import contextlib import warnings from enum import auto, Enum from itertools import accumulate, chain from typing import ( Any, Dict, Generator, Iterator, List, NamedTuple, no_type_check, Optional, Sequence, Set, Tuple, Union, ) import torch import torch.distributed as dist import torch.nn as nn import torch.nn.functional as F from torch import Tensor from torch.distributed._tensor import DTensor from torch.distributed.fsdp._common_utils import ( _set_fsdp_flattened, HandleTrainingState, ) from ._fsdp_extensions import _ext_post_unflatten_transform, _ext_pre_flatten_transform from ._utils import ( _alloc_storage, _free_storage, _no_dispatch_record_stream, _same_storage, p_assert, ) __all__ = [ "FlatParameter", "FlatParamHandle", "FlatParamShardMetadata", "ParamInfo", "SharedParamInfo", "HandleShardingStrategy", ] """ [Note: Fully Sharded Module] We define the "fully sharded module" to be the original ``nn.Module`` that owns a ``FlatParamHandle``. It is the *single* module logically responsible for the *single* unshard/reshard pair for the handle's ``FlatParameter`` for a given forward or backward pass. The fully sharded module should be passed to the ``FlatParamHandle`` constructor. For the wrapper code path: - The ``FullyShardedDataParallel`` module wrapping the fully sharded module runs the unshard/reshard on behalf of the fully sharded module by overriding ``nn.Module.forward``. - The fully sharded module is exactly the module passed to the ``FullyShardedDataParallel`` constructor's ``module`` argument. For the non-wrapper code path: - Hooks registered on the fully sharded module run the unshard/reshard. - The fully sharded module may either be the direct argument to ``fully_shard`` or a submodule chosen by the provided wrapping policy. """ class ParamInfo(NamedTuple): """Information for an original module parameter.""" param_name: str # unprefixed module: nn.Module module_name: str class SharedParamInfo(NamedTuple): """ Additional information for a shared parameter. For each shared parameter, we designate one module and its parameter variable to be the primary owner, determined as the first one encountered in the parameter walk. These are prefixed with "prim". The primary module and parameter do not have their own :class:`SharedParamInfo` instance. """ param_name: str # unprefixed module: nn.Module module_name: str prim_param_name: str # unprefixed prim_module: nn.Module prim_module_name: str class FlatParamShardMetadata(NamedTuple): """ This holds metadata specific to this rank's shard of the flattened parameter. Attributes: param_names (Tuple[str, ...]): Prefixed parameter names of this rank's shard of the parameters; see :class:`FlatParameter`. param_shapes (Tuple[torch.Size, ...]): Parameter shapes of this rank's shard of the parameters; see :class:`FlatParameter`. param_numels (Tuple[int, ...]): Parameter numels of this rank's shard of the parameters; see :class:`FlatParameter`. param_offsets (Tuple[Tuple[int, int], ...]): [start, end] offsets (in units of numels) giving this rank's part of each flattened original module parameter. """ param_names: Tuple[str, ...] param_shapes: Tuple[torch.Size, ...] param_numels: Tuple[int, ...] param_offsets: Tuple[Tuple[int, int], ...] # TODO (awgu): Prefix these with "Handle" for now to avoid circular imports and # inadvertent misuses; coalesce with those in fully_sharded_data_parallel.py # later class HandleShardingStrategy(Enum): FULL_SHARD = auto() SHARD_GRAD_OP = auto() NO_SHARD = auto() HYBRID_SHARD = auto() _HYBRID_SHARD_ZERO2 = auto() class FlatParameter(nn.Parameter): """ This is the flattened parameter used by :class:`FullyShardedDataParallel`. It is comprised of one or more original parameters, which are flattened and concatenated to construct the flattened parameter. Under the current design, this parameter logically represents both the unsharded and sharded flattened parameter, and its data changes storages dynamically. - In the :class:`FullyShardedDataParallel` constructor, the parameter is initialized as unsharded and then sharded in-place. - At runtime, the parameter is lazily (re)-initialized. The sharded parameter data is saved in ``self._local_shard``, and a new ``Tensor`` ``self._full_param_padded`` is created, which is the all-gather destination and owns the unsharded parameter