new_checkpoint.py

import torch
import warnings
import weakref
from weakref import ReferenceType
from typing import Any, Callable, ContextManager, Iterable, List, Tuple, Dict, Optional, DefaultDict
from collections import defaultdict
import uuid
import contextlib

__all__ = [
    "checkpoint", "checkpoint_sequential", "CheckpointFunction",
    "check_backward_validity", "detach_variable", "get_device_states",
    "set_device_states", "noop_context_fn", "set_checkpoint_early_stop",
    "DefaultDeviceType"
]

def detach_variable(inputs: Tuple[Any, ...]) -> Tuple[torch.Tensor, ...]:
    if isinstance(inputs, tuple):
        out = []
        for inp in inputs:
            if not isinstance(inp, torch.Tensor):
                out.append(inp)
                continue


            x = inp.detach()
            x.requires_grad = inp.requires_grad
            out.append(x)
        return tuple(out)
    else:
        raise RuntimeError(
            "Only tuple of tensors is supported. Got Unsupported input type: ", type(inputs).__name__)


def check_backward_validity(inputs: Iterable[Any]) -> None:
    if not any(inp.requires_grad for inp in inputs if isinstance(inp, torch.Tensor)):
        warnings.warn("None of the inputs have requires_grad=True. Gradients will be None")


def _get_device_module(device='cuda'):
    device_module = getattr(torch, device)
    return device_module

class DefaultDeviceType(object):
    r"""
    A class that manages the default device type for checkpointing.
    If no non-CPU tensors are present, the default device type will
    be used. The default value is 'cuda'. The device type is used in
    the checkpointing process when determining which device states
    to save and restore for recomputation.
    """
    _default_device_type = "cuda"

    @staticmethod
    def set_device_type(device: str = "cuda"):
        """
        Set the default device type for checkpointing.

        Args:
            device (str): The device type to be set as default. Default is 'cuda'.
        """
        DefaultDeviceType._default_device_type = device

    @staticmethod
    def get_device_type() -> str:
        """
        Get the current default device type for checkpointing.

        Returns:
            str: The current default device type.
        """
        return DefaultDeviceType._default_device_type

def _infer_device_type(*args):
    device_types = list({arg.device.type for arg in args
                        if isinstance(arg, torch.Tensor) and not arg.device.type == "cpu"})
    if len(device_types) > 1:
        warnings.warn("Tensor arguments, excluding CPU tensors, are detected on at least two types of devices. "
                      "Device state will only be saved for devices of a single device type, and the remaining "
                      "devices will be ignored. Consequently, if any checkpointed functions involve randomness, "
                      "this may result in incorrect gradients. (Note that if CUDA devices are among the devices "
                      "detected, it will be prioritized; otherwise, the first device encountered will be selected.)")
    if len(device_types) == 0:
        return DefaultDeviceType.get_device_type()
    elif "cuda" in device_types:
        return "cuda"
    else:
        return device_types[0]

# We can't know if the run_fn will internally move some args to different devices,
# which would require logic to preserve rng states for those devices as well.
# We could paranoically stash and restore ALL the rng states for all visible devices,
# but that seems very wasteful for most cases.  Compromise:  Stash the RNG state for
# the device of all Tensor args.
#
# To consider:  maybe get_device_states and set_device_states should reside in torch/random.py?
def get_device_states(*args) -> Tuple[List[int], List[torch.Tensor]]:
    # This will not error out if "arg" is a CPU tensor or a non-tensor type because
    # the conditionals short-circuit.
    fwd_device_ids = list({arg.get_device() for arg in args
                          if isinstance(arg, torch.Tensor) and not arg.device.type == "cpu"})

    fwd_device_states = []
    device_module = _get_device_module(_infer_device_type(*args))

    for device_id in fwd_device_ids:
        with device_module.device(device_id):
            fwd_device_states.append(device_module.get_rng_state())

