fully_sharded_data_parallel.py

# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the BSD license found in the
# LICENSE file in the root directory of this source tree.

import contextlib
import copy
from enum import Enum, auto
import functools
from math import inf
from typing import TYPE_CHECKING, Any, Callable, Dict, Generator, List, NamedTuple, Optional, Tuple, Union

import torch
from torch.autograd import Variable
import torch.distributed as dist
from torch.distributed import ProcessGroup
import torch.nn as nn
from torch.nn import Parameter
import torch.nn.functional as F

from fairscale.nn.misc import FlattenParamsWrapper
from fairscale.optim.utils import calc_grad_norm
from fairscale.utils.containers import apply_to_tensors
from fairscale.utils.parallel import chunk_and_pad, validate_process_group
from fairscale.utils.reduce_scatter_bucketer import ReduceScatterBucketer
from fairscale.utils.state_dict import replace_by_prefix_

if TYPE_CHECKING:
    from collections import OrderedDict  # noqa: F401


class TrainingState(Enum):
    """
    Simple enum to indicate what state FSDP is in. Used for asserting
    to make sure APIs are called in the correct state.

    TODO (Min): It would be nice to capture the stepping state as well.
        Maybe we can use the model.zero_grad() call, but not sure if it
        is called if optim.zero_grad() is used instead.
        It would be nice to have clear state transition be explicit like:

        zero_grad -> fwd -> bwd -> optionally accum grad by repeating
        fwd/bwd -> stepping -> loop back to zero_grad
    """

    IDLE = auto()
    FORWARD = auto()
    BACKWARD = auto()
    SUMMON_FULL_PARAMS = auto()


class FullyShardedDataParallel(nn.Module):
    """
    A wrapper for sharding Module parameters across data parallel workers. This
    is inspired by `Xu et al.`_ as well as the ZeRO Stage 3 from DeepSpeed_.

    .. _`Xu et al.`: https://arxiv.org/abs/2004.13336
    .. _DeepSpeed: https://www.deepspeed.ai/

    Usage::

        torch.cuda.set_device(device_id)
        sharded_module = FullyShardedDataParallel(my_module)
        optim = torch.optim.Adam(sharded_module.parameters(), lr=0.0001)
        x = sharded_module(x, y=3, z=torch.Tensor([1]))
        loss = x.sum()
        loss.backward()
        optim.step()

    It is also possible to shard individual layers separately and have an outer
    wrapper handle any leftover parameters. This can be helpful to further
    reduce GPU memory usage, reduce system memory usage when initializing large
    models and to improve training speed by overlapping the all-gather step
    across the forward pass. For example::

        sharded_model = FullyShardedDataParallel(
            nn.Sequential(  # doesn't have to be nn.Sequential
                nn.Linear(5, 100),
                FullyShardedDataParallel(nn.Linear(100, 100)),
                FullyShardedDataParallel(nn.Linear(100, 100)),
                nn.Linear(100, 5),
            )
        )

    .. warning::

        The optimizer must be initialized *after* the module has been wrapped,
        since FSDP will shard parameters in-place and this will break any
        previously initialized optimizers.

    .. warning::

        If you wrap every parameter inside a nested FSDP and leaving the outer
        FSDP empty without any parameter, checkpointing activation may trigger
        an assert on the backward pass. The solution is to leave some parameters
        to the outer FSDP.

    Args:
        module (nn.Module):
            module to checkpoint
        process_group (Optional):
            process group for sharding
        reshard_after_forward (bool, Optional):
            if ``True``, reshard parameters after the forward pass. This saves
            memory but slows training. This is only relevant when resharding
            individual layers.
        mixed_precision (bool, Optional):
            if ``True``, inputs, activations and gradients will be kept in FP16;
            computation and communication will occur in FP16; and a (sharded)
            master copy of the model weights will be maintained in FP32.
        fp32_reduce_scatter (bool, Optional):
            if ``True``, then reduce-scatter gradients in FP32. This is only
            relevant when *``mixed_precision``* is ``True``.
        flatten_parameters (bool, Optional):
            if ``True``, flatten parameters into a single contiguous tensor,
            which improves training speed.
        cpu_offload (bool, Optional):
            if ``True``, offload FP32 params to CPU. This is only relevant when
            *``mixed_precision``* is ``True``.
        compute_dtype (torch.dtype, Optional):
            dtype for full parameters for computation. This defaults to
            ``torch.float32`` unless *``mixed_precision``* is set, in which case
            it defaults to ``torch.float16``.
        buffer_dtype (torch.dtype, Optional):
            dtype for buffers for computation. This defaults to ``compute_dtype``.
        move_grads_to_cpu (bool, Optional):
            move gradient shard to CPU after reduction. This is useful when
            combined with CPU-based optimizers. It defaults to the value of
            *``cpu_offload``*.
        bucket_cap_mb (int, Optional):
            FSDP will bucket parameters so that gradient reduction can
            potentially overlap with backward computation. bucket_cap_mb
            controls the bucket size in MegaBytes (MB). Buckets are sub-divided
            based on world_size, so the max shard size is roughly
            ``bucket_cap_mb / world_size``. Values <= 0 disable bucketing.
            Default: 25.
    """

    def __init__(
        self,
        module: nn.Module,
        process_group: Optional[ProcessGroup] = None,
        reshard_after_forward: bool = True,
        mixed_precision: bool = False,
        fp32_reduce_scatter: bool = False,
        flatten_parameters: bool = True,
        cpu_offload: bool = False,
        compute_dtype: Optional[torch.dtype] = None,
        buffer_dtype: Optional[torch.dtype] = None,
        move_grads_to_cpu: Optional[bool] = None,
        bucket_cap_mb: int = 25,
    ):
        super().__init__()
        self.process_group = process_group or dist.new_group()
        self.rank = self.process_group.rank()
        self.world_size = self.process_group.size()
        self.reshard_after_forward = reshard_after_forward
        self.mixed_precision = mixed_precision
        self.fp32_reduce_scatter = fp32_reduce_scatter
        self.flatten_parameters = flatten_parameters
        self.cpu_offload = cpu_offload
        self.compute_dtype = compute_dtype or (torch.float16 if mixed_precision else torch.float32)
        self.buffer_dtype = buffer_dtype or self.compute_dtype
        self.move_grads_to_cpu = cpu_offload if move_grads_to_cpu is None else move_grads_to_cpu
        self.bucket_cap_mb = bucket_cap_mb

        if self.fp32_reduce_scatter and not self.mixed_precision:
            raise ValueError("fp32_reduce_scatter requires mixed_precision=True")
        if self.cpu_offload and not self.mixed_precision:
            raise ValueError("cpu_offload requires mixed_precision=True")

        compute_device = torch.device("cuda") if self.cpu_offload else next(module.parameters()).device
        validate_process_group(compute_device, self.process_group)