storage thereafter. (See :meth:`FlatParamHandle.init_flat_param_attributes`.) - Throughout runtime, the parameter data changes storages as needed, e.g. to the sharded flattened parameter, reduced-precision sharded flattened parameter, or the unsharded flattened parameter. Attributes: _unpadded_unsharded_size (torch.Size): Unsharded flattened parameter's size without padding. _padded_unsharded_size (torch.Size): Unsharded flattened parameter's size with padding. This is only set for sharded strategies since they require padding for the all-gather. _sharded_size (torch.Size): Sharded flattened parameter's size with padding. This is also set for ``NO_SHARD``, in which case it is the same as the unsharded sizes. (We omit "padded" because there is no analogous unpadded one.) _param_infos (Tuple[ParamInfo, ...]): Each parameter's parameter info entry; see :class:`ParamInfo`. _numels (Tuple[int, ...]): Each parameter's numel. _shapes (Tuple[torch.Size, ...]): Each parameter's shape. _fqns (Tuple[str, ...]): The original parameters' FQNs prefixed from the owning handle's ``_fully_sharded_module``. The names are guaranteed to be unique within the subtree rooted at that module. _num_params (int): Number of original parameters flattened into this flattened parameter; this is the length of ``_param_infos``, ``_numels``, ``_shapes``, and ``_fqns``. _shared_param_infos (Tuple[SharedParamInfo, ...]): Shared parameter info entries; see :class:`SharedParamInfo`. _param_extensions (Tuple[Optional[Any], ...]): Parameter extensions (i.e. some per-parameter state) used to customize pre-flatten and post-unflatten behavior. This is experimental, and users should not depend on its existence in the future. _modules (Set[nn.Module]): Modules that contain some original parameter that is flattened into the ``FlatParameter``. _shard_param_offsets (List[Tuple[int, int])): [start, end] offsets (in units of numel) giving this rank's part of each flattened original module parameter; for any parameter ``p`` that is not sharded across ranks, this will be [0, ``p.numel()``-1]. _shard_indices (Tuple[int, int]): [start, end] indices (in units of parameters) for this rank's shard of the original model parameters, where the parameters follow the order in which they were originally flattened; this indexes appropriately into any data structure that follows the flattening order (e.g. ``_param_infos``, ``_numels``, etc.). _shard_numel_padded (int): Numel padded for this rank's sharded flattened parameter. _local_shard (Tensor): Sharded flattened parameter with padding if using a sharded strategy. If using ``NO_SHARD``, then this is the unpadded unsharded flattened parameter, and there is no notion of a sharded flattened parameter or padded unsharded flattened parameter. _full_param_padded (Tensor): Unsharded flattened parameter with padding. This is not defined for ``NO_SHARD``. When using mixed precision for parameters, this has the low precision. _full_prec_full_param_padded (Tensor): Full precision unsharded flattened parameter with padding. This is used for unsharding outside of computation when using mixed precision for parameters. This is never defined for ``NO_SHARD``. _post_backward_hook_state (Tuple[AccumulateGrad, RemovableHandle]): Flattened parameter's :class:`AccumulateGrad` object and post-backward hook handle. _mp_shard (Tensor): Low precision sharded flattened parameter with padding. This is only defined when parameter mixed precision is enabled. For ``NO_SHARD``, this is used for computation. _cpu_grad (Tensor): Sharded gradient with padding stored on CPU. This is only defined when offloading parameters is enabled. _saved_grad_shard (Tensor): Sharded gradient with padding from previous iterations for gradient accumulation without :meth:`no_sync`. _params (Optional[List[nn.Parameter]]): The original parameter variables if ``use_orig_params=True`` and ``None`` otherwise. _shared_params (Optional[List[nn.Parameter]]): The original shared parameter variables if ``use_orig_params=True`` and ``None`` otherwise. _tensors (Optional[List[Optional[Tensor]]]): This saves the ``Tensor`` views created in the forward and tracked by autograd when ``use_orig_params=True`` and is ``None`` otherwise. This is to preserve those ``Tensor`` variables for the backward to ensure that the ``FlatParameter`` 