    return fwd_device_ids, fwd_device_states


def set_device_states(devices, states) -> None:
    device_module = _get_device_module(_infer_device_type(*states))
    for device, state in zip(devices, states):
        with device_module.device(device):
            device_module.set_rng_state(state)


def _get_autocast_kwargs(device="cuda"):

    if device == "cuda":
        device_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
                                  "dtype": torch.get_autocast_gpu_dtype(),
                                  "cache_enabled": torch.is_autocast_cache_enabled()}
    else:
        device_module = _get_device_module(device)
        device_autocast_kwargs = {"enabled": device_module.is_autocast_enabled(),
                                  "dtype": device_module.get_autocast_dtype(),
                                  "cache_enabled": torch.is_autocast_cache_enabled()}

    cpu_autocast_kwargs = {"enabled": torch.is_autocast_cpu_enabled(),
                           "dtype": torch.get_autocast_cpu_dtype(),
                           "cache_enabled": torch.is_autocast_cache_enabled()}

    return device_autocast_kwargs, cpu_autocast_kwargs

class CheckpointFunction(torch.autograd.Function):

    @staticmethod
    def forward(ctx, run_function, preserve_rng_state, *args):
        check_backward_validity(args)
        ctx.run_function = run_function
        ctx.preserve_rng_state = preserve_rng_state
        # Accommodates the (remote) possibility that autocast is enabled for cpu AND gpu.
        ctx.device = _infer_device_type(*args)
        ctx.device_autocast_kwargs, ctx.cpu_autocast_kwargs = _get_autocast_kwargs(ctx.device)
        if preserve_rng_state:
            ctx.fwd_cpu_state = torch.get_rng_state()
            # Don't eagerly initialize the cuda context by accident.
            # (If the user intends that the context is initialized later, within their
            # run_function, we SHOULD actually stash the cuda state here.  Unfortunately,
            # we have no way to anticipate this will happen before we run the function.)
            ctx.had_device_in_fwd = False
            device_module = _get_device_module(ctx.device)
            if device_module._initialized:
                ctx.had_device_in_fwd = True
                ctx.fwd_devices, ctx.fwd_device_states = get_device_states(*args)

        # Save non-tensor inputs in ctx, keep a placeholder None for tensors
        # to be filled out during the backward.
        ctx.inputs = []
        ctx.tensor_indices = []
        tensor_inputs = []
        for i, arg in enumerate(args):
            if torch.is_tensor(arg):
                tensor_inputs.append(arg)
                ctx.tensor_indices.append(i)
                ctx.inputs.append(None)
            else:
                ctx.inputs.append(arg)

        ctx.save_for_backward(*tensor_inputs)
        with torch.no_grad():
            outputs = run_function(*args)
        return outputs

    @staticmethod
    def backward(ctx, *args):
        if not torch.autograd._is_checkpoint_valid():
            raise RuntimeError(
                "Checkpointing is not compatible with .grad() or when an `inputs` parameter"
                " is passed to .backward(). Please use .backward() and do not pass its `inputs`"
                " argument.")
        # Copy the list to avoid modifying original list.
        inputs = list(ctx.inputs)
        tensor_indices = ctx.tensor_indices
        tensors = ctx.saved_tensors
        device_module = _get_device_module(ctx.device)

        # Fill in inputs with appropriate saved tensors.
        for i, idx in enumerate(tensor_indices):
            inputs[idx] = tensors[i]

        # Stash the surrounding rng state, and mimic the state that was
        # present at this time during forward.  Restore the surrounding state
        # when we're done.
        rng_devices = []
        if ctx.preserve_rng_state and ctx.had_device_in_fwd:
            rng_devices = ctx.fwd_devices
        # with torch.random.fork_rng(devices=rng_devices, enabled=ctx.preserve_rng_state, device_type=ctx.device):
        with torch.random.fork_rng(devices=rng_devices, enabled=ctx.preserve_rng_state):
            if ctx.preserve_rng_state:
                torch.set_rng_state(ctx.fwd_cpu_state)
                if ctx.had_device_in_fwd:
                    set_device_states(ctx.fwd_devices, ctx.fwd_device_states)
            detached_inputs = detach_variable(tuple(inputs))
            with torch.enable_grad(), \
                 device_module.amp.autocast(**ctx.device_autocast_kwargs), \
                 torch.cpu.amp.autocast(**ctx.cpu_autocast_kwargs):
                outputs = ctx.run_function(*detached_inputs)

        if isinstance(outputs, torch.Tensor):
            outputs = (outputs,)

