pytorch/torch/cuda/graphs.py

# mypy: allow-untyped-defs
import contextlib
import gc
import typing

import torch

from .._utils import _dummy_type


if not hasattr(torch._C, "_CudaStreamBase"):
    # Define dummy base classes
    torch._C.__dict__["_CUDAGraph"] = _dummy_type("_CUDAGraph")
    torch._C.__dict__["_graph_pool_handle"] = _dummy_type("_graph_pool_handle")
    torch._C.__dict__["_cuda_isCurrentStreamCapturing"] = _dummy_type(
        "_cuda_isCurrentStreamCapturing"
    )

from torch._C import (  # noqa: F401
    _cuda_isCurrentStreamCapturing,
    _CUDAGraph,
    _graph_pool_handle,
)


def is_current_stream_capturing():
    r"""Return True if CUDA graph capture is underway on the current CUDA stream, False otherwise.

    If a CUDA context does not exist on the current device, returns False without initializing the context.
    """
    return _cuda_isCurrentStreamCapturing()


# Python shim helps Sphinx process docstrings more reliably.
def graph_pool_handle():
    r"""Return an opaque token representing the id of a graph memory pool.

    See :ref:`Graph memory management<graph-memory-management>`.

    .. warning::
        This API is in beta and may change in future releases.
    """
    return _graph_pool_handle()


# Python shim helps Sphinx process docstrings more reliably.
class CUDAGraph(torch._C._CUDAGraph):
    r"""Wrapper around a CUDA graph.

    .. warning::
        This API is in beta and may change in future releases.
    """

    def __new__(cls):
        return super().__new__(cls)

    def capture_begin(self, pool=None, capture_error_mode="global"):
        r"""Begin capturing CUDA work on the current stream.

        Typically, you shouldn't call ``capture_begin`` yourself.
        Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`,
        which call ``capture_begin`` internally.

        Arguments:
            pool (optional): Token (returned by :func:`~torch.cuda.graph_pool_handle` or
                :meth:`other_Graph_instance.pool()<torch.cuda.CUDAGraph.pool>`) that hints this graph may share memory
                with the indicated pool.  See :ref:`Graph memory management<graph-memory-management>`.
            capture_error_mode (str, optional): specifies the cudaStreamCaptureMode for the graph capture stream.
                Can be "global", "thread_local" or "relaxed". During cuda graph capture, some actions, such as cudaMalloc,
                may be unsafe. "global" will error on actions in other threads, "thread_local" will only error for
                actions in the current thread, and "relaxed" will not error on these actions. Do NOT change this setting
                unless you're familiar with `cudaStreamCaptureMode <https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85>`_
        """  # noqa: B950
        super().capture_begin(pool=pool, capture_error_mode=capture_error_mode)

    def capture_end(self):
        r"""End CUDA graph capture on the current stream.

        After ``capture_end``, ``replay`` may be called on this instance.

        Typically, you shouldn't call ``capture_end`` yourself.
        Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`,
        which call ``capture_end`` internally.
        """
        super().capture_end()

    def replay(self):
        r"""Replay the CUDA work captured by this graph."""
        super().replay()

    def reset(self):
        r"""Delete the graph currently held by this instance."""
        super().reset()

    def pool(self):
        r"""Return an opaque token representing the id of this graph's memory pool.

        This id can optionally be passed to another graph's ``capture_begin``,
        which hints the other graph may share the same memory pool.
        """
        return super().pool()

    def enable_debug_mode(self):
        r"""Enable debugging mode for CUDAGraph.debug_dump."""
        return super().enable_debug_mode()

    def debug_dump(self, debug_path):
        r"""
        Arguments:
            debug_path (required): Path to dump the graph to.

        Calls a debugging function to dump the graph if the debugging is
        enabled via CUDAGraph.enable_debug_mode()
        """
        return super().debug_dump(debug_path)


class graph:
    r"""Context-manager that captures CUDA work into a :class:`torch.cuda.CUDAGraph` object for later replay.

    See :ref:`CUDA Graphs <cuda-graph-semantics>` for a general introduction,
    detailed use, and constraints.