        # Only handle params which are not already sharded. This enables
        # sharding individual layers of a Module, with an outer wrapper to
        # shard any leftover parameters.
        params = list(p for p in module.parameters() if not hasattr(p, "_is_sharded"))

        self._has_params = len(params) > 0
        if not self._has_params:
            self.flatten_parameters = False

        if self.flatten_parameters:
            self._fsdp_wrapped_module: nn.Module = FlattenParamsWrapper(module, param_list=params)
            del module  # free original module in case it helps garbage collection
            self.params = [self._fsdp_wrapped_module.flat_param]
        else:
            self._fsdp_wrapped_module = module
            self.params = params

        # Shard module parameters in place
        self._shard_parameters_()

        # Make sure all parameters are sharded.
        for n, p in self.named_parameters():
            assert hasattr(p, "_is_sharded"), f"found unsharded parameter: {n} ; {p.size()}"

        self._reset_lazy_init()

        # Flag to indicate if we require gradient reduction in the backward
        # pass. This will be False when inside the no_sync context manager.
        self._require_backward_grad_sync: bool = True

        # Enum to indicate if we're in the forward/backward pass, idle, etc.
        self.training_state = TrainingState.IDLE

        # Flag to indicate if the full params are gathered.
        self.has_full_params: bool = False

        # Register hook after state_dict() to remove the "_fsdp_wrapped_module."
        # prefix and before load_state_dict() to add it back.
        self._register_state_dict_hook(_post_state_dict_hook)
        self._register_load_state_dict_pre_hook(_pre_load_state_dict_hook)

        # Flag to indicate whether state_dict() should automatically summon the
        # full params. This defaults to True, but may be set to False if the
        # user explicitly requests the local state dict via local_state_dict().
        self._return_full_state_dict = True

    @property
    def module(self) -> nn.Module:
        return self._fsdp_wrapped_module  # note: may be a FlattenParamsWrapper instance

    @torch.no_grad()
    def _all_buffers_to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
        """Move all buffers to the specified device and dtype, recursively."""
        cast_fn = functools.partial(cast_buffers_, device=device, dtype=dtype)
        self.apply(cast_fn)

    @property
    def params_with_grad(self) -> List[Parameter]:
        """[p for p in self.parameters() if p.grad is not None] """
        return [p for p in self.parameters() if p.grad is not None]

    @torch.no_grad()
    def clip_grad_norm_(
        self,
        max_norm: Union[float, int],
        norm_type: Union[float, int] = 2.0,
        # filter_params_fn: Callable[[Any], Any] = None,
    ) -> torch.Tensor:
        """
        Clip all gradients at this point in time. The norm is computed over all
        gradients together, as if they were concatenated into a single vector.
        Gradients are modified in-place.

        Args:
            max_norm (float or int): max norm of the gradients
            norm_type (float or int): type of the used p-norm. Can be ``'inf'``
                for infinity norm.

        Returns:
            Total norm of the parameters (viewed as a single vector).

        .. note:: This is analogous to `torch.nn.utils.clip_grad_norm_` but
            handles the partitioning and multiple devices per rank under the
            hood. The default torch util is not applicable here, because each
            rank only has a partial view of all the grads in the model, so
            calling it in the OSS context would lead to different scaling being
            applied per subset of model parameters.

        .. warning:: This needs to be called on all ranks, since synchronization
            primitives will be used.
        """
        # We don't call torch.cuda.synchronize() here, since clipping can be
        # inside the train loop and we probably don't want to force a GPU-CPU sync.
        # _lazy_init should be sufficient, since it will force the other streams
        # to sync with the default stream (via _wait_for_previous_optim_step).
        self._lazy_init()
        assert self._is_root, "clip_grad_norm should only be called on the root (parent) instance"
        self.assert_state(TrainingState.IDLE)

        max_norm = float(max_norm)
        norm_type = float(norm_type)
        params_with_grad = self.params_with_grad
        if not self.children_share_process_group:
            raise NotImplementedError(
                "clip_grad_norm requires that all params share one process group. clip_grad_by_value_ should work"
            )
        # Computes the max norm for this shard's gradients and sync's across workers
        local_norm = calc_grad_norm(params_with_grad, norm_type).cuda()
        if norm_type == inf:
            total_norm = local_norm
            dist.all_reduce(total_norm, op=torch.distributed.ReduceOp.MAX, group=self.process_group)
        else:
            total_norm = local_norm ** norm_type
            dist.all_reduce(total_norm, group=self.process_group)
            total_norm = total_norm ** (1.0 / norm_type)

        if self.move_grads_to_cpu:
            total_norm = total_norm.cpu()
        # Now multiply each grad by (max_norm/total_norm), same as torch 1.7 https://tinyurl.com/3wtxhhqq)
        clip_coef = torch.tensor(max_norm, dtype=total_norm.dtype, device=total_norm.device) / (total_norm + 1e-6)
        if clip_coef < 1:

            # multiply by clip_coef
            for p in params_with_grad:
                p.grad.detach().mul_(clip_coef.to(p.grad.device))  # type: ignore

        return total_norm

    @torch.no_grad()
    def _shard_parameters_(self) -> None:
        """
        At initialization we wrap a module with full parameters and shard the
        parameters in-place. Sharding is implemented by viewing each parameter
        as a 1D Tensor and retaining only a single slice, where the slice size
        is determined by the number of data parallel workers.

        Wrapping modules with many small parameters (or with a very large data
        parallel world size) will result in many small parameter shards and slow
        performance. In this case it's better to set *``flatten_parameters``* to
        ``True``, so that all of the small parameters in the module are combined
        into a single contiguous Tensor and sharded once.