        # run backward() with only tensor that requires grad
        outputs_with_grad = []
        args_with_grad = []
        for i in range(len(outputs)):
            if torch.is_tensor(outputs[i]) and outputs[i].requires_grad:
                outputs_with_grad.append(outputs[i])
                args_with_grad.append(args[i])
        if len(outputs_with_grad) == 0:
            raise RuntimeError(
                "none of output has requires_grad=True,"
                " this checkpoint() is not necessary")
        torch.autograd.backward(outputs_with_grad, args_with_grad)
        grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else None
                      for inp in detached_inputs)

        return (None, None) + grads


def noop_context_fn():
    return contextlib.nullcontext(), contextlib.nullcontext()


def checkpoint(
    function,
    *args,
    use_reentrant: Optional[bool] = False,
    context_fn: Callable[[], Tuple[ContextManager, ContextManager]] = noop_context_fn,
    swap_list: List = [],
    act_pack: Dict = {},
    **kwargs
):
    r"""Checkpoint a model or part of the model

    Checkpointing is a technique that trades compute for memory. Instead of
    storing all intermediate activations of the entire computation graph for
    the backward pass, the checkpointed part omits saving intermediate
    activations and recomputes them during the backward pass. This can be
    applied to any part of a model.

    There are currently two checkpointing implementations available, determined
    by the :attr:`use_reentrant` parameter. It is recommended that you use
    ``use_reentrant=False``. Please refer the note below for a discussion of
    their differences.

    .. warning::

        If the :attr:`function` invocation during the backward pass differs
        from the forward pass, e.g., due to a global variable, the checkpointed
        checkpointed version may not be equivalent, potentially causing an
        error being raised or leading to silently incorrect gradients.

    .. warning::

        If you are using the ``use_reentrant=True`` variant (this is currently
        the default), please refer to the note below for important
        considerations and potential limitations.

    .. note::

        The reentrant variant of checkpoint (``use_reentrant=True``) and
        the non-reentrant variant of checkpoint (``use_reentrant=False``)
        differ in the following ways:

        * Non-reentrant checkpoint stops recomputation as soon as all needed
          intermediate activations have been recomputed. This feature is enabled
          by default, but can be disabled with :func:`set_checkpoint_early_stop`.
          Reentrant checkpoint always recomputes :attr:`function` in its
          entirety during the backward pass.

       * The reentrant variant does not record the autograd graph during the
          forward pass, as it runs with the forward pass under
          :func:`torch.no_grad`. The non-reentrant version does record the
          autograd graph, allowing one to perform backward on the graph within
          checkpointed regions.

        * The reentrant checkpoint only supports the
          :func:`torch.autograd.backward` API for the backward pass without its
          `inputs` argument, while the non-reentrant version supports all ways
          of performing the backward pass.

        * At least one input and output must have ``requires_grad=True`` for the
          reentrant variant. If this condition is unmet, the checkpointed part
          of the model will not have gradients. The non-reentrant version does
          not have this requirement.

        * The reentrant version does not consider tensors in nested structures
          (e.g., custom objects, lists, dicts, etc) as participating in
          autograd, while the non-reentrant version does.

        * The reentrant checkpoint does not support checkpointed regions with
          detached tensors from the computational graph, whereas the
          non-reentrant version does. For the reentrant variant, if the
          checkpointed segment contains tensors detached using ``detach()`` or
          with :func:`torch.no_grad`, the backward pass will raise an error.
          This is because ``checkpoint`` makes all the outputs require gradients
          and this causes issues when a tensor is defined to have no gradient in
          the model. To avoid this, detach the tensors outside of the
          ``checkpoint`` function.