    Arguments:
        cuda_graph (torch.cuda.CUDAGraph): Graph object used for capture.
        pool (optional): Opaque token (returned by a call to :func:`~torch.cuda.graph_pool_handle()` or
            :meth:`other_Graph_instance.pool()<torch.cuda.CUDAGraph.pool>`) hinting this graph's capture
            may share memory from the specified pool. See :ref:`Graph memory management<graph-memory-management>`.
        stream (torch.cuda.Stream, optional): If supplied, will be set as the current stream in the context.
            If not supplied, ``graph`` sets its own internal side stream as the current stream in the context.
        capture_error_mode (str, optional): specifies the cudaStreamCaptureMode for the graph capture stream.
            Can be "global", "thread_local" or "relaxed". During cuda graph capture, some actions, such as cudaMalloc,
            may be unsafe. "global" will error on actions in other threads, "thread_local" will only error for
            actions in the current thread, and "relaxed" will not error on actions. Do NOT change this setting
            unless you're familiar with `cudaStreamCaptureMode <https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85>`_
        collect_garbage (bool, optional): If True, call torch.cuda.synchronize() followed by gc.collect() to free
            memory before starting graph capture. Users almost always this to be True, but since the introduction of
            conditional nodes in cuda graphs, it is possible that more than one stream may be capturing at once.
            Since cudaDeviceSynchronize() synchronizes all streams, including capturing streams, previously started
            stream captures will be invalidated. This is not desirable.

    .. note::
        For effective memory sharing, if you pass a ``pool`` used by a previous capture and the previous capture
        used an explicit ``stream`` argument, you should pass the same ``stream`` argument to this capture.

    .. warning::
        This API is in beta and may change in future releases.

    .. _cudaStreamCaptureMode:
        https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85
    """  # noqa: B950

    default_capture_stream: typing.Optional["torch.cuda.Stream"] = None

    def __init__(
        self, cuda_graph, pool=None, stream=None, capture_error_mode: str = "global"
    ):
        # Lazy-init of default_capture_stream helps avoid circular-import errors.
        # Not thread safe, but graphs already have the general (explicitly documented)
        # restriction that only one capture may be underway at a time in the process.
        if self.__class__.default_capture_stream is None:
            self.__class__.default_capture_stream = torch.cuda.Stream()

        self.pool = () if pool is None else (pool,)
        self.capture_stream = (
            stream if stream is not None else self.__class__.default_capture_stream
        )
        assert self.capture_stream is not None
        self.stream_ctx = torch.cuda.stream(self.capture_stream)
        self.cuda_graph = cuda_graph
        self.capture_error_mode = capture_error_mode

    def __enter__(self):
        # Free as much memory as we can for the graph.
        torch.cuda.synchronize()
        gc.collect()
        torch.cuda.empty_cache()

        # Stackoverflow seems comfortable with this pattern
        # https://stackoverflow.com/questions/26635684/calling-enter-and-exit-manually#39172487
        self.stream_ctx.__enter__()

        self.cuda_graph.capture_begin(
            *self.pool, capture_error_mode=self.capture_error_mode
        )

    def __exit__(self, exc_type, exc_value, traceback):
        self.cuda_graph.capture_end()
        self.stream_ctx.__exit__(exc_type, exc_value, traceback)
        # returning None should propagate exceptions from either capture_end or stream_ctx.__exit__()


@contextlib.contextmanager
def _graph_no_gc(cuda_graph, pool, stream, capture_error_mode):
    """This is an internal function used to do stream capture without
    calling torch.cuda.synchronize(), gc.collect(), and
    torch.cuda.empty_cache(). Unfortunately, cudagraph trees runs its
    eager warmup inside of the context manager
    _use_cuda_memory_pool_manager(), which makes captures_underway in
    CUDACachingAllocator.cpp non-empty. We need this in order to
    warmup conditional higher order operators, like torch.cond() and
    torch.while_loop(). torch.cuda.empty_cache() will fail if
    captures_underway is non-empty. Removing torch.cuda.synchronize()
    and gc.collect() is not strictly speaking required, but they are
    expensive an unnecessary operations.
    """
    stream_ctx = torch.cuda.stream(stream)
    pool = () if pool is None else (pool,)
    with stream_ctx:
        cuda_graph.capture_begin(*pool, capture_error_mode=capture_error_mode)
        try:
            yield
        finally:
            cuda_graph.capture_end()


def make_graphed_callables(
    callables, sample_args, num_warmup_iters=3, allow_unused_input=False, pool=None
):
    r"""Accept callables (functions or :class:`nn.Module<torch.nn.Module>`\ s) and returns graphed versions.

    Each graphed callable's forward pass runs its source callable's
    forward CUDA work as a CUDA graph inside a single autograd node.

    The graphed callable's forward pass also appends
    a backward node to the autograd graph. During backward, this node runs the
    callable's backward work as a CUDA graph.

    Therefore, each graphed callable should be a drop-in replacement for its source callable
    in an autograd-enabled training loop.

    See :ref:`Partial-network capture<partial-network-capture>` for detailed use and constraints.