        After this initial sharding is complete, the user can initialize a
        ``torch.optim.Optimizer`` in the usual way, i.e.::

        .. code-block:: python

            optim = torch.optim.Adam(sharded_module.parameters(), lr=0.0001)

        The optimizer will see only a single slice of parameters and will thus
        allocate less memory for optimizer state, avoiding redundancy across
        data parallel workers.
        """
        for p in self.params:
            assert not hasattr(p, "_is_sharded")
            assert p.is_floating_point()
            if self.mixed_precision:
                assert p.dtype == torch.float32

            # If world_size is 1, then we all-reduce grads instead of sharding.
            p._is_sharded = self.world_size > 1
            p._orig_size = p.data.size()

            if not p._is_sharded:
                continue
            p._is_sharded = True

            # Replace p.data with the relevant shard.
            orig_data = p.data
            p.data = self._get_shard(p.data)
            free_storage_(orig_data)

    def _get_shard(self, tensor: torch.Tensor) -> torch.Tensor:
        """Return the local shard of a given full tensor."""
        # Shard using torch.chunk to match all-gather/reduce-scatter.
        chunks = list(torch.flatten(tensor).chunk(self.world_size))
        while len(chunks) < self.world_size:
            chunks.append(chunks[0].new_empty(0))

        # Determine number of padding elements.
        num_to_pad = chunks[0].numel() - chunks[self.rank].numel()
        assert num_to_pad >= 0, num_to_pad

        shard = chunks[self.rank].clone()
        if num_to_pad > 0:
            shard = F.pad(shard, [0, num_to_pad])
        return shard

    def extra_repr(self) -> str:
        return (
            f"rank={self.rank}, world_size={self.world_size}, "
            f"reshard_after_forward={self.reshard_after_forward}, "
            f"mixed_precision={self.mixed_precision}, "
            f"fp32_reduce_scatter={self.fp32_reduce_scatter}, "
            f"flatten_parameters={self.flatten_parameters}, "
            f"cpu_offload={self.cpu_offload}, "
            f"compute_dtype={self.compute_dtype}, "
            f"move_grads_to_cpu={self.move_grads_to_cpu}"
        )

    def __getattr__(self, name: str) -> Any:
        """Forward missing attributes to wrapped module."""
        try:
            return super().__getattr__(name)  # defer to nn.Module's logic
        except AttributeError:
            return getattr(self.module, name)

    def __getstate__(self) -> Dict[str, str]:
        """Serialize the state of the current FullyShardedDataParallel instance.

        Some properties are not serializable (e.g., process groups, streams), so
        we remove them and try to reconstruct them in :func:`__setstate__`.
        """
        state = copy.copy(self.__dict__)
        state["is_sharded"] = [p._is_sharded for p in self.params]
        state["orig_sizes"] = [p._orig_size for p in self.params]
        if state["process_group"] is not None:
            state["process_group"] = "MISSING"  # process_group isn't pickleable
        self._reset_lazy_init()
        return state

    def __setstate__(self, state: Dict[str, Any]) -> None:
        """Intercept state setting and perform needed changes on params."""
        super().__setstate__(state)

        def fixup(p: Parameter, is_sharded: bool, size: torch.Size) -> Parameter:
            assert isinstance(p, Parameter)
            p.data = p.data.clone()  # move tensors out of shared memory
            p._is_sharded = is_sharded
            p._orig_size = size
            return p

        self.params = [
            fixup(p, is_sharded, size) for p, is_sharded, size in zip(self.params, self.is_sharded, self.orig_sizes)
        ]
        del self.is_sharded
        del self.orig_sizes
        self._reset_lazy_init()

    # TODO (Min): figuring out how to do typing for this overloaded function.
    def state_dict(self, *args: Any, **kwargs: Any) -> "OrderedDict[str, torch.Tensor]":  # type: ignore
        """
        Returns the whole (unsharded) state of the module. Parameters are not
        sharded, so the resulting state_dict can be loaded directly by the
        wrapped Module without any sharding-specific logic. Returned tensors
        will be full precision (e.g., FP32).

        .. warning:: This needs to be called on all ranks, since synchronization
            primitives will be used.
        """
        torch.cuda.synchronize()
        self._lazy_init()
        if self.mixed_precision:
            # Buffers dtype stays consistent with parameters.
            self._all_buffers_to(dtype=torch.float32)

        if self._return_full_state_dict:
            if self.training_state != TrainingState.SUMMON_FULL_PARAMS:
                with self.summon_full_params(volatile=True):
                    state_dict = super().state_dict(*args, **kwargs)
            else:
                state_dict = super().state_dict(*args, **kwargs)
        else:
            if self.flatten_parameters:
                assert isinstance(self.module, FlattenParamsWrapper)
                state_dict = self.module.flat_state_dict(*args, **kwargs)
            else:
                state_dict = super().state_dict(*args, **kwargs)

        if self.cpu_offload:
            for k in state_dict.keys():
                state_dict[k] = state_dict[k].cpu()

        if self.mixed_precision:
            # In case we are in mixed precision, restore buffers back to fp16.
            self._all_buffers_to(dtype=self.buffer_dtype)
        return state_dict

    # TODO (Min): figuring out how to do typing for this overloaded function.
    def local_state_dict(self, *args, **kwargs):  # type: ignore
        """
        Returns the local (sharded) state of the module. Parameters are sharded,
        so the resulting state_dict can only be loaded after the Module has been
        wrapped with FullyShardedDataParallel.
        """
        with contextlib.ExitStack() as stack:
            # Tell any nested FSDP instances not to auto summon full params.
            for module in self.modules():  # includes self
                if isinstance(module, FullyShardedDataParallel):
                    stack.enter_context(module._no_return_full_state_dict())
            return self.state_dict(*args, **kwargs)

    @contextlib.contextmanager
    def _no_return_full_state_dict(self) -> Generator:
        backup = self._return_full_state_dict
        self._return_full_state_dict = False
        try:
            yield
        finally:
            self._return_full_state_dict = backup

    def load_state_dict(
        self, state_dict: Union[Dict[str, torch.Tensor], "OrderedDict[str, torch.Tensor]"], strict: bool = True
    ) -> NamedTuple:
        """
        Load a whole (unsharded) state_dict.