    Args:
        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
            ``(activation, hidden)``, :attr:`function` should correctly use the
            first input as ``activation`` and the second input as ``hidden``
        preserve_rng_state(bool, optional):  Omit stashing and restoring
            the RNG state during each checkpoint.
            Default: ``True``
        use_reentrant(bool, optional): Use checkpointing
            implementation that requires re-entrant autograd.
            If ``use_reentrant=False`` is specified, ``checkpoint`` will use an
            implementation that does not require re-entrant autograd. This
            allows ``checkpoint`` to support additional functionality, such as
            working as expected with ``torch.autograd.grad`` and support for
            keyword arguments input into the checkpointed function. Note that future
            versions of PyTorch will default to ``use_reentrant=False``.
            Default: ``True``
        context_fn(Callable, optional): A callable returning a tuple of two
            context managers. The function and its recomputation will be run
            under the first and second context managers respectively.
            This argument is only supported if ``use_reentrant=False``.
        args: tuple containing inputs to the :attr:`function`

    Returns:
        Output of running :attr:`function` on :attr:`*args`
    """
    if use_reentrant is None:
        warnings.warn(
            "torch.utils.checkpoint: please pass in use_reentrant=True or "
            "use_reentrant=False explicitly. The default value of use_reentrant "
            "will be updated to be False in the future. To maintain current "
            "behavior, pass use_reentrant=True. It is recommended that you use "
            "use_reentrant=False. Refer to docs for more details on the "
            "differences between the two variants.")
        use_reentrant = True
    # Hack to mix *args with **kwargs in a python 2.7-compliant way
    preserve = kwargs.pop('preserve_rng_state', True)
    if kwargs and use_reentrant:
        raise ValueError("Unexpected keyword arguments: " + ",".join(arg for arg in kwargs))

    if use_reentrant:
        if context_fn is not noop_context_fn:
            raise ValueError("Passing context_fn is only supported when use_reentrant=False.")
        return CheckpointFunction.apply(function, preserve, *args)
    else:
        return _checkpoint_without_reentrant(
            function,
            preserve,
            context_fn,
            swap_list,
            act_pack,
            *args,
            **kwargs,
        )


def checkpoint_sequential(functions, segments, input, use_reentrant=True, **kwargs):
    r"""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 not store
    the intermediate  activations. The inputs of each checkpointed segment will
    be saved for re-running the segment in the backward pass.

    .. warning::
        If you are using the ``use_reentrant=True` variant (this is the
        default), please see :func:`~torch.utils.checkpoint.checkpoint` for
        the important considerations and limitations of this variant. It is
        recommended that you use ``use_reentrant=False``.

    .. warning:
        Since PyTorch 1.4, it allows only one Tensor as the input and
        intermediate outputs, just like :class:`torch.nn.Sequential`.

    Args:
        functions: A :class:`torch.nn.Sequential` or 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 :attr:`functions`
        preserve_rng_state(bool, optional):  Omit stashing and restoring
            the RNG state during each checkpoint.
            Default: ``True``
        use_reentrant(bool, optional): Use checkpointing
            implementation that requires re-entrant autograd.
            If ``use_reentrant=False`` is specified, ``checkpoint`` will use an
            implementation that does not require re-entrant autograd. This
            allows ``checkpoint`` to support additional functionality, such as
            working as expected with ``torch.autograd.grad`` and support for
            keyword arguments input into the checkpointed function.
            Default: ``True``

    Returns:
        Output of running :attr:`functions` sequentially on :attr:`*inputs`

