Checkpointing is implemented by rerunning a forward-pass segment for
each checkpointed segment during backward. This can cause persistent
states like the RNG state to be advanced than they would without
checkpointing. By default, checkpointing includes logic to juggle
the RNG state such that checkpointed passes making use of RNG
(through dropout for example) have deterministic output as
compared to non-checkpointed passes. The logic to stash and restore
RNG states can incur a moderate performance hit depending on the runtime
of checkpointed operations. If deterministic output compared to
non-checkpointed passes is not required, supply
checkpoint_sequential to omit stashing and
restoring the RNG state during each checkpoint.
The stashing logic saves and restores the RNG state for the current device
and the device of all cuda Tensor arguments to the
However, the logic has no way to anticipate if the user will move
Tensors to a new device within the
run_fn itself. Therefore, if you move
Tensors to a new device (“new” meaning not belonging to the set of
[current device + devices of Tensor arguments]) within
output compared to non-checkpointed passes is never guaranteed.
checkpoint(function, *args, use_reentrant=True, **kwargs)¶
Checkpoint a model or part of the model
Checkpointing works by trading compute for memory. Rather than storing all intermediate activations of the entire computation graph for computing backward, the checkpointed part does not save intermediate activations, and instead recomputes them in backward pass. It can be applied on any part of a model.
Specifically, in the forward pass,
functionwill run in
torch.no_grad()manner, i.e., not storing the intermediate activations. Instead, the forward pass saves the inputs tuple and the
functionparameter. In the backwards pass, the saved inputs and
functionis retrieved, and the forward pass is computed on
functionagain, now tracking the intermediate activations, and then the gradients are calculated using these activation values.
The output of
functioncan contain non-Tensor values and gradient recording is only performed for the Tensor values. Note that if the output consists of nested structures (ex: custom objects, lists, dicts etc.) consisting of Tensors, these Tensors nested in custom structures will not be considered as part of autograd.
functioninvocation during backward does anything different than the one during forward, e.g., due to some global variable, the checkpointed version won’t be equivalent, and unfortunately it can’t be detected.
use_reentrant=Trueis specified, then if the checkpointed segment contains tensors detached from the computational graph by detach() or torch.no_grad(), the backward pass will raise an error. This is because checkpoint makes all the outputs require gradients which causes issues when a tensor is defined to have no gradient in the model. To circumvent this, detach the tensors outside of the checkpoint function. Note that the checkpointed segment can contain tensors detached from the computational graph if
use_reentrant=Trueis specified, at least one of the inputs needs to have
requires_grad=Trueif grads are needed for model inputs, otherwise the checkpointed part of the model won’t have gradients. At least one of the outputs needs to have
requires_grad=Trueas well. Note that this does not apply if
use_reentrant=Trueis specified, checkpointing currently only supports
torch.autograd.backward()and only if its inputs argument is not passed.
torch.autograd.grad()is not supported. If
use_reentrant=Falseis specified, checkpointing will work with
function – describes what to run in the forward pass of the model or part of the model. It should also know how to handle the inputs passed as the tuple. For example, in LSTM, if user passes
functionshould correctly use the first input as
activationand the second input as
preserve_rng_state (bool, optional, default=True) – Omit stashing and restoring the RNG state during each checkpoint.
use_reentrant (bool, optional, default=True) – Use checkpointing implementation that requires re-entrant autograd. If
checkpointwill use an implementation that does not require re-entrant autograd. This allows
checkpointto support additional functionality, such as working as expected with
torch.autograd.grad. Note that future versions of PyTorch will default to
args – tuple containing inputs to the
Output of running
checkpoint_sequential(functions, segments, input, **kwargs)¶
A helper function for checkpointing sequential models.
Sequential models execute a list of modules/functions in order (sequentially). Therefore, we can divide such a model in various segments and checkpoint each segment. All segments except the last will run in
torch.no_grad()manner, i.e., not storing the intermediate activations. The inputs of each checkpointed segment will be saved for re-running the segment in the backward pass.
checkpoint()on how checkpointing works.
functions – A
torch.nn.Sequentialor the list of modules or functions (comprising the model) to run sequentially.
segments – Number of chunks to create in the model
input – A Tensor that is input to
Output of running
>>> model = nn.Sequential(...) >>> input_var = checkpoint_sequential(model, chunks, input_var)