    If you pass a tuple of several callables, their captures will use the same memory pool.
    See :ref:`Graph memory management<graph-memory-management>` for when this is appropriate.

    Arguments:
        callables (torch.nn.Module or Python function, or tuple of these): Callable or callables to graph.
            See :ref:`Graph memory management<graph-memory-management>` for when passing a tuple of callables
            is appropriate.  If you pass a tuple of callables, their order in the tuple must be the same order
            they'll run in the live workload.
        sample_args (tuple of Tensors, or tuple of tuples of Tensors): Samples args for each callable.
            If a single callable was passed, ``sample_args`` must be a single tuple of argument Tensors.
            If a tuple of callables was passed, ``sample_args`` must be tuple of tuples of argument Tensors.
        num_warmup_iters (int): The number of warmup iterations. Currently, ``DataDistributedParallel`` needs
            11 iterations for warm up. Default: ``3``.
        allow_unused_input (bool): If False, specifying inputs that were not used when computing outputs
            (and therefore their grad is always zero) is an error. Defaults to False.
        pool (optional): Token (returned by :func:`~torch.cuda.graph_pool_handle` or
            :meth:`other_Graph_instance.pool()<torch.cuda.CUDAGraph.pool>`) that hints this graph may share memory
            with the indicated pool.  See :ref:`Graph memory management<graph-memory-management>`.
    .. note::
        The ``requires_grad`` state of each Tensor in ``sample_args`` must match the state
        that's expected for the corresponding real input in the training loop.

    .. warning::
        This API is in beta and may change in future releases.

    .. warning::
        ``sample_args`` for each callable must contain only Tensors. Other types are not allowed.

    .. warning::
        Returned callables do not support higher order differentiation (e.g., double backward).

    .. warning::
        In any :class:`~torch.nn.Module` passed to :func:`~make_graphed_callables`, only parameters
        may be trainable. Buffers must have ``requires_grad=False``.

    .. warning::
        After you pass a :class:`torch.nn.Module` through :func:`~make_graphed_callables`,
        you may not add or remove any of that Module's parameters or buffers.

    .. warning::
        :class:`torch.nn.Module`\s passed to :func:`~torch.cuda.make_graphed_callables` must not have module hooks
        registered on them at the time they are passed. However, registering hooks on modules *after* passing them
        through :func:`~torch.cuda.make_graphed_callables` is allowed.

    .. warning::
        When running a graphed callable, you must pass its arguments in the same order and format
        they appeared in that callable's ``sample_args``.

    .. warning::
        The automatic mixed precision is supported in :func:`~torch.cuda.make_graphed_callables` only with disabled
        caching. The context manager `torch.cuda.amp.autocast()` must have `cache_enabled=False`.
    """
    if torch.is_autocast_enabled() and torch.is_autocast_cache_enabled():
        raise RuntimeError(
            "make_graphed_callables does not support the autocast caching. Please set `cache_enabled=False`."
        )

    just_one_callable = False

    if not isinstance(callables, tuple):
        just_one_callable = True
        callables = (callables,)
        sample_args = (sample_args,)

    flatten_sample_args = []

    for c, args in zip(callables, sample_args):
        if isinstance(c, torch.nn.Module):
            assert (
                len(c._backward_hooks) == 0
                and len(c._forward_hooks) == 0
                and len(c._forward_pre_hooks) == 0
            ), (
                "Modules must not have hooks registered at the time they are passed. However, registering hooks "
                + "on modules after passing them through make_graphed_callables is allowed."
            )
            assert all(b.requires_grad is False for b in c.buffers()), (
                "In any :class:`~torch.nn.Module` passed to "
                + ":func:`~make_graphed_callables`, only parameters may be trainable. All buffers must have "
                + "``requires_grad=False``."
            )
        flatten_arg = torch.utils._pytree.arg_tree_leaves(*args)
        flatten_sample_args.append(tuple(flatten_arg))
        assert all(isinstance(arg, torch.Tensor) for arg in flatten_arg), (
            "In the beta API, sample_args "
            + "for each callable must contain only Tensors. Other types are not allowed."
        )