    Example:
        >>> # xdoctest: +SKIP("stub")
        >>> model = nn.Sequential(...)
        >>> input_var = checkpoint_sequential(model, chunks, input_var)
    """
    # Hack for keyword-only parameter in a python 2.7-compliant way
    preserve = kwargs.pop('preserve_rng_state', True)
    if kwargs:
        raise ValueError("Unexpected keyword arguments: " + ",".join(arg for arg in kwargs))

    def run_function(start, end, functions):
        def forward(input):
            for j in range(start, end + 1):
                input = functions[j](input)
            return input
        return forward

    if isinstance(functions, torch.nn.Sequential):
        functions = list(functions.children())

    segment_size = len(functions) // segments
    # the last chunk has to be non-volatile
    end = -1
    for start in range(0, segment_size * (segments - 1), segment_size):
        end = start + segment_size - 1
        input = checkpoint(
            run_function(start, end, functions),
            input,
            use_reentrant=use_reentrant,
            preserve_rng_state=preserve
        )
    return run_function(end + 1, len(functions) - 1, functions)(input)

# NOTE [ Nestable Checkpoint ]
#
# The semantics of nested checkpoint can be defined by two basic rules.
# Following the two rules leads to an important implication that is central
# to motivating the design.
#
# Rule 1. Saved tensors are managed by inner-most checkpoint only and hidden
#         from any outer layers of checkpoint.
#
# Rule 2. The inputs of inner checkpoints are treated as tensors saved to its
#         parent checkpoint.
#
# Implication: To recompute any given saved tensor, we need to recompute all of
#              the checkpoints wrapping it.
#
# Why is this implied? To unpack a saved tensor X during backward we need to
# recompute the inner-most checkpoint (#1), and in order to recompute that
# checkpoint I need to have its inputs, which are managed by that checkpoint's
# parent (#2), which thus also needs to be recomputed first. Continue this line
# of reasoning and we realize that in order to unpack X, all checkpoints that
# were active at the time X was saved need to be recomputed. (unless we have
# already done so in that backward for some other saved tensor).
#
# In practice, we use a noop autograd Function to save inputs as saved tensors.
# During unpack calling ctx.saved_tensor triggers the parent checkpoint to
# recompute.
#
# Rule 3. We should start recomputation as if there are no checkpoints currently
#         active. Checkpoints encountered during recomputation are still
#         respected.
#
# When we start recomputation, we push the saved variable hook meant for
# recomputation on the stack. See examples in Rule 6 for more context.
#
#                                  * * * *
#
# Beyond the basic semantics specific to nested checkpoint, we impose several
# more constraints that may apply to checkpointing in general.
#
# Rule 4. Lifetime of recomputed tensors
#
#         Recomputed tensors are considered specific to particular invocations
#         of backward and are always cleared immediately as they are unpacked
#         Particularly, we require this to happen even if retain_graph=True.
#
# [ Implementation details of Rule 4 ]
#
# If we were okay with recomputed tensors staying alive after backward is run
# with retain_graph=True, we would store recomputed variables as the values of a
# WeakKeyDictionary and pack strong references to the keys, so that as we
# backward, those packed keys would be cleared as long as retain_graph=False.
# Clearing the packed key clears the corresponding entry in the WKD.
#
# If we wish recomputed variables to be immediately cleared as we unpack them in
# the retain_graph=True case, we cannot rely on the packed keys to be cleared by
# backward automatically. Instead of packing the strong reference to the key
# directly, we pack a container object, which we manually clear as we unpack.
#
# An important detail is that if a second backward happens, the second
# recomputation needs to reset the container with a newly created key.
#
# Rule 5. Stop recomputation as soon as we've recomputed the saved tensors we
#         know we need.
#
# [ Implementation details of Rule 5 ]
#
# During recomputation, raise an exception if the number of recomputed tensors
# matches the number of tensors that we expected to recompute. We wrap the
# recomputation call with a try-catch to catch this specific exception. See
# Rule #6 below for some examples.
#
# Rule 6. We support doing backward inside checkpoint context
#
# This section is just a bunch of random examples that we'd like to support,
# and comments on how that forced us to make certain design decisions.
#
# [ Basic case ]
#
# def fn(x):
#   y = x.sin()
#   z = y.cos()
#   gx, = torch.autograd.grad(z, x, retains_grad=True)
#   return gx, z
#
# out = checkpoint(fn)(inp)
# out.backward()
#
# Because z is saved by cos while checkpoint is enabled, it would not be
# actually saved, and so the .grad() call inside must trigger a recomputation.
#
# During recomputation the "inner pack hook" has two responsibilities:
#
# 1) As usual, populating the WeakKeyDictionary storing recomputed tensors
# 2) Pack the actual tensor (detached) so that one may perform backward on the
#    recomputed graph. The tensors saved to this graph will live until the end
#    of recomputation, or die earlier if someone performs backward with
#    retain_graph=False.
#
# More generally performing backward on the recomputed graph occurs in the
# following cases:
# - If backward is performed inside forward,
#   - During the original forward IF early-stop is disabled
#   - During the original backward
# - If there are multiple .grad()/.backward() calls, we would perform backward
#   on the recomputed graph even if early-stop is enabled (see the example below)
#
# [ Multiple backwards ]
#
# The example below shows what happens if during recomputation we find that some
# of the tensors we are trying to recompute have already been cleared.
#
# Spoiler: we don't do anything special, we just skip over them!
#
# def fn(x):
#   y = x.sin()                           # (1)
#   z = y.cos()                           # (2)
#   gx, = torch.autograd.grad(z, x)       # (3)
#   w = x.sin()                           # (4)
#   v = w.cos()                           # (5)
#   gx2, = torch.autograd.grad(v, x)      # (6)
#   return x * gx * gx2
#
# out = checkpoint(fn)(inp)
#
# In the code above fn is computed (potentially partially) 4 times in total.
#
# 1. Don't save x and y since we are inside a checkpoint.
# 2. Trigger a recompute of fn as we reach (3) since x and y weren't saved.
# 3. If early stop is enabled, stop at (2)
# 4. Continue original forward at (4), not saving x and w.
# 5. (5) triggers a recompute of fn
# 6. During recompute, we see that in the original graph, gx has already
#    cleared x and y since backward is run at (3) without retain_graph=True
#    We save x and w, however.
# 7. Continue with returning