        .. warning:: This needs to be called on all ranks, since synchronization
            primitives will be used.
        """
        if self._return_full_state_dict:
            with self.summon_full_params():
                return self.module.load_state_dict(state_dict, strict)
        else:
            torch.cuda.synchronize()
            self._lazy_init()
            return self.module.load_state_dict(state_dict, strict)

    def load_local_state_dict(
        self, state_dict: Union[Dict[str, torch.Tensor], "OrderedDict[str, torch.Tensor]"], strict: bool = True
    ) -> NamedTuple:
        """Load a local (sharded) state_dict."""
        with contextlib.ExitStack() as stack:
            # Tell any nested FSDP instances not to auto summon full params.
            for module in self.modules():  # includes self
                if isinstance(module, FullyShardedDataParallel):
                    stack.enter_context(module._no_return_full_state_dict())
            output = self.load_state_dict(state_dict, strict)
        return output

    @contextlib.contextmanager
    def no_sync(self) -> Generator:
        """
        A context manager to disable gradient synchronizations across DDP
        processes. Within this context, gradients will be accumulated on module
        variables, which will later be synchronized in the first
        forward-backward pass after exiting the context.

        .. note:: This may result in higher memory usage because we will
            accumulate the full model gradients (instead of gradient shards)
            until the eventual sync.
        """
        self._lazy_init()
        assert self._is_root, "no_sync on inner FSDP is not supported"
        self.assert_state(TrainingState.IDLE)
        # This instance may wrap other FullyShardedDataParallel instances and we
        # need to set all of them to accumulate gradients.
        old_flags = []
        for m in self.modules():  # includes self
            if isinstance(m, FullyShardedDataParallel):
                old_flags.append((m, m._require_backward_grad_sync))
                m._require_backward_grad_sync = False
        try:
            yield
        finally:
            for m, old_flag in old_flags:
                m._require_backward_grad_sync = old_flag

    @contextlib.contextmanager
    def summon_full_params(self, recurse: bool = True, volatile: bool = False) -> Generator:
        """
        A context manager to expose full params for the current FSDP instance.
        Can be useful *after* forward/backward for a model to get the params for
        additional processing or checking. Parameters will be gathered in full
        precision (e.g., FP32).

        .. note:: This can be used on inner FSDPs.

        .. note:: This can *not* be used within a forward or backward pass. Nor
            can forward and backward be started from within this context.

        .. note:: The full parameters will be freed after the context manager
            exits; it is up to the caller to clone them if needed.

        .. note:: The full parameters can be modified, but only the portion
            corresponding to the local param shard will persist after the
            context manager exits (unless ``volatile=True``, in which case there
            are no guarantees about persistence).

        Args:
            recurse (bool, Optional): recursively summon all params for nested
                FSDP instances (default: True)
            volatile (bool, Optional): if ``True``, modifications to params are
                not guaranteed persist after the context manager exists;
                enabling this can be slightly more efficient (default: False)
        """
        if recurse:
            with contextlib.ExitStack() as stack:
                # Summon all params for any nested FSDP instances.
                for module in self.modules():
                    if isinstance(module, FullyShardedDataParallel):
                        stack.enter_context(module.summon_full_params(recurse=False, volatile=volatile))
                # Yield to the caller, with full params in all nested instances.
                yield
            # Exiting from the ExitStack will re-shard params.
            return
        else:
            torch.cuda.synchronize()
            self._lazy_init()
            self.assert_state(TrainingState.IDLE)
            # Set the state so that we assert when trying to go into
            # forward/backward.
            self.training_state = TrainingState.SUMMON_FULL_PARAMS
            full_tensors = self._rebuild_full_params(full_precision=True)
            assert full_tensors is not None
            with contextlib.ExitStack() as stack:
                if self.flatten_parameters and self.module.is_flattened:
                    # Update flattened views to point to fully-sized tensors. We
                    # use self.params[0] instead of full_tensors since the
                    # latter may contain padding.
                    assert len(self.params) == 1
                    assert isinstance(self.module, FlattenParamsWrapper)
                    stack.enter_context(self.module.unflatten_params(recurse=False, flat_param=self.params[0]))
                try:
                    yield
                finally:
                    stack.close()
                    assert len(full_tensors) == len(self.params)
                    for p, (full_tensor, safe_to_free) in zip(self.params, full_tensors):
                        if not volatile:
                            # Copy any changes made to the full params back into
                            # the corresponding local shards.
                            local_shard = self._get_shard(full_tensor)
                            p._fp32_shard.copy_(local_shard.view_as(p._fp32_shard))
                        if safe_to_free:
                            free_storage_(full_tensor)
                    self.has_full_params = False
                    self._use_fp32_param_shard()
                    self.training_state = TrainingState.IDLE

    def _reset_lazy_init(self) -> None:
        """Reset instance so :func:`_lazy_init` will run on the next forward."""
        self._is_root: Optional[bool] = None
        self._queue_wait_for_post_backward_closure: Optional[Callable] = None
        self._streams: Dict[str, torch.cuda.Stream] = {}
        self._reducer: Optional[ReduceScatterBucketer] = None

    def _lazy_init(self) -> None:
        """Initialization steps that should happen lazily, typically right
           before the first forward pass.
        """
        # Initialize param attributes lazily, in case the param's dtype or
        # device changes after __init__.
        for p in self.params:
            self._init_param_attributes(p)

        # Initialize _is_root and setup streams. These steps would ideally
        # happen in __init__, but _is_root can only be determined after the
        # entire model hierarchy is setup, thus we run it lazily.
        if self._is_root is None:
            self._set_is_root()
            self._setup_streams()

        if self.cpu_offload:  # Buffers stay on GPU, and don't get sharded
            self._all_buffers_to(device=torch.device("cuda"), dtype=self.buffer_dtype)
        else:
            self._all_buffers_to(dtype=self.buffer_dtype)

        if self._is_root:
            # Don't free the full params for the outer-most (root) instance,
            # since those params will be needed immediately after for the
            # backward pass.
            self.reshard_after_forward = False