's ``AccumulateGrad`` object does not change in which case the post-backward hook does not run. This is relevant for cases like reentrant activation checkpointing. _is_grad_none (Optional[List[bool]]): A mask over the original parameters' gradients indicating if it is logically ``None`` or not if ``use_orig_params=True`` and ``None`` otherwise. This is needed because only some of the parameters may have ``None`` gradient, in which case the ``FlatParameter`` gradient must be non-``None`` and must use zeros to approximate those original ``None`` gradients. This mask informs FSDP to set the original parameter gradients to ``None`` (instead of zeros) as needed. """ def _init_metadata( self, param_infos: List[ParamInfo], numels: List[int], shapes: List[torch.Size], fqns: List[str], shared_param_infos: List[SharedParamInfo], param_extensions: List[Any], params: Optional[List[nn.Parameter]], shared_params: Optional[List[nn.Parameter]], ) -> None: """ Initializes attributes holding metadata about the original parameters comprising the flattened parameter. We expose this method separate from the constructor to keep the constructor only responsible for the flattened parameter's tensor data. This method should only be called once per model, while the constructor may be called multiple times, e.g. when reloading from a checkpoint, in which case only the tensor data needs to be passed to the constructor. Since :meth:`load_state_dict` is implemented via :meth:`copy_`, the metadata is correctly assumed to be unchanged. Args: See the Attributes in the class docstring. """ assert len(param_infos) == len(numels) assert len(param_infos) == len(shapes) assert len(param_infos) == len(fqns) assert len(param_infos) == len(param_extensions) self._num_params = len(param_infos) self._param_infos = tuple(param_infos) self._numels = tuple(numels) self._shapes = tuple(shapes) self._fqns = tuple(fqns) self._shared_param_infos = tuple(shared_param_infos) self._param_extensions = tuple(param_extensions) self._modules = {pi.module for pi in self._param_infos}.union( {spi.module for spi in self._shared_param_infos} ) assert (params is None) == (shared_params is None) if params is not None: assert shared_params is not None and len(shared_params) == len( shared_param_infos ) self._params: Optional[List[nn.Parameter]] = params self._shared_params: Optional[List[nn.Parameter]] = shared_params # Mark the original parameters to avoid flattening them into # another `FlatParameter` during recursive construction for param in chain(self._params, self._shared_params): _set_fsdp_flattened(param) self._is_grad_none: Optional[List[bool]] = [ False for _ in range(len(params)) ] self._tensors: Optional[List[Optional[Tensor]]] = [ None for _ in range(len(self._params)) ] else: self._params = None self._shared_params = None self._is_grad_none = None self._tensors = None self._unpadded_unsharded_size = self.size() _set_fsdp_flattened(self) # Tracks whether the `FlatParameter`'s post-backward hook has been # called to modify the behavior of the post-backward callback self._post_backward_called = False class FlatParamHandle: """ This handle manages a flattened parameter (:class:`FlatParameter`). This includes sharding and view management. Args: params (Sequence[nn.Parameter]): The parameters to use for the flattened parameter. fully_sharded_module (nn.Module): See [Note: Fully Sharded Module]. device (torch.device): The compute and communication device, which should be a non-CPU device. We refer to it as the compute device. sharding_strategy (ShardingStrategy): Sharding strategy to apply to this handle's ``FlatParameter``. offload_params (bool): Whether to offload the handle's ``FlatParameter`` to CPU. mp_param_dtype (Optional[torch.dtype]): Parameter mixed precision setting passed to the FSDP constructor. mp_reduce_dtype (Optional[torch.dtype]): Gradient reduction mixed precision setting passed to the FSDP constructor. keep_low_precision_grads (bool): Whether to keep gradients in low precision. use_orig_params (bool): If ``True``, then FSDP preserves the original parameter variables and returns them from ``named_parameters()`` (e.g. to support different optimizer hyperparameters within one :class:`FlatParameter`). If ``False``, then FSDP reconstructs the parameter every iteration and returns the :class:`FlatParameter` s from ``named_parameters()``. """ ################## # INITIALIZATION # ################## def __init__( self, params: Sequence[nn.Parameter], fully_sharded_module: nn.Module, device: torch.device, sharding_strategy: HandleShardingStrategy, offload_params: bool, mp_param_dtype: Optional[torch.dtype], mp_reduce_dtype: Optional[torch.dtype], keep_low_precision_grads: bool, process_group: dist.ProcessGroup, use_orig_params: bool, ): super().