_enable_checkpoint_early_stop = True

@contextlib.contextmanager
def set_checkpoint_early_stop(enable: bool):
    """Context manager that sets whether checkpoint should stop recomputation
    early.

    By default, non-reentrant checkpoint stops recomputation as soon as it
    has computed all needed Tensors. This context manager can be used to disable
    that feature if it is problematic for your specific application.

    This context manager only needs to be active when forward is run. It does
    not need to be active during backward.

    Example::

    >>> # xdoctest: +SKIP(failing)
    >>> message = "saved tensors default hooks are disabled"
    >>> with set_checkpoint_early_stop(False):
    ...     # Any checkpoint under this context manager will respect this
    ...     # context manager, even if its backward is performed outside.
    ...     out = checkpoint(fn, inputs)
    ...
    >>> out.backward()
    """
    global _enable_checkpoint_early_stop
    try:
        prev = _enable_checkpoint_early_stop
        _enable_checkpoint_early_stop = enable
        yield
    finally:
        _enable_checkpoint_early_stop = prev

class _Handle():
    pass

class _Holder():
    def __init__(self):
        self.handles: Dict[int, Optional[_Handle]] = dict()

class _NoopSaveInputs(torch.autograd.Function):
    @staticmethod
    def forward(*args):
        return torch.empty((0,))

    @staticmethod
    def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
        # Only tensors can be saved with ctx.save_for_backward, everything else
        # is captured by get_args, which is saved directly on ctx
        tensor_indices, tensors = zip(*[(i, o) for i, o in enumerate(inputs) if isinstance(o, torch.Tensor)])
        idx2saved_idx = {b: a for a, b in enumerate(tensor_indices)}
        # args but with tensors replaced with None as placeholders
        args = [None if isinstance(o, torch.Tensor) else o for o in inputs]

        def get_args(saved_tensors):
            # restore the placeholders with the original tensors grabbed from
            # ctx.saved_tensors (which may be saved on a parent checkpoint if
            # this checkpoint is nested, and that would trigger a recursive
            # unpack!)
            ret = [saved_tensors[idx2saved_idx[i]] if i in tensor_indices else o for i, o in enumerate(args)]
            # grab the tail since we also saved the dummy to avoid having to explicitly
            # handle the case where there are no tensor inputs
            return ret[1:]

        ctx.get_args = get_args
        ctx.save_for_backward(*tensors)