            # Due to the use of streams, we need to make sure the previous
            # ``optim.step()`` is done before we all-gather parameters.
            self._wait_for_previous_optim_step()

    @torch.no_grad()
    def _init_param_attributes(self, p: Parameter) -> None:
        """
        We manage several attributes on each Parameter instance. The first two
        are set by :func:`_shard_parameters_`:

            ``_is_sharded``: ``True`` if the Parameter is sharded or ``False``
                if the Parameter is intentionally not sharded (in which case we
                will all-reduce grads for this param).
            ``_orig_size``: the size of the original Parameter (before sharding)

        The remaining attributes are set here:
            ``_fp32_shard``: a single shard of the parameters in full precision
                (typically FP32, but this is dependent on the dtype of the model
                as it's passed in by the user). This can be on CPU or GPU
                depending on the value of *``cpu_offload``*.
            ``_fp16_shard``: if *``mixed_precision``* is ``True``, this will be
                a single shard of the parameters in FP16, used for all-gather.
            ``_full_param_padded``: the full weight (padded to be evenly
                divisible by ``world_size``), used for computation in the
                forward and backward pass. This will be resized in place and
                only materialized (via all-gather) as needed.
        """
        assert hasattr(p, "_is_sharded") and hasattr(p, "_orig_size")
        if hasattr(p, "_fp32_shard"):
            return

        # Compute device defaults to CUDA when *cpu_offload* is enabled, or the
        # param's current device otherwise (could be CPU).
        compute_device = torch.device("cuda") if self.cpu_offload else p.device

        # A single shard of the parameters in full precision.
        p._fp32_shard = p.data

        if self.mixed_precision:
            assert p._fp32_shard.dtype == torch.float32

            if self.cpu_offload:
                assert p._fp32_shard.device == torch.device("cpu")
                # If we plan to keep the FP32 parameters on CPU, then pinning
                # memory allows us to later use non-blocking transfers when moving
                # the FP32 param shard to compute_device.
                p._fp32_shard = p._fp32_shard.pin_memory()
                p.data = p._fp32_shard

            # In mixed precision mode, we maintain a reduced precision
            # (typically FP16) parameter shard on compute_device for performing
            # the computation in the forward/backward pass. We resize the
            # storage to size 0 at init (here) and re-materialize (by copying
            # from _fp32_shard) as needed.
            p._fp16_shard = torch.zeros_like(p._fp32_shard, device=compute_device, dtype=self.compute_dtype)
            free_storage_(p._fp16_shard)
        else:
            p._fp16_shard = None  # use _fp32_shard

        # We also maintain a full-sized parameter of type self.compute_dtype
        # (FP16 for mixed_precision or FP32 otherwise). We resize the
        # storage to size 0 at init (here) and only materialize as needed. The
        # storage may contain padding elements so that it is evenly divisible by
        # world_size, although these padding elements will be removed before the
        # relevant computation.
        if p._is_sharded:
            p._full_param_padded = torch.zeros(
                p.data.numel() * self.world_size, device=compute_device, dtype=self.compute_dtype
            )
            free_storage_(p._full_param_padded)

        if self.move_grads_to_cpu:
            # We can optionally move the grad shard to CPU during the backward
            # pass. In this case, it's important to pre-allocate the CPU grad
            # shard in pinned memory so that we can do a non-blocking transfer.
            p._cpu_grad = torch.zeros_like(p.data, device="cpu").pin_memory()

    def _set_is_root(self) -> None:
        """If ``True``, implies that no other :class:`FullyShardedDataParallel`
        instance wraps this one. Called once by :func:`_lazy_init`.
        Also sets self.children_share_process_group = True if all child
        instances share the same process group. If some child instances use a
        different process group, self.clip_grad_norm_ will raise an error.
        """
        if self._is_root is not None:
            return
        # No FullyShardedDataParallel instance wraps this, else _is_root would be set to False.
        self._is_root = True
        assert self._queue_wait_for_post_backward_closure is None
        self._queue_wait_for_post_backward_closure = self._queue_wait_for_post_backward
        # As the root, we now set all children instances to False and
        # give them a closure to try to queue a wait_for_post_backward.
        self.children_share_process_group = True
        for n, m in self.named_modules():
            # `n != ""` excludes self.
            if n != "" and isinstance(m, FullyShardedDataParallel):
                assert m._is_root is None
                m._is_root = False
                # When root instance doesn't have params, allow children instances
                # to queue the post_backward hook.
                #
                # TODO (Min): we should think if we can have a empty param at the root
                #             so that root always have a callback on the backward graph.
                if not self._has_params:
                    assert m._queue_wait_for_post_backward_closure is None
                    m._queue_wait_for_post_backward_closure = self._queue_wait_for_post_backward
                if m.process_group != self.process_group:
                    self.children_share_process_group = False

    def _setup_streams(self) -> None:
        """Create streams to overlap data transfer and computation."""
        if len(self._streams) > 0 or not self._is_root:
            return
        # Stream to move main FP32 params (may be on CPU) to FP16 for forward.
        self._streams["fp32_to_fp16"] = torch.cuda.Stream()
        # Stream for all-gathering parameters.
        self._streams["all_gather"] = torch.cuda.Stream()
        # Stream for overlapping grad reduction with the backward pass.
        self._streams["post_backward"] = torch.cuda.Stream()
        # Helper for bucketing reduce-scatter ops. This is also shared with
        # children instances to improve bucket utilization.
        self._reducer = ReduceScatterBucketer(self.bucket_cap_mb)
        # We share streams with all children instances, which allows them to
        # overlap transfers across the forward pass without synchronizing with
        # the default stream.
        for n, m in self.named_modules():
            if n != "" and isinstance(m, FullyShardedDataParallel):
                m._streams = self._streams
                m._reducer = self._reducer

    def _wait_for_previous_optim_step(self) -> None:
        """
        The outer-most :class:`FullyShardedDataParallel` instance (i.e., the root
        instance) needs to synchronize with the default stream to ensure the
        previous optimizer step is done.
        """
        if self.mixed_precision:
            self._streams["fp32_to_fp16"].wait_stream(torch.cuda.current_stream())
        else:
            self._streams["all_gather"].wait_stream(torch.cuda.current_stream())

    def forward(self, *args: Any, **kwargs: Any) -> torch.Tensor:
        self._lazy_init()

