PyTorch Checkpoint 机制原理与源码解析

由于需要重复计算，所以随机状态的一致性是需要重视的。由于前向传播的部分在反向过程中仍会计算一次，所以如果不使用原始的随机状态的话，会导致重新计算和原本正常计算过程中的随机状态不同，而影响模型的行为。

另外在这段代码的注释中提到了一点有趣的地方：

由于无法获悉被 checkpoint 处理的操作是否在运算中间会将一些参数移动到不同的设备上，这可能需要手动保存这些设备对应的随机状态。当前的实现直接保存了所有可见设备上的随机状态，但是这样有时可能是不必要的，但是目前尚没有较好的解决策略。

所以按照文档的意思，就是在说如果没有这样的移动，那就可以不用保存随机状态咯？这一点其实有些令人疑惑。

# 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]]:
    """获取不同输入对应的 GPU 设备的随机数生成器的状态"""
    # This will not error out if "arg" is a CPU tensor or a non-tensor type because
    # the conditionals short-circuit.
    fwd_gpu_devices = list(set(arg.get_device() for arg in args
                               if isinstance(arg, torch.Tensor) and arg.is_cuda))

    fwd_gpu_states = []
    for device in fwd_gpu_devices:
        with torch.cuda.device(device):
            fwd_gpu_states.append(torch.cuda.get_rng_state())

    return fwd_gpu_devices, fwd_gpu_states

def set_device_states(devices, states) -> None:
    """针对不同的设备设置随机数生成器的状态"""
    for device, state in zip(devices, states):
        with torch.cuda.device(device):
            torch.cuda.set_rng_state(state)

核心 Function

可以看到，这里的 Checkpoint 本身就是基于 PyTorch 的 torch.autograd.Function 实现的一个扩展算子，所以该部分代码也涉及到了 Function 的诸多功能。

阅读它既可以帮助我们同时复习一下相关的知识，又能进一步了解更复杂的处理逻辑该如何搭建。

class CheckpointFunction(torch.autograd.Function):

    @staticmethod
    def forward(ctx, run_function, preserve_rng_state, *args):
        check_backward_validity(args)
        # 暂存前向传播函数
        ctx.run_function = run_function
        ctx.preserve_rng_state = preserve_rng_state
        # 用来保存当前模型的混合精度的状态，以用在反向传播中
        ctx.had_autocast_in_fwd = torch.is_autocast_enabled()
        if preserve_rng_state:  # 保存目标模块前向传播之前，此时 CPU 和 GPU 的随机数生成器的状态
            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_cuda_in_fwd = False
            if torch.cuda._initialized:  
                # PyTorch 提供的一个内部变量，用于判定 CUDA 状态是否已经被初始化了
                # torch.cuda.is_initialized 中就用到了该变量
                ctx.had_cuda_in_fwd = True
                # 保存输入变量涉及的各个 GPU 设备的随机状态
                ctx.fwd_gpu_devices, ctx.fwd_gpu_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)

        # save_for_backward() 中保存反向传播中需要用到的输入和输出 tensor 量。
        # 由于在反向传播中需要重新计算记录梯度的 output，所以就不要保存 output 了。
        # 并且后面的计算也不需要在梯度模式下计算。
        ctx.save_for_backward(*tensor_inputs)

        with torch.no_grad():  
            # 不保存梯度的前向传播操作，也就是说这里的 output 是不会记录中间变量，无法直接计算梯度的。
            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 # 获取前向传播中保存的输入 tensor