    @staticmethod
    def backward(ctx, *grad_outputs):
        raise AssertionError("Did not expect to backward on this graph")

class _CheckpointFrame():
    def __init__(self, recompute_fn, early_stop):
        self.recompute_fn = recompute_fn
        self.input_saver = None
        self.weak_holders: List[ReferenceType] = []
        # We store this as a weakkeydictionary so that in the case of a partial
        # backward, the entries in the dict are cleared alongside the Holder
        # which will be removed when the SavedVariable is cleared.
        self.recomputed: DefaultDict[int, weakref.WeakKeyDictionary[_Handle, torch.Tensor]] = \
            defaultdict(weakref.WeakKeyDictionary)
        # We need both recomp_counter and recomputed since they can diverge
        # https://github.com/pytorch/pytorch/pull/90105#discussion_r1135889885
        self.recomp_counter: DefaultDict[int, int] = defaultdict(int)
        self.is_recomputed: DefaultDict[int, bool] = defaultdict(bool)

        # See Rule 5
        self.early_stop = early_stop

# See Rule 5
class _StopRecomputationError(Exception):
    pass

class _recomputation_hook(torch.autograd.graph.saved_tensors_hooks):
    def __init__(self, target_frame_ref: ReferenceType, gid: int):
        def pack_hook(x):
            target_frame = target_frame_ref()
            assert target_frame is not None
            recomp_idx = target_frame.recomp_counter[gid]
            target_frame.recomp_counter[gid] += 1

            if recomp_idx >= len(target_frame.weak_holders):
                # We run into this case when early stop is not enabled and do
                # grad within checkpoint.
                return x.detach()
            holder = target_frame.weak_holders[recomp_idx]()

            if holder is not None:
                # See Rule 6: [ Multiple backwards ] above
                if holder.handles.get(gid, None) is None:
                    holder.handles[gid] = _Handle()
                target_frame.recomputed[gid][holder.handles[gid]] = x.detach()

            if target_frame.early_stop and \
               target_frame.recomp_counter[gid] == len(target_frame.weak_holders):
                raise _StopRecomputationError()
            # See Rule 6: [ Basic case ] above
            return x.detach()

        def unpack_hook(x):
            # See Rule 6: [ Basic case ] above for an example of when the graph
            # created during recomputation could be backwarded.
            return x

        super().__init__(pack_hook, unpack_hook)

class _checkpoint_hook(torch.autograd.graph.saved_tensors_hooks):
    def __init__(self, frame):
        def pack_hook(_unused_x):
            # See Rule 4 above
            holder = _Holder()
            frame.weak_holders.append(weakref.ref(holder))
            return holder

        def unpack_hook(holder):
            gid = torch._C._current_graph_task_id()
            if gid == -1:
                # generate a temporary id if we trigger unpack outside of a backward call
                gid = int(uuid.uuid4())

            if not frame.is_recomputed[gid]:
                ctx = frame.input_saver.grad_fn
                args = ctx.get_args(ctx.saved_tensors)

                try:
                    with _recomputation_hook(weakref.ref(frame), gid), torch.autograd.enable_grad():
                        frame.recompute_fn(*args)
                        if frame.early_stop:
                            raise AssertionError("if early stop is enabled, we don't expect to reach here")
                except _StopRecomputationError:
                    pass
                frame.is_recomputed[gid] = True

            if holder.handles[gid] is None:
                raise RuntimeError(
                    "If you are calling ctx.saved_tensor in backward, make sure to do so only once. "
                    "Otherwise please open an issue with details on your use case."
                )
            if holder.handles[gid] not in frame.recomputed[gid]:
                raise RuntimeError(
                    "Attempt to retrieve a tensor saved by autograd multiple times without checkpoint"
                    " recomputation being triggered in between, this is not currently supported. Please"
                    " open an issue with details on your use case."
                )
            ret = frame.recomputed[gid][holder.handles[gid]]
            holder.handles[gid] = None
            return ret

        super().__init__(pack_hook, unpack_hook)