        # Start of a forward pass.
        self.training_state = TrainingState.FORWARD

        if self._is_root and self.mixed_precision:
            args, kwargs = cast_inputs_to_fp16(*args, **kwargs)

        # All-gather full parameters. This will also transfer FP32 parameters to
        # ``self.compute_dtype`` (e.g., FP16 if *mixed_precision* is ``True``).
        self._rebuild_full_params()

        # Register backward hooks to reshard params and reduce-scatter grads.
        # These need to be re-registered every forward pass.
        self._register_post_backward_hooks()

        outputs = self.module(*args, **kwargs)

        if self.reshard_after_forward:
            self._free_full_params()

        # Switch to main FP32 param shard. We maintain this invariant throughout
        # the code, i.e., ``p.data == p._fp32_shard`` after each function. This
        # also ensures that after the first forward, the optimizer state will be
        # initialized with the correct dtype and (sharded) size, since optimizer
        # state is typically initialized lazily in ``optim.step()``.
        self._use_fp32_param_shard()

        # Register pre-backward hooks to all-gather the params for the backward
        # pass (if needed).
        outputs = self._register_pre_backward_hooks(outputs)

        # Done with a forward pass.
        self.training_state = TrainingState.IDLE

        return outputs

    def _register_pre_backward_hooks(self, outputs: Any) -> Any:
        """Register pre-backward hook to run before the wrapped module's
        backward. Hooks should be attached to all outputs from the forward."""
        if not torch.is_grad_enabled():
            return outputs  # don't register hooks if grad isn't enabled

        pre_backward_hook_has_run = [False]

        def _pre_backward_hook(*unused: Any) -> None:
            if pre_backward_hook_has_run[0]:
                return  # only run once
            pre_backward_hook_has_run[0] = True

            # Start of a backward pass.
            self.training_state = TrainingState.BACKWARD

            # All-gather full parameters.
            if self.reshard_after_forward:
                self._rebuild_full_params()
            else:
                self._use_full_params()

            # Make sure p.grad has the correct size/device (or set it to None).
            self._prep_grads_for_backward()

        def _register_hook(t: torch.Tensor) -> torch.Tensor:
            if t.requires_grad:
                t.register_hook(_pre_backward_hook)
            return t

        # Attach hooks to Tensor outputs.
        outputs = apply_to_tensors(_register_hook, outputs)

        return outputs

    def _register_post_backward_hooks(self) -> None:
        """Register backward hooks to reshard params and reduce-scatter grads."""
        if not torch.is_grad_enabled():
            return  # don't register grad hooks if grad isn't enabled
        if self._is_root:
            # This actually means that only root instance has this field
            # defined. Accidentally accessing this field will assert on all
            # other instances, giving us a nice bug checker.
            self._post_backward_callback_queued = False
        for p in self.params:
            if p.requires_grad:
                if hasattr(p, "_shard_bwd_hook"):
                    p._shard_bwd_hook[1].remove()  # remove existing handle
                p_tmp = p.expand_as(p)
                grad_acc = p_tmp.grad_fn.next_functions[0][0]
                handle = grad_acc.register_hook(functools.partial(self._post_backward_hook, p))
                p._shard_bwd_hook = (grad_acc, handle)

    @torch.no_grad()
    def _post_backward_hook(self, param: Parameter, *unused: Any) -> None:
        """
        At the start of :func:`_post_backward_hook`, ``param.grad`` contains the
        full gradient for the local batch. The reduce-scatter op will replace
        ``param.grad`` with a single shard of the summed gradient across all
        GPUs. This shard will align with the current GPU rank. For example::

            before reduce_scatter:
                param.grad (GPU #0): [1, 2, 3, 4]
                param.grad (GPU #1): [5, 6, 7, 8]

            after reduce_scatter:
                param.grad (GPU #0): [6, 8]    # 1+5, 2+6
                param.grad (GPU #1): [10, 12]  # 3+7, 4+8

        The local GPU's ``optim.step`` is responsible for updating a single
        shard of params, also corresponding to the current GPU's rank. This
        alignment is created by :func:`_shard_parameters_`, which ensures that
        the local optimizer only sees the relevant parameter shard.
        """
        self.assert_state(TrainingState.BACKWARD)
        if param.grad is None:
            return
        if param.grad.requires_grad:
            raise RuntimeError("FullyShardedDataParallel only works with gradients that don't require grad")

        if not self._is_root or self._require_backward_grad_sync:
            # Free full params. As a special case, we don't free the full params
            # on the root instance when in a ``no_sync`` context (as indicated
            # by ``self._require_backward_grad_sync``), since we will need the
            # params again immediately for the next forward.
            self._free_full_params([param])

        if self.mixed_precision:
            # This is a no-op if reshard_after_forward is True, since we already
            # free the param shard when rebuilding the full params in the
            # pre_backward_hook.
            self._free_fp16_param_shard([param])

        # Switch to FP32 shard after backward.
        self._use_fp32_param_shard([param])

        # (try to) Enqueue a callback at the end of the backward pass to ensure that all
        # post-backward work has finished. We only need one callback and all instances
        # of FSDP (root and children) make this attempt here to queue to ensure it is queued
        # no matter which instance(s) has(have) params.
        assert self._queue_wait_for_post_backward_closure is not None or not self._is_root
        if self._queue_wait_for_post_backward_closure is not None:
            self._queue_wait_for_post_backward_closure()

        if not self._require_backward_grad_sync:
            return

        # Wait for all work in the current stream to finish, then start the
        # reductions in post_backward stream.
        self._streams["post_backward"].wait_stream(torch.cuda.current_stream())
        with torch.cuda.stream(self._streams["post_backward"]):
            orig_grad_data = param.grad.data

            if self.mixed_precision and self.fp32_reduce_scatter:
                # Cast grad to FP32.
                param.grad.data = param.grad.data.to(param.dtype)

            if self.world_size > 1:
                # Average grad by world_size for consistency with PyTorch DDP.
                param.grad.data.div_(self.world_size)

            callback_fn = functools.partial(self._post_reduction_hook, param)
            if param._is_sharded:
                assert param._is_sharded
                assert self._reducer is not None
                grad_chunks = chunk_and_pad(param.grad.data, self.world_size)
                self._reducer.reduce_scatter_async(grad_chunks, group=self.process_group, callback_fn=callback_fn)
            else:
                # Currently the only way for _is_sharded to be False is if
                # world_size == 1. This could be relaxed in the future, in which
                # case grads should be all-reduced here.
                assert self.world_size == 1
                callback_fn(param.grad.data)