        # 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_cuda_in_fwd:
            rng_devices = ctx.fwd_gpu_devices
        
        # 使用之前前向传播开始之前保存的随机数生成器的状态来进行一次一模一样的前向传播过程
        with torch.random.fork_rng(devices=rng_devices, enabled=ctx.preserve_rng_state):
            # 使用上下文管理器保护原始的随机数生成器的状态，内部处理后在进行复原
            if ctx.preserve_rng_state:
                torch.set_rng_state(ctx.fwd_cpu_state)
                if ctx.had_cuda_in_fwd:
                    set_device_states(ctx.fwd_gpu_devices, ctx.fwd_gpu_states)
            # 这里将 inputs 从计算图中剥离开，但是其属性 requires_grad 和原来是一样的，这么做的目的是为了截断反向传播的路径。
            # 从整个操作目的来看，由于我们需要重新计算输出，并将梯度回传到输入上，所以输入本身需要可以记录梯度。
            # 但是这里的回传不可以影响到 checkpoint 之外更靠前的那些操作，
            # backward 之后会将之前保存的中间变量释放掉，而我们仅仅是为了计算当前一小块结构，所以梯度回传需要截断。
            detached_inputs = detach_variable(tuple(inputs))  # 会变成叶子结点，grad 和 grad_fn 均重置为 None
            # 处理完随机状态之后，就该准备着手重新前向传播了。
            # 这次前向传播是在梯度模式 (torch.enable_grad()) 下执行的。此时会保存中间变量。
            with torch.enable_grad(), torch.cuda.amp.autocast(ctx.had_autocast_in_fwd):
                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)):
            # 记录需要计算梯度的输出 outputs[i] 以及对应的回传回来的有效梯度 args[i]
            if torch.is_tensor(outputs[i]) and outputs[i].requires_grad:
                outputs_with_grad.append(outputs[i])
                args_with_grad.append(args[i])
        # 检查需要计算梯度的输出，如果没有输出需要计算梯度，那么实际上就说明这个模块是不参与梯度计算的，
        # 也就是说，该模块不需要使用 checkpoint 来调整。
        if len(outputs_with_grad) == 0:
            raise RuntimeError(
                "none of output has requires_grad=True,"
                " this checkpoint() is not necessary")
        # 该操作对被包裹的目标操作计算反向传播，即计算回传到输入 detached_inputs 上的梯度。
        # 由于输入的 tensor 已被从整体梯度图中剥离，所以可以看做是一个叶子结点，可以在反向传播之后获得其梯度，并且中间变量也会随之释放。
        # 另外这里反传计算梯度也不会导致将更靠前的结构中暂时保存来计算梯度的参数给释放掉。
        torch.autograd.backward(outputs_with_grad, args_with_grad)
        # 如果前面不执行 detach()，这里的 inp.grad 会被直接释放并置为 None，这并不符合预期
        grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else None
                      for inp in detached_inputs)

        # 这里返回的梯度与当前类的 forward 输入一一对应，
        # 由于这里的 forward 包含着本不需要梯度的两个参数 run_function、preserve_rng_state，故对应回传 None 即可。
        return (None, None) + grads

这里实际上就是在原始的操作和整体的计算图之间添加了一个中间层，用于信息的交互：

1. 原始模型的数据传输到被包裹的目标层的时候，数据进入 checkpoint 的 forward() 中，被 checkpoint 进行检查和记录后，再送入目标层中；
2. 目标层在非梯度模式下执行前向传播。该模式下，新创建的 tensor 都是不会记录梯度信息的；
3. 目标层的结果通过 checkpoint 的前向传播输出，送入模型后续的其他结构中；
4. 执行反向传播，损失求导，链式回传，计算梯度；
5. 回传回来的对应于 checkpoint 输出的梯度被送入其对应的反向传播函数，即 checkpoint 的 backward()。
6. 梯度送入 checkpoint 中后，需要进一步将梯度回传到目标层的输入上。由于在 checkpoint 的 forward 中目标层本身前向传播是处于非梯度状态下，所以回传路径上缺少目标层中操作的梯度子图。于是为了获取这部分信息，需要先梯度状态下对目标层进行一次前向传播，通过将回传回来的梯度和目标层的输出一起执行 torch.autograd.backward(outputs_with_grad, args_with_grad)，从而获得对应输入的梯度信息。
7. 将对应目标操作输入的梯度信息按照 checkpoint 本身 Function 的 backward 的需求，使用 None 对其他辅助参数的梯度占位后进行返回。
8. 返回的对应于其他模块的输出量的梯度，被沿着反向传播的路径送入对应操作的 backward 中，一层一层回传累加到各个叶子节点上。

定义好操作后，进行一个简单的包装，同时处理一下默认参数，补充了更细致的文档：

def checkpoint(function, *args, use_reentrant: bool = True, **kwargs):
    r"""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, :attr:`function` will run in
    :func:`torch.no_grad` manner, i.e., not storing the intermediate
    activations. Instead, the forward pass saves the inputs tuple and the
    :attr:`function` parameter. In the backwards pass, the saved inputs and
    :attr:`function` is retrieved, and the forward pass is computed on
    :attr:`function` again, now tracking the intermediate activations, and then
    the gradients are calculated using these activation values.
    这一段详细介绍了 checkpoint 的核心技术，也就是在非梯度模式下执行目标操作的前向传播，只保留输入和结构参数，省去了中间激活的保存。反向传播时在梯度模式下重新计算这些激活，重建这部分反向图，进而实现了梯度的正常回传。
    