# NB: this helper wraps fn before calling checkpoint_impl. kwargs and
#     saving/restoring of global state is handled here.
def _checkpoint_without_reentrant(
    fn,
    preserve_rng_state=True,
    context_fn: Callable[[], Tuple[ContextManager, ContextManager]] = noop_context_fn,
    swap_list=[],
    act_pack={},
    *args,
    **kwargs
):
    """Checkpointining without re-entrant autograd
    Args:
        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
            ``(activation, hidden)``, :attr:`function` should correctly use the
            first input as ``activation`` and the second input as ``hidden``
        preserve_rng_state(bool, optional):  Omit stashing and restoring
            the RNG state during each checkpoint.
            Default: ``True``
        context_fn(Callable, optional): A callable returning a tuple of two
            context managers. The function and its recomputation will be run
            under the first and second context managers respectively.
        *args: Arguments to pass in to the given ``function``.
        **kwargs: Keyword arguments to pass into the given ``function``.
    """
    import functools
    device = _infer_device_type(*args)
    device_module = _get_device_module(device)
    context_fn_partial = functools.partial(context_fn, swap_list, act_pack)
    forward_context, recompute_context = context_fn_partial()
    # Accommodates the (remote) possibility that autocast is enabled for cpu AND gpu.
    device_autocast_kwargs, cpu_autocast_kwargs = _get_autocast_kwargs(device=device)

    if preserve_rng_state:
        fwd_cpu_state = torch.get_rng_state()
        # Don't eagerly initialize the cuda context by accident.
        # (If the user intends that the context is initialized later, within their
        # run_function, we SHOULD actually stash the cuda state here.  Unfortunately,
        # we have no way to anticipate this will happen before we run the function.
        # If they do so, we raise an error.)
        had_device_in_fwd = False
        if device_module._initialized:
            had_device_in_fwd = True
            fwd_devices, fwd_device_states = get_device_states(*args)

    def recompute_fn(*inputs):
        kwargs, *args = inputs
        # This will be called later during recomputation. This wrapping enables
        # the necessary global state to be captured.
        rng_devices = []
        if preserve_rng_state and had_device_in_fwd:
            rng_devices = fwd_devices
        # with torch.random.fork_rng(devices=rng_devices, enabled=preserve_rng_state, device_type=device):
        with torch.random.fork_rng(devices=rng_devices, enabled=preserve_rng_state):
            if preserve_rng_state:
                torch.set_rng_state(fwd_cpu_state)
                if had_device_in_fwd:
                    set_device_states(fwd_devices, fwd_device_states)

            with device_module.amp.autocast(**device_autocast_kwargs), \
                 torch.cpu.amp.autocast(**cpu_autocast_kwargs), \
                 recompute_context:
                fn(*args, **kwargs)

    new_frame = _CheckpointFrame(recompute_fn, _enable_checkpoint_early_stop)
    dummy = torch.empty((0,), requires_grad=True)
    new_frame.input_saver = _NoopSaveInputs.apply(dummy, kwargs, *args)

    # When ambient grad_mode is False
    if new_frame.input_saver.grad_fn is None:
        return fn(*args, **kwargs)
    
    with _checkpoint_hook(new_frame), \
         forward_context:
        ret = fn(*args, **kwargs)

    if device_module._initialized and preserve_rng_state and not had_device_in_fwd:
        # Device was not initialized before running the forward, so we didn't
        # stash the device state.
        raise RuntimeError(
            "PyTorch's device state was initialized in the forward pass "
            "of a Checkpoint, which is not allowed. Please open an issue "
            "if you need this feature.")

    return ret