            # After _post_backward_hook returns, orig_grad_data will eventually
            # go out of scope, at which point it could otherwise be freed for
            # further reuse by the main stream while the div/reduce_scatter/copy
            # are underway in the post_backward stream. See:
            # github.com/NVIDIA/apex/blob/master/apex/parallel/distributed.py
            orig_grad_data.record_stream(self._streams["post_backward"])

    def _post_reduction_hook(self, param: Parameter, reduced_grad: torch.Tensor) -> None:
        """Hook to call on each param after the reduce-scatter."""
        assert torch.cuda.current_stream() == self._streams["post_backward"]
        assert param.grad is not None
        self.assert_state(TrainingState.BACKWARD)
        param.grad.data = reduced_grad
        # Cast grad to param's dtype (typically FP32). Note: we do this
        # before the move_grads_to_cpu step so that this entire hook remains
        # non-blocking. The downside is a bit more D2H transfer in that case.
        if self.mixed_precision:
            param.grad.data = param.grad.data.to(dtype=param.data.dtype)
        # Optionally move gradients to CPU, typically used if one is running
        # the optimizer on the CPU.
        if self.move_grads_to_cpu:
            param._cpu_grad.copy_(param.grad.data, non_blocking=True)
            param.grad.data = param._cpu_grad
        # Don't let this memory get reused until after the transfers.
        reduced_grad.record_stream(torch.cuda.current_stream())

    def _queue_wait_for_post_backward(self) -> None:
        """Try to queue a `wait_for_post_backward` callback.

        Only called on root and only queue one callback. But can be called by
        children FSDPs via a closure in case the root instance doesn't own any
        params.
        """
        assert self._is_root
        self.assert_state(TrainingState.BACKWARD)
        if not self._post_backward_callback_queued:
            self._post_backward_callback_queued = True
            Variable._execution_engine.queue_callback(self._wait_for_post_backward)

    @torch.no_grad()
    def _wait_for_post_backward(self) -> None:
        """Wait for post-backward to finish. Only called on root instance."""
        assert self._is_root
        self.assert_state(TrainingState.BACKWARD)
        if self._require_backward_grad_sync:
            # Flush any unreduced buckets in the post_backward stream.
            with torch.cuda.stream(self._streams["post_backward"]):
                assert self._reducer is not None
                self._reducer.flush()
            torch.cuda.current_stream().wait_stream(self._streams["post_backward"])
            if self.move_grads_to_cpu:
                # Wait for the non-blocking GPU -> CPU grad transfers to finish.
                torch.cuda.current_stream().synchronize()
        # A backward pass is done, update root and nested FSDP's flags.
        for m in self.modules():  # includes self
            if isinstance(m, FullyShardedDataParallel):
                m.assert_state(TrainingState.BACKWARD)
                m.training_state = TrainingState.IDLE

    @torch.no_grad()
    def _rebuild_full_params(self, full_precision: bool = False) -> Optional[List[Tuple[torch.Tensor, bool]]]:
        """
        Gather all shards of params.

        Args:
            full_precision (bool, Optional): by default params will be gathered
                in ``compute_dtype`` (e.g., FP16), unless *full_precision* is
                ``True``, in which case they will be gathered in full precision
                (e.g., FP32), possibly in fresh storage.

        Returns:
            A list of tuples, where the first element is the full-sized param
            and the second element is a bool indicating if it's safe for the
            caller to free the full-sized param. This will be ``None`` if
            ``full_precision=False`` and the full params are already gathered.
        """
        output_tensors: List[Tuple[torch.Tensor, bool]] = []

        def update_p_data(custom_output_tensor: Optional[torch.Tensor] = None) -> None:
            if custom_output_tensor is not None:
                assert p._is_sharded
                p.data = custom_output_tensor
                output_tensors.append((p.data, True))
            elif not p._is_sharded:
                if self.mixed_precision and not full_precision:
                    p.data = p._fp16_shard
                    output_tensors.append((p.data, True))
                else:
                    # Here p.data == p._fp32_shard, so it's not safe to free.
                    output_tensors.append((p.data, False))
            else:
                p.data = p._full_param_padded
                output_tensors.append((p.data, True))
            # Trim any padding and reshape to match original size.
            p.data = p.data[: p._orig_size.numel()].view(p._orig_size)

        # Early exit if we already have full params and don't need full precision.
        if self.has_full_params and not full_precision:
            for p in self.params:
                update_p_data()
            return output_tensors

        self.has_full_params = True

        with torch.cuda.stream(self._streams["all_gather"]):
            if self.mixed_precision and not full_precision:
                self._cast_fp32_param_shards_to_fp16()

            for p in self.params:
                if not p._is_sharded:  # e.g., when world_size == 1
                    update_p_data()
                else:
                    # If self.cpu_offload and full_precision, we need to cast
                    # the FP32 CPU param to CUDA for the all-gather.
                    p_data = p.data.to(p._full_param_padded.device)

                    p_size = p._full_param_padded.size()
                    assert p_size.numel() % self.world_size == 0
                    if not self.mixed_precision or not full_precision:
                        if p._full_param_padded.storage().size() != p_size.numel():
                            # Allocate based on full size from all shards.
                            alloc_storage_(p._full_param_padded, size=p_size)
                        output_tensor = p._full_param_padded
                    else:
                        # Allocate fresh tensor in full precision.
                        output_tensor = p_data.new_zeros(p_size)

                    # Fill output_tensor with (p.data for each shard in self.world_size)
                    chunks = list(output_tensor.chunk(self.world_size))
                    dist.all_gather(chunks, p_data, group=self.process_group)

                    # Set p.data = output_tensor (with padding trimmed)
                    update_p_data(output_tensor)

                    if self.mixed_precision and not full_precision:
                        self._free_fp16_param_shard([p])
        torch.cuda.current_stream().wait_stream(self._streams["all_gather"])
        return output_tensors

    @torch.no_grad()
    def _use_full_params(self) -> None:
        """
        Switch p.data pointers to use the full params.