    # If a callable is an nn.Module, its graph's full input surface is the args the user explicitly
    # passes to forward (ie, its sample_args) AND the module's parameter attributes.
    per_callable_len_user_args = [len(args) for args in flatten_sample_args]
    per_callable_module_params = [
        tuple(c.parameters()) if isinstance(c, torch.nn.Module) else ()
        for c in callables
    ]
    per_callable_static_input_surfaces = [
        flatten_sample_args[i] + per_callable_module_params[i]
        for i in range(len(callables))
    ]

    fwd_graphs = [torch.cuda.CUDAGraph() for _ in range(len(callables))]
    bwd_graphs = [torch.cuda.CUDAGraph() for _ in range(len(callables))]

    mempool = graph_pool_handle() if pool is None else pool

    # Warmup
    # Hopefully prevents cudnn benchmarking and other lazy-initialization cuda work
    # from ending up in any captures.
    torch.cuda.synchronize()
    with torch.cuda.stream(torch.cuda.Stream()):
        for func, args, static_input_surface in zip(
            callables, sample_args, per_callable_static_input_surfaces
        ):
            grad_inputs, outputs, outputs_grad = None, None, None
            for _ in range(num_warmup_iters):
                outputs = torch.utils._pytree.tree_leaves(func(*args))
                outputs_grad = tuple(o for o in outputs if o.requires_grad)
                if len(outputs_grad) > 0:
                    grad_inputs = torch.autograd.grad(
                        outputs=outputs_grad,
                        inputs=tuple(
                            i for i in static_input_surface if i.requires_grad
                        ),
                        grad_outputs=tuple(
                            torch.empty_like(o) for o in outputs if o.requires_grad
                        ),
                        only_inputs=True,
                        allow_unused=allow_unused_input,
                    )
            for v in [outputs, outputs_grad, grad_inputs]:
                del v

    torch.cuda.synchronize()

    # All captures here share a mempool. To avoid replays corrupting each other's memory,
    # the safest approach is to capture all passes in the same order they'll run:
    # fwd 1, fwd 2, ... fwd N, then bwd N, bwd N-1, ... bwd 1.

    # Capture forward graphs
    per_callable_static_outputs = []
    per_callable_output_unflatten_spec = []
    for func, args, fwd_graph in zip(callables, sample_args, fwd_graphs):
        with torch.cuda.graph(fwd_graph, pool=mempool):
            outputs = func(*args)

        flatten_outputs, spec = torch.utils._pytree.tree_flatten(outputs)
        per_callable_static_outputs.append(tuple(flatten_outputs))
        per_callable_output_unflatten_spec.append(spec)

    # Capture backward graphs in reverse order
    per_callable_static_grad_outputs = []
    per_callable_static_grad_inputs = []
    for static_input_surface, static_outputs, bwd_graph in zip(
        reversed(per_callable_static_input_surfaces),
        reversed(per_callable_static_outputs),
        reversed(bwd_graphs),
    ):
        # For now, assumes all static_outputs require grad
        # assert all(o.requires_grad for o in static_outputs), "Outputs of graphed callables must require grad."
        static_grad_outputs = tuple(
            torch.empty_like(o) if o.requires_grad else None for o in static_outputs
        )

        outputs_grad = tuple(o for o in static_outputs if o.requires_grad)
        grad_inputs = None
        if len(outputs_grad) > 0:
            with torch.cuda.graph(bwd_graph, pool=mempool):
                grad_inputs = torch.autograd.grad(
                    outputs=outputs_grad,
                    inputs=tuple(i for i in static_input_surface if i.requires_grad),
                    grad_outputs=tuple(o for o in static_grad_outputs if o is not None),
                    only_inputs=True,
                    allow_unused=allow_unused_input,
                )

        # Constructs a tuple suitable for returning from Graphed.backward:
        # Pads out the actually-needed grads with Nones in gradient slots for inputs that don't require grad.
        # I couldn't think of a slick one-liner for this pattern.
        static_grad_inputs = []
        grad_idx = 0
        for arg in static_input_surface:
            if arg.requires_grad and grad_inputs is not None:
                static_grad_inputs.append(grad_inputs[grad_idx])
                grad_idx += 1
            else:
                static_grad_inputs.append(None)  # type: ignore[arg-type]
        static_grad_inputs = tuple(static_grad_inputs)  # type: ignore[assignment]

        per_callable_static_grad_outputs.append(static_grad_outputs)
        per_callable_static_grad_inputs.append(static_grad_inputs)

    # Reverses the most recent two lists
    per_callable_static_grad_outputs.reverse()
    per_callable_static_grad_inputs.reverse()
    # Now for every per_callable list, per_callable_*[i] holds the stuff for the ith callable.