__init__() self.device = device self.process_group = process_group self.rank = process_group.rank() self.world_size = process_group.size() self._sharding_strategy = sharding_strategy self._offload_params = offload_params self._use_orig_params = use_orig_params self._keep_low_precision_grads = keep_low_precision_grads self._training_state = HandleTrainingState.IDLE self._debug_level = dist.get_debug_level() self._fully_sharded_module = fully_sharded_module self._init_flat_param(params, fully_sharded_module, use_orig_params) self._orig_param_dtype = self.flat_param.dtype self._use_unsharded_views(as_params=False) self._init_param_reduce_dtypes(mp_param_dtype, mp_reduce_dtype) def _init_flat_param( self, params: Sequence[Optional[nn.Parameter]], module: nn.Module, use_orig_params: bool, ) -> None: """ Initializes the flattened parameter ``self.flat_param`` by flattening the parameters in ``params`` into a single :class:`FlatParameter` and saves relevant metadata. Shared parameters are only included in the flattened parameter once. This checks that all comprising parameters have the same dtype and ``requires_grad`` and does not support nested construction of :class:`FlatParameter` s. Args: See the Args in the class docstring. """ params_set = set(params) params_set.discard(None) if len(params_set) == 0: raise ValueError( "Cannot initialize a `FlatParameter` from an empty parameter list" ) param_infos: List[ParamInfo] = [] numels: List[int] = [] shapes: List[torch.Size] = [] fqns: List[str] = [] shared_param_infos: List[SharedParamInfo] = [] shared_param_memo: Dict[nn.Parameter, Tuple[nn.Module, str, str]] = {} params_to_flatten: List[Union[torch.Tensor, nn.Parameter]] = [] shared_params: List[Union[torch.Tensor, nn.Parameter]] = [] param_extensions: List[Any] = [] dtype: Optional[torch.dtype] = None requires_grad: Optional[bool] = None for submodule_name, submodule in module.named_modules(): for param_name, param in submodule.named_parameters(recurse=False): if param not in params_set: continue if param in shared_param_memo: # shared reference prim_module, prim_module_name, prim_param_name = shared_param_memo[ param ] shared_params.append(param) shared_param_infos.append( SharedParamInfo( param_name, submodule, submodule_name, prim_param_name, prim_module, prim_module_name, ) ) else: if type(param) is FlatParameter: raise ValueError("`FlatParameter` does not support nesting") if dtype is not None and param.dtype != dtype: raise ValueError( "`FlatParameter` requires uniform dtype but got " f"{dtype} and {param.dtype}" ) if dtype is None and not param.is_floating_point(): raise ValueError("Integer parameters are unsupported") if ( requires_grad is not None and param.requires_grad != requires_grad ): raise ValueError( "`FlatParameter` requires uniform `requires_grad`" ) param, extension = _ext_pre_flatten_transform(param) param_extensions.append(extension) dtype = param.dtype requires_grad = param.requires_grad shared_param_memo[param] = (submodule, submodule_name, param_name) params_to_flatten.append(param) param_infos.append(ParamInfo(param_name, submodule, submodule_name)) numels.append(param.numel()) shapes.append(param.shape) fqn = ( submodule_name + "." + param_name if submodule_name else param_name ) fqns.append(fqn) assert requires_grad is not None, ( "Passed-in `params` were not found in the module tree\n" f"params: {params}\nmodule: {module}" ) self.flat_param = FlatParamHandle.flatten_params( params_to_flatten, requires_grad ) # For `use_orig_params=True`, ensure that the logical parameters are # `nn.Parameter`s (and not plain `torch.Tensor`) def convert_to_params( tensors: List[Union[torch.Tensor, nn.Parameter]] ) -> List[nn.Parameter]: return [ t if isinstance(t, nn.Parameter) else nn.Parameter(t) for t in tensors ] self.flat_param._init_metadata( param_infos, numels, shapes, fqns, shared_param_infos, param_extensions, convert_to_params(params_to_flatten) if use_orig_params else None, convert_to_params(shared_params) if use_orig_params else None, ) @staticmethod def flatten_params( params: Sequence[torch.Tensor], requires_grad: bool, ) -> FlatParameter: """ Flattens the parameters in ``params`` into a single :class:`FlatParameter`. This should be the only way used to construct :class:`FlatParameter` s. We expose this factory method for checkpointing (e.g. sharded state dict). The flattened parameter's metadata should only be initialized once (see :meth:`_init_metadata`), but its tensor data may be reloaded. """ with torch.no_grad(): flat_params = [ p.detach().reshape(-1) if isinstance(p, nn.Parameter) else p.reshape(-1) for p in params ] flat_param_data = torch.cat(flat_params, dim=0) flat_param = FlatParameter(flat_param_data, requires_grad=requires_grad) return flat_param def _init_param_reduce_dtypes( self, mp_param_dtype: Optional[torch.dtype], mp_reduce_dtype: Optional[torch.dtype], ) -> None: """ Precondition: ``self.flat_param`` is set via :meth:`_init_flat_param`. This ensures that this handle's parameters have a single dtype. Postcondition: This sets ``self._fwd_bwd_param_dtype`` and ``self._reduce_dtype``. If ``mp_param_dtype`` or ``mp_reduce_dtype`` is ``None``, then we assume the original parameter dtype. One special case is if ``mp_param_dtype`` is not ``None`` and ``mp_reduce_dtype`` is ``None``, in which case we assume the gradient reduction dtype matches the forward/backward parameter dtype. """ # Save whether these dtypes were specified so that we permit the # parameter dtype to change up until the lazy initialization self._low_prec_param_dtype_specified = mp_param_dtype is not None self._low_prec_reduce_dtype_specified = mp_reduce_dtype is not None if ( self._low_prec_param_dtype_specified and not self._low_prec_reduce_dtype_specified ): # Special case: infer gradient reduction mixed precision self._fwd_bwd_param_dtype = mp_param_dtype self._reduce_dtype = self._fwd_bwd_param_dtype else: self._fwd_bwd_param_dtype = mp_param_dtype or self._orig_param_dtype self._reduce_dtype = mp_reduce_dtype or self._orig_param_dtype assert self._fwd_bwd_param_dtype is not None assert self._reduce_dtype is not None ################################### # SHARD INITIALIZATION & METADATA # ################################### @torch.no_grad() def shard(self): """ Shards the handle's ``FlatParameter``. In terms of memory, this allocates new memory for the sharded flattened parameter and frees the unsharded flattened parameter's storage. Postcondition: ``self.flat_param`` is the sharded flattened parameter. Shard metadata attributes are set for all sharding strategies. ``process_group``, ``rank``, and ``world_size`` attributes are set if using a sharded strategy. """ flat_param = self.flat_param if not self.uses_sharded_strategy: self._init_shard_metadata(0, 0, flat_param.numel() - 1) else: p_assert( flat_param.storage_offset() == 0, "The `FlatParameter` is not the sole occupant of its storage", ) orig_storage = flat_param._typed_storage() sharded_flat_param, numel_padded = FlatParamHandle._get_shard( flat_param, self.rank, self.world_size ) flat_param.set_(sharded_flat_param) # type: ignore[call-overload] start = sharded_flat_param.numel() * self.rank end = sharded_flat_param.numel() * (self.rank + 1) - 1 # inclusive self._init_shard_metadata(numel_padded, start, end) if orig_storage._size() > 0: orig_storage._resize_(0) if self._use_orig_params: self._use_sharded_views() def _init_shard_metadata( self, numel_padded: int, start: int, end: int, ) -> None: """ Initializes shard-related metadata for this rank's shard of the flattened parameter: ``_sharded_size``, ``_shard_param_offsets``, ``_shard_indices``, and ``_shard_numel_padded``. Args: numel_padded (int): Numel padded for this rank's sharded flattened parameter. start (int): Start index in the sharded flattened parameter assigned to this rank. end (int): End index (inclusive) in the sharded flattened parameter assigned to this rank. If this exceeds the sharded flattened parameter's numel, then it is truncated. Precondition: ``self.flat_param`` 's data is the sharded flattened parameter. """ self.flat_param._sharded_size = self.flat_param.size() # type: ignore[attr-defined] sharded_flat_param_numel = self.flat_param.numel() # includes `numel_padded` p_assert(start >= 0 and start <= end, f"start: {start} end: {end}") p_assert( numel_padded <= sharded_flat_param_numel, f"numel_padded: {numel_padded} " f"sharded_flat_param_numel: {sharded_flat_param_numel}", ) ( self.flat_param._shard_param_offsets, # type: ignore[attr-defined] self.flat_param._shard_indices, # type: ignore[attr-defined] ) = self._get_shard_metadata(start, end) self.flat_param._shard_numel_padded = numel_padded # type: ignore[attr-defined] def _get_shard_metadata( self, start: int, end: int, ) -> Tuple[Tuple[Tuple[int, int], ...], Tuple[int, int]]: """ Computes the shard metadata based on ``start`` and ``end``, which give the closed interval of the unsharded flattened parameter specifying the shard. Args: start (int): Start index (in units of numel) of this rank's shard of the flattened parameter. end (int): End index (in units of numel and inclusive) of this rank's shard of the flattened parameter. Return: Tuple[Tuple[Tuple[int, int], ...], Tuple[int, int]]: See ``_shard_param_offsets`` and ``_shard_indices`` in :class:`FlatParameter` 's docstring. """ flat_param_offsets = self._get_flat_param_offsets() # Indices of the original parameters in this rank's sharded flattened # parameter shard_param_indices_range = [] # elements will be consecutive # [start, end] offsets giving this rank's part of the flattened # original module parameter (which will be [0, `p.numel()`-1] for any # parameter that is not sharded across ranks) shard_param_offsets = [] for i, (param_start, param_end) in enumerate(flat_param_offsets): if start > param_end or end < param_start: continue if start <= param_start: intra_param_start = 0 else: intra_param_start = start - param_start intra_param_end = min(param_end, end) - param_start shard_param_indices_range.append(i) shard_param_offsets.append( (intra_param_start, intra_param_end) ) # both inclusive if len(shard_param_indices_range) == 0: shard_param_indices = (0, 0) assert len(shard_param_offsets) == 0 else: shard_param_indices = ( shard_param_indices_range[0], shard_param_indices_range[-1], ) assert ( len(shard_param_offsets) == shard_param_indices[-1] - shard_param_indices[0] + 1 ) return tuple(shard_param_offsets), shard_param_indices @staticmethod def _get_unpadded_shard( tensor: Tensor, rank: int, world_size: int, ) -> Tuple[Tensor, int]: """ Returns the shard of ``tensor`` without any padding for the given ``rank`` and ``world_size`` and the numel to pad for that shard. If ``tensor`` is already flattened or may be viewed in the flattened shape (which is true in the expected usage), then this method does not allocate any new tensor memory. """ chunks = torch.flatten(tensor).chunk(world_size) if len(chunks) < (rank + 1): # This rank gets an empty chunk fully padded with zeros since there # are not enough chunks across ranks chunk = chunks[0].new_empty(0) else: chunk = chunks[rank] numel_to_pad = chunks[0].numel() - chunk.numel() assert ( numel_to_pad >= 0 ), "Chunk's size should be at most the first chunk's size" return chunk, numel_to_pad @staticmethod def _get_shard( tensor: Tensor, rank: int, world_size: int, ) -> Tuple[Tensor, int]: """ Returns the shard of ``tensor`` with padding for the given ``rank`` and ``world_size`` and the numel padded for that shard. This method allocates new memory (via :meth:`clone`) since the unsharded ``tensor`` may be deallocated after this method returns. """ chunk, numel_to_pad = FlatParamHandle._get_unpadded_shard( tensor, rank, world_size ) shard = chunk.clone() if numel_to_pad > 0: shard = F.pad(shard, [0, numel_to_pad]) return shard, numel_to_pad @staticmethod def _get_sharded_size(tensor: Tensor, rank: int, world_size: int) -> torch.Size: """ Returns the shape of ``tensor`` after sharding including padding. This requires ``tensor`` to have 1D shape and ensures that the returned shape is 1D. """ assert len(tensor.shape) == 1, f"{tensor.shape}" unpadded_sharded_tensor, numel_to_pad = FlatParamHandle._get_unpadded_shard( tensor, rank, world_size ) unpadded_sharded_size = unpadded_sharded_tensor.size() assert len(unpadded_sharded_size) == 1, f"{unpadded_sharded_size}" return torch.Size([unpadded_sharded_size[0] + numel_to_pad]) def _get_flat_param_offsets(self) -> List[Tuple[int, int]]: """Returns [start, end] offsets of each original parameter's flattened data in the unsharded flattened parameter (without padding).""" cumulative_sum = list(accumulate(self.flat_param._numels)) starts = [0] + cumulative_sum[:-1] ends = [end - 1 for end in cumulative_sum] # inclusive param_offsets = list(zip(starts, ends)) return param_offsets def shard_metadata( self, ) -> FlatParamShardMetadata: """Returns shard-related metadata specific to this rank's shard of the flattened parameter.""" assert hasattr(self.flat_param, "_shard_indices") and hasattr( self.flat_param, "_shard_param_offsets" ), "Shard metadata has not been initialized" shard_param_start_index = self.flat_param._shard_indices[0] # type: ignore[attr-defined] shard_param_end_index = self.flat_param._shard_indices[1] # type: ignore[attr-defined] sl = ( slice(shard_param_start_index, shard_param_end_index + 1) if shard_param_start_index <= shard_param_end_index else slice(0, 0) ) return FlatParamShardMetadata( self.flat_param._fqns[sl], self.flat_param._shapes[sl], self.flat_param._numels[sl], self.flat_param._shard_param_offsets[:], # type: ignore[attr-defined] ) @no_type_check @torch.no_grad() def init_flat_param_attributes(self) -> None: """ This initializes some attributes on the handle's ``FlatParameter``. This should be called during lazy initialization since it requires the parameter to be on the compute device if not offloading to CPU and we want to give users the chance to move the parameter appropriately after the FSDP constructor. For each tensor attribute on the ``FlatParameter``, see the unshard and reshard methods in this class for the allocation and free pattern. """ flat_param = self.flat_param if flat_param.dtype != self._orig_param_dtype: # Entering this branch means that the user changed the parameter # dtype after FSDP initialization, in which case we may need to # refresh some saved dtype attributes (dtypes specified as a part # of mixed precision take precedence). if not self._low_prec_param_dtype_specified: self._fwd_bwd_param_dtype = flat_param.dtype # For `reduce_dtype`, require `param_dtype` was not specified since # then we infer the `reduce_dtype` from the specified `param_dtype` if ( not self._low_prec_reduce_dtype_specified and not self._low_prec_param_dtype_specified ): self._reduce_dtype = flat_param.dtype self._orig_param_dtype = flat_param.dtype cpu_device = torch.device("cpu") if self._offload_params: p_assert( flat_param.device == cpu_device, f"Expects the `FlatParameter` to be on CPU when parameter CPU " f"offloading is enabled, not {flat_param.device}", ) else: self._check_on_compute_device(self.flat_param) flat_param._local_shard = flat_param.data if self._offload_params: # Pin the memory for faster H2D transfer flat_param._local_shard = flat_param._local_shard.pin_memory() # Pre-allocate the sharded gradient on CPU to enable non-blocking # D2H transfer during the backward pass flat_param._cpu_grad = torch.zeros_like( flat_param._local_shard, device=cpu_device ).pin_memory() if self._uses_param_mixed_precision: # For parameter mixed precision, we maintain a low precision # sharded tensor on the compute device to be all-gathered (for # sharded strategies) or directly used (for `NO_SHARD`) for # computation. flat_param._mp_shard = torch.zeros_like( flat_param._local_shard, device=self.device, dtype=self._fwd_bwd_param_dtype, ) _free_storage(flat_param._mp_shard) if self.uses_sharded_strategy: # We maintain a padded unsharded tensor that serves as the # all-gather destination and owns the original parameter storages. unsharded_param_dtype = ( self._fwd_bwd_param_dtype if self._uses_param_mixed_precision else flat_param.dtype ) # use low precision if parameter mixed precision is enabled padded_unsharded_numel = flat_param.numel() * self.world_size flat_param._full_param_padded = torch.zeros( padded_unsharded_numel, device=self.device, dtype=unsharded_param_dtype, ) flat_param._padded_unsharded_size = flat_param._full_param_padded.size() _free_storage(flat_param._full_param_padded) if self._uses_param_mixed_precision: # For parameter mixed precision, we maintain a full precision # padded unsharded tensor for when we force full precision. flat_param._full_prec_full_param_padded = torch.zeros( padded_unsharded_numel, device=self.device, dtype=flat_param.dtype, # full