        Note: this assumes full params are already gathered.
        """
        assert self.has_full_params
        for p in self.params:
            if not p._is_sharded:
                if self.mixed_precision:
                    assert p._fp16_shard.storage().size() != 0
                    p.data = p._fp16_shard
            else:
                assert p._full_param_padded.storage().size() != 0
                p.data = p._full_param_padded[: p._orig_size.numel()].view(p._orig_size)

    @torch.no_grad()
    def _prep_grads_for_backward(self) -> None:
        """Make sure p.grad has the correct size/device, otherwise set it to None."""
        for p in self.params:
            if p.grad is not None and (p.grad.size() != p._orig_size or p.grad.device != p.data.device):
                p.grad = None

    @torch.no_grad()
    def _free_full_params(self, params: Optional[List[Parameter]] = None) -> None:
        """Free up storage for full parameters."""
        if params is None:
            params = self.params
        self.has_full_params = False
        current_stream = torch.cuda.current_stream()
        with torch.cuda.stream(self._streams["all_gather"]):
            for p in params:
                if not p._is_sharded:  # e.g., world_size == 1
                    if self.mixed_precision:
                        self._free_fp16_param_shard([p])
                    continue
                # There may be external references to the Tensor Storage that we
                # can't modify, such as references that are created by
                # ctx.save_for_backward in the forward pass. Thus when we
                # unshard parameters, we should reuse the original Tensor
                # Storage object and unshard it in-place. For now, just resize
                # the Storage to 0 to save memory.
                p._full_param_padded.record_stream(current_stream)
                free_storage_(p._full_param_padded)

    @torch.no_grad()
    def _use_fp32_param_shard(self, params: Optional[List[Parameter]] = None) -> None:
        """Use FP32 shard for a list of params."""
        if params is None:
            params = self.params
        for p in params:
            p.data = p._fp32_shard

    @torch.no_grad()
    def _cast_fp32_param_shards_to_fp16(self, params: Optional[List[Parameter]] = None) -> None:
        """Cast FP32 param shard to FP16 for a list of params."""
        if params is None:
            params = self.params
        with torch.cuda.stream(self._streams["fp32_to_fp16"]):
            for p in params:
                assert p._fp16_shard is not None
                alloc_storage_(p._fp16_shard, size=p._fp32_shard.size())
                p._fp16_shard.copy_(
                    # If cpu_offload is True, this will be non-blocking because
                    # _fp32_shard is pinned, otherwise it's a no-op.
                    p._fp32_shard.to(p._fp16_shard.device, non_blocking=True)
                )
                p.data = p._fp16_shard
        torch.cuda.current_stream().wait_stream(self._streams["fp32_to_fp16"])

    @torch.no_grad()
    def _free_fp16_param_shard(self, params: Optional[List[Parameter]] = None) -> None:
        """Free storage for FP16 shards for a list of params."""
        if params is None:
            params = self.params
        current_stream = torch.cuda.current_stream()
        for p in params:
            if p._fp16_shard is not None:
                # _fp16_shard is allocated in _fp32_to_fp16_stream, so we can't
                # free it until the work in the current stream completes.
                p._fp16_shard.record_stream(current_stream)
                free_storage_(p._fp16_shard)

    def assert_state(self, state: TrainingState) -> None:
        """Assert we are in the given state."""
        assert (
            self.training_state == state
        ), f"expected to be in state {state} but current state is {self.training_state}"


@torch.no_grad()
def cast_inputs_to_fp16(*args: Any, **kwargs: Any) -> Tuple[Any, Any]:
    """
    Cast any Tensors in *args or **kwargs to FP16.
    """

    def fn(x: torch.Tensor) -> torch.Tensor:
        if x.dtype is torch.float32:
            return x.half()
        return x

    return apply_to_tensors(fn, args), apply_to_tensors(fn, kwargs)


def cast_buffers_(
    module: nn.Module, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None
) -> None:
    """Cast all of module.named_buffers to device and floating point buffers to dtype."""
    # if buffers are already on the right device and/or dtype this is just python loop cost
    assert dtype in {torch.float32, torch.float16}  # assumes compute_dtype == float16
    for key, buf in module.named_buffers(recurse=False):
        if buf is not None:
            buf = buf.to(device=device)
            if torch.is_floating_point(buf):
                buf = buf.to(dtype=dtype)
            setattr(module, key, buf)


def free_storage_(data: torch.Tensor) -> None:
    """Free underlying storage of a Tensor."""
    if data.storage().size() > 0:
        # Since we're modifying the Tensor's Storage directly, make sure the Tensor
        # is the sole occupant of the Storage.
        assert data.storage_offset() == 0
        data.storage().resize_(0)


@torch.no_grad()
def alloc_storage_(data: torch.Tensor, size: torch.Size) -> None:
    """Allocate storage for a tensor."""
    if data.storage().size() == size.numel():  # no need to reallocate
        return
    assert data.storage().size() == 0
    data.storage().resize_(size.numel())


def _post_state_dict_hook(
    module: nn.Module, state_dict: "OrderedDict[str, torch.Tensor]", prefix: str, *args: Any
) -> "OrderedDict[str, torch.Tensor]":
    if module.training_state == TrainingState.SUMMON_FULL_PARAMS:
        # We copy the state_dict since full param will be freed after
        # we exit the summon_full_params() context.
        for key in state_dict.keys():
            state_dict[key] = state_dict[key].clone()

    # Remove "_fsdp_wrapped_module." prefix
    replace_by_prefix_(state_dict, prefix + "_fsdp_wrapped_module.", prefix)
    return state_dict


def _pre_load_state_dict_hook(
    state_dict: Union[Dict[str, torch.Tensor], "OrderedDict[str, torch.Tensor]"], prefix: str, *args: Any
) -> None:
    replace_by_prefix_(state_dict, prefix, prefix + "_fsdp_wrapped_module.")