    def make_graphed_autograd_function(
        fwd_graph,
        bwd_graph,
        module_params,
        len_user_args,
        output_unflatten_spec,
        static_input_surface,
        static_outputs,
        static_grad_outputs,
        static_grad_inputs,
    ):
        class Graphed(torch.autograd.Function):
            @staticmethod
            def forward(ctx, *inputs):
                # At this stage, only the user args may (potentially) be new tensors.
                for i in range(len_user_args):
                    if static_input_surface[i].data_ptr() != inputs[i].data_ptr():
                        static_input_surface[i].copy_(inputs[i])
                fwd_graph.replay()
                assert isinstance(static_outputs, tuple)
                return tuple(o.detach() for o in static_outputs)

            @staticmethod
            @torch.autograd.function.once_differentiable
            def backward(ctx, *grads):
                assert len(grads) == len(static_grad_outputs)
                for g, grad in zip(static_grad_outputs, grads):
                    if g is not None:
                        # don't copy if autograd gods have been kind and the
                        # incoming grad is already in the right place
                        if g.data_ptr() != grad.data_ptr():
                            g.copy_(grad)
                bwd_graph.replay()

                # Input args that didn't require grad expect a None gradient.
                assert isinstance(static_grad_inputs, tuple)
                return tuple(
                    b.detach() if b is not None else b for b in static_grad_inputs
                )

        def functionalized(*user_args):
            # Runs the autograd function with inputs == all inputs to the graph that might require grad
            # (explicit user args + module parameters)
            # Assumes module params didn't change since capture.
            flatten_user_args = torch.utils._pytree.arg_tree_leaves(*user_args)
            out = Graphed.apply(*(tuple(flatten_user_args) + module_params))
            return torch.utils._pytree.tree_unflatten(out, output_unflatten_spec)

        return functionalized

    # Put together the final graphed callables
    ret = []
    for i, func in enumerate(callables):
        graphed = make_graphed_autograd_function(
            fwd_graphs[i],
            bwd_graphs[i],
            per_callable_module_params[i],
            per_callable_len_user_args[i],
            per_callable_output_unflatten_spec[i],
            per_callable_static_input_surfaces[i],
            per_callable_static_outputs[i],
            per_callable_static_grad_outputs[i],
            per_callable_static_grad_inputs[i],
        )

        if isinstance(func, torch.nn.Module):

            def make_graphed_forward(func, graph_training_state, graphed, orig_fwd):
                def new_fwd(*user_args):
                    # If the module's training-or-eval state matches what we graphed,
                    # run the graph, otherwise run the original forward method
                    if func.training == graph_training_state:
                        return graphed(*user_args)
                    else:
                        return orig_fwd(*user_args)

                return new_fwd

            func.forward = make_graphed_forward(func, func.training, graphed, func.forward)  # type: ignore[assignment]
            ret.append(func)
        else:
            ret.append(graphed)

    if just_one_callable:
        return ret[0]

    return tuple(ret)


@contextlib.contextmanager
def thread_cuda_stream_capture_mode(new_mode):
    r"""Changes current thread's stream capture mode to `new_mode` upon __enter__ and resets the mode upon __exit__.

    The only documentation on a thread's stream capture mode is here:
    https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85

    However, it is a little bit inadequate, so here is a more in-depth description.

    Both CPU threads and capturing cuda streams have a capture mode. A
    cuda stream's capture mode is set at cudaStreamBeginCapture() and
    can never be changed. Meanwhile all CPU threads start with a capture
    mode of cudaStreamCaptureModeGlobal, which can be changed at any
    time.

    Whenever a thread executes an unsafe CUDA action while CUDA
    streams are capturing, it follows the following logic to determine
    whether to invalidate those streams:

    if capture_mode_this_thread == cudaStreamCaptureModeRelaxed:
        never invalidate any capturing cuda streams whatsoever.
    elif capture_mode_this_thread == cudaStreamCaptureModeThreadLocal:
        invalidate any cuda streams for which cudaStreamBeginCapture() was called by this
        thread, except for streams whose capture mode is cudaStreamCaptureModeRelaxed.
    elif capture_mode_this_thread == cudaStreamCaptureModeGlobal:
        invalidate all cuda streams that are currently capturing on any thread,
        except for streams whose capture mode is cudaStreamCaptureModeRelaxed and for
        streams for which cudaStreamCaptureBegin() was called with
        cudaStreamCaptureModeThreadLocal on a thread other than this one.

    In practice, changed the current capture mode to
    cudaStreamCaptureModeRelaxed in particular is helpful for enabling
    developers to do "unsafe" things that we know are safe in our
    case.
    """
    cudart = torch.cuda.cudart()
    old_mode = cudart.cudaThreadExchangeStreamCaptureMode(new_mode)
    try:
        yield
    finally:
        cudart.cudaThreadExchangeStreamCaptureMode(old_mode)