ``True``.
preserve_end_of_sequence (bool, optional): Preserve the end-of-sequence of the batch when
truncating. Useful when input formats include a unique end-of-sequence token.
Ignored if ``truncate=False``. Default: ``False``.
E.g., if ``batch["input_ids"]`` is ``[[10, 11, 12, 13, 14, 15]]``
and ``curr_seq_length=3``, ``"input_ids"`` in the returned batch would be
``[[10, 11, 12]]`` with ``preserve_end_of_sequence=False`` and would be
``[[10, 11, 15]]`` with ``preserve_end_of_sequence=True``. This behavior applies to any
batch tensor with 2 or more dimensions.
"""
def __init__(
self,
duration: float = 0.3,
min_seq_length: int = 8,
max_seq_length: int = 1024,
step_size: int = 8,
truncate: bool = True,
preserve_end_of_sequence: bool = False,
):
self.duration = duration
self.min_seq_length = min_seq_length
self.max_seq_length = max_seq_length
self.step_size = step_size
self.truncate = truncate
self.preserve_end_of_sequence = preserve_end_of_sequence
if self.duration < 0 or self.duration > 1:
raise ValueError(f'Duration must be between 0 and 1, got: {self.duration}')
if self.max_seq_length < self.min_seq_length:
raise ValueError(
f'max_seq_length={self.max_seq_length} must be '
f'greater than min_seq_length={self.min_seq_length}',
)
self._activated = False
self._original_model = None
def _activate_model(self, state: State, logger: Logger) -> None:
"""Does a forward and a backward pass on a dummy input.
The purpose of activating the model is to prevent OOMs. This happens two ways.
First, this prevents GPU memory from being reallocated when the sequence
length increases.
Second, it detects if the batch*max_sequence_length size will cause an OOM and
decreases state.device_train_microbatch_size accordingly. This logic mirrors the
``device_train_microbatch_size='auto'`` logic in :class:`.Trainer`.
"""
assert self._original_model is not None, 'original model should be set on Event.INIT'
try:
# Both PyTorch and FFCV dataloaders define a `batch_size` attribute
# This exception would mainly be raised if the user is passing in a custom
# iterable
per_gpu_macrobatch = getattr(state.dataloader, 'batch_size')
except AttributeError as e:
raise AttributeError(
'Sequence Length Warmup requires the `state.dataloader` to have a `batch_size` attribute.',
) from e
if per_gpu_macrobatch is None:
raise RuntimeError('Sequence Length Warmup algorithm requires constant batch size.')
# truncate all sequence-shaped tensors to the max sequence length
batch_clone = {k: torch.clone(v) for k, v in state.batch.items()}
for k, v in batch_clone.items():
if v.ndim < 2:
raise ValueError(
f'Sequence Length Warmup requires that all tensors are sequence-shaped. '
f'Tensor "{k}" has shape {v.shape}.',
)
batch_clone[k] = v[:, :self.max_seq_length].contiguous()
# In-line to avoid circular dependency
from composer.trainer.trainer import _adjust_device_train_microbatch_size, _is_cuda_oom
# This loop tries to do a forward/backward pass using the current microbatch size.
# If it hits an OOM error, it halves `state.device_train_microbatch_size` and tries again
# until it succeeds.
while True:
model_inputs = {k: v[:state.device_train_microbatch_size] for k, v in batch_clone.items()}
found_cuda_oom = 0 # int since bool BOR not supported on all torch.distributed backends
try:
# Start by running a forward and backward pass
# of the maximum sequence length to allocate cache.
with get_precision_context(state.precision, state.precision_config):
outputs = state.model.forward(model_inputs)
loss = self._original_model.loss(outputs, model_inputs)
# Check if other ranks OOMed after forward pass when using auto microbatching. This may
# happen when close to memory limit or with uneven memory usage across ranks
if state.auto_microbatching:
# Check if any other rank hit an OOM
found_cuda_oom_tensor = state.device.tensor_to_device(torch.tensor([0], dtype=torch.uint8))
dist.all_reduce(found_cuda_oom_tensor, reduce_operation='MAX')
found_cuda_oom = found_cuda_oom_tensor.item()
# Signal current rank is still in batch
all_ranks_finished_tensor = state.device.tensor_to_device(torch.tensor([0], dtype=torch.uint8))
dist.all_reduce(all_ranks_finished_tensor, reduce_operation='MIN')
if found_cuda_oom == 1:
raise RuntimeError('CUDA out of memory encountered on a different rank')
# Since use_grad_scaling is in the Trainer, and we
# don't care about the loss values, skip scaling
for loss_item in ensure_tuple(loss):
loss_item.backward()
# Zero any gradients created by the backward pass
for optimizer in state.optimizers:
optimizer.zero_grad()
# This error/state.device_train_microbatch_size handling mimics the logic in trainer._train_batch().
except RuntimeError as e:
if state.auto_microbatching and _is_cuda_oom(e):
log.debug((f"Rank {dist.get_global_rank()} OOM'd."))