    The output of :attr:`function` can 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.
    因为 checkpoint 的 backward 实现的逻辑中，直接遍历目标操作的输出（会被自定转换成元组类型）并确定那些需要回流梯度的输出。如果输出中包含其他的非 tensor 结构，就会导致在遍历过程中这些输出被忽略掉。不过也确实，这样直接简化处理虽然使得灵活性下降，但是却也避免了代码过于复杂。
    
    .. warning::
        Checkpointing currently only supports :func:`torch.autograd.backward`
        and only if its `inputs` argument is not passed. :func:`torch.autograd.grad`
        is not supported.
    
    .. warning::
        If :attr:`function` invocation 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.
        尽量保证目标操作在反向计算期间和前向期间的操作的一致性。
        因为在 checkpoint 会在反向中重新计算一次前向，这可能会带来一些由于无法检测到的不确定因素而造成的与常规版本的差异。
        
    .. warning::
        If 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.
        不要在目标操作中包含 detach 或者非梯度模式的处理。
        **在我的实际测试中似乎并没有这个问题？**或许这里应该看一下 pytorch 提供的测试案例。
        
    .. warning::
        At least one of the inputs needs to have :code:`requires_grad=True` if
        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
        :code:`requires_grad=True` as well.
        要保证至少有一个输入是 requires_grad 的，这样才可以保证这部分操作可以被记录梯度。
        也要保证输出至少有一个需要计算梯度。

    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, default=True):  Omit stashing and restoring
            the RNG state during each checkpoint.
        args: tuple containing inputs to the :attr:`function`

    Returns:
        Output of running :attr:`function` on :attr:`*args`
    """
    # Hack to mix *args with **kwargs 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))

    return CheckpointFunction.apply(function, preserve, *args)

应用案例

Checkpoint for Sequential

PyTorch 源码中给了一个很直接的应用案例，就是将 checkpoint 应用于 Sequential 搭建起来的模型。按照分段数 segments 指定的，将模型划分为多段。

def checkpoint_sequential(functions, segments, input, **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 run in
    :func:`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.

    See :func:`~torch.utils.checkpoint.checkpoint` on how checkpointing works.

    .. warning::
        Checkpointing currently only supports :func:`torch.autograd.backward`
        and only if its `inputs` argument is not passed. :func:`torch.autograd.grad`
        is not supported.

    .. warning:
        At least one of the inputs needs to have :code:`requires_grad=True` if
        grads are needed for model inputs, otherwise the checkpointed part of the
        model won't have gradients.

    .. 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, default=True):  Omit stashing and restoring
            the RNG state during each checkpoint.

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

    Example:
        >>> 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()) 
        # 获取 Sequential 的子模块，这里使用 children 方法，仅获取最外层

    segment_size = len(functions) // segments
    # the last chunk has to be non-volatile（为什么？似乎加上也是可以的）
    end = -1
    for start in range(0, segment_size * (segments - 1), segment_size):
        end = start + segment_size - 1
        # 迭代式的将各个子模块集合使用 checkpoint 包装并前向传播。
        input = checkpoint(run_function(start, end, functions), input,
                           preserve_rng_state=preserve)
    # 剩余的结构不再使用 checkpoint
    return run_function(end + 1, len(functions) - 1, functions)(input)

总结

PyTorch 的 Checkpoint 机制是深度学习训练中显存优化的重要工具。它通过牺牲一定的计算时间来换取显存空间的节省，特别适用于深层网络或 Batch Size 受限的场景。在使用时，开发者需注意以下几点：

1. 梯度要求：确保至少有一个输入和一个输出设置了 requires_grad=True，否则 Checkpoint 将无效。
2. 随机性一致性：开启 preserve_rng_state 以保证反向重计算时的随机状态与前向一致，这对包含 Dropout 等随机操作的模型至关重要。
3. 兼容性限制：不支持 .grad() 接口，需使用 .backward()；且被 Checkpoint 包裹的段内不应包含额外的 detach() 或 no_grad() 操作。
4. 适用场景：适合内存敏感但计算资源相对充裕的情况，如训练 Transformer 类大模型或高分辨率图像分割网络。

合理运用 Checkpoint 机制，可以在有限的硬件资源下训练更大的模型，是构建高效深度学习系统的关键技术之一。