found_cuda_oom = 1
elif state.auto_microbatching and ('cuda' in str(e).lower() or 'c10' in str(e).lower()):
raise ValueError(
textwrap.dedent(
'Encountered non-addressable cuda error while using auto microbatching. '
'If this repeatedly occurs, set `device_train_microbatch_size` manually.',
),
) from e
else:
raise
if state.auto_microbatching:
all_ranks_finished = False
while not all_ranks_finished:
# Propagate across all ranks if any rank hit CUDA OOM
found_cuda_oom_tensor = state.device.tensor_to_device(
torch.tensor([found_cuda_oom], dtype=torch.uint8),
)
dist.all_reduce(found_cuda_oom_tensor, reduce_operation='MAX')
found_cuda_oom = found_cuda_oom_tensor.item()
# Check if any rank is still not done with the batch. This may happen if only a
# subset of ranks OOM, leaving some batches still in the forward pass
all_ranks_finished_tensor = state.device.tensor_to_device(torch.tensor([1], dtype=torch.uint8))
dist.all_reduce(all_ranks_finished_tensor, reduce_operation='MIN')
all_ranks_finished = all_ranks_finished_tensor.item() == 1
if found_cuda_oom == 1:
_adjust_device_train_microbatch_size(state)
# Skip return and rerun after handling oom
continue
# Activate and return if we've completed without OOMing.
self._activated = True
return
def match(self, event: Event, state: State) -> bool:
return (event == Event.INIT and self._original_model is None) or event == Event.AFTER_DATALOADER
def apply(self, event: Event, state: State, logger: Logger) -> Optional[int]:
if event == Event.INIT:
if not isinstance(state.model, HuggingFaceModel):
raise RuntimeError(
textwrap.dedent(
f"""\
{type(self).__name__} requires state.model to be of type {HuggingFaceModel.__name__}, not of type {type(state.model)}""",
),
)
self._original_model = state.model
return
assert state.dataloader is not None, 'dataloader should be set on AFTER_DATALOADER'
assert state.max_duration is not None, 'max_duration should be set on AFTER_DATALOADER'
# in order to avoid OOMs, we do a forward and a backward pass on a dummy input.
if not self._activated:
self._activate_model(state, logger)
if state.max_duration.unit == TimeUnit.EPOCH:
if state.dataloader_len is None:
raise RuntimeError('Sequential Length Warmup requires the dataloader to be sized.')
num_optimization_steps = int(state.dataloader_len) * state.max_duration.value
elif state.max_duration.unit == TimeUnit.BATCH:
num_optimization_steps = state.max_duration.value
else:
raise NotImplementedError(
textwrap.dedent(
"""\
To use sequential length warmup, the max_duration must be in epochs or batches.
Specifying the `max_duration` in tokens or samples for use with sequential
length warmup will be supported in a future Composer release. See
https://github.com/mosaicml/composer/issues/226.""",
),
)
num_warmup_steps = int(num_optimization_steps * self.duration) # in batches
# assume the full sequence length is the unaltered sequence length
num_update_steps = (self.max_seq_length - self.min_seq_length) // self.step_size
update_every_n_steps = num_warmup_steps // num_update_steps
curr_seq_len = self.step_size * (int(state.timestamp.batch) // update_every_n_steps) + self.min_seq_length
curr_seq_len = max(curr_seq_len, self.min_seq_length)
curr_seq_len = min(curr_seq_len, self.max_seq_length)
state.batch = set_batch_sequence_length(state.batch, curr_seq_len, self.truncate, self.preserve_end_of_sequence)
batch_size = state.batch['input_ids'].shape[0]
logger.log_metrics({
'seq_length_warmup/curr_seq_len': curr_seq_len,
'seq_length_warmup/curr_bs': batch_size,
})