precision ) _free_storage(flat_param._full_prec_full_param_padded) ################### # UNSHARD/RESHARD # ################### def pre_unshard(self) -> bool: """ Returns: ``False`` if this is a no-op and ``True`` otherwise. Postcondition: ``self.flat_param`` 's data is on the device for communication and is what should be all-gathered. This means that it matches the dtype of the expected unsharded parameter. """ ret = False if self._use_orig_params: ret = self._writeback_orig_params() if ( self.uses_sharded_strategy and not self._offload_params and not self.needs_unshard() ): pass # no-op elif self._uses_param_mixed_precision and not self._force_full_precision: self._use_low_precision_shard() ret = True elif self._offload_params and self.flat_param.device != self.device: # NOTE: This creates a new tensor distinct from any attributes. self.flat_param_to(self.device, non_blocking=True) ret = True self._check_on_compute_device(self.flat_param) return ret def _use_low_precision_shard(self): """ Allocates the low precision shard directly on the compute device and switches to using the low precision sharded flattened parameter. """ self._check_low_precision_shard() flat_param = self.flat_param _alloc_storage( flat_param._mp_shard, flat_param._local_shard.size() # type: ignore[attr-defined] ) # `copy_()` implicitly casts to the low precision flat_param._mp_shard.copy_( # type: ignore[attr-defined] flat_param._local_shard.to( # type: ignore[attr-defined] self.device, non_blocking=True ) ) # Invariant: `_mp_shard` is always on the compute device. flat_param.data = flat_param._mp_shard # type: ignore[attr-defined] def unshard(self): """ Runs the unshard logic. This includes all-gathering the flattened parameter and switching to using the unsharded flattened parameter. If the handle does not need unsharding, then this only switches to using the unsharded flattened parameter. For ``NO_SHARD``, this is a no-op. If FSDP is in :meth:`summon_full_params` and the handle uses parameter mixed precision, then the parameter is forced to full precision. """ if not self.needs_unshard(): # Even when not needing an unshard, we should switch to using # the unsharded flattened parameter unsharded_flat_param = ( self._get_padded_unsharded_flat_param() if self.uses_sharded_strategy else self.flat_param ) self._use_unsharded_flat_param(unsharded_flat_param) return unsharded_flat_param = self._alloc_padded_unsharded_flat_param() padded_unsharded_flat_param = self._all_gather_flat_param(unsharded_flat_param) self._use_unsharded_flat_param(padded_unsharded_flat_param) def needs_unshard(self) -> bool: """Returns if the handle's flattened parameter needs to be unsharded.""" if not self.uses_sharded_strategy: return False unsharded_flat_param = self._get_padded_unsharded_flat_param() already_unsharded = ( unsharded_flat_param._typed_storage()._size() == unsharded_flat_param.numel() ) return not already_unsharded def _alloc_padded_unsharded_flat_param(self): """ Allocates the *padded* unsharded flattened parameter. The unpadded unsharded flattened parameter is always a view into the padded one. This padded parameter is saved to a different attribute on the ``FlatParameter`` depending on if we force full precision. """ self._check_sharded_strategy() flat_param = self.flat_param unsharded_flat_param = self._get_padded_unsharded_flat_param() self._check_storage_freed(unsharded_flat_param) _alloc_storage(unsharded_flat_param, flat_param._padded_unsharded_size) # type: ignore[attr-defined] return unsharded_flat_param def _get_padded_unsharded_flat_param(self) -> torch.Tensor: """ Returns a reference to the padded unsharded flattened parameter depending on the calling context. This should only be called if using a sharded strategy. """ self._check_sharded_strategy() flat_param = self.flat_param if self._force_full_precision: # When parameter mixed precision is enabled, we use a different # tensor as the all-gather destination to preserve the invariant # that `_full_param_padded` is in the low precision unsharded_flat_param = flat_param._full_prec_full_param_padded # type: ignore[attr-defined] p_assert( unsharded_flat_param.dtype != self._fwd_bwd_param_dtype, f"Expects full precision but got {self._fwd_bwd_param_dtype}", ) else: unsharded_flat_param = flat_param._full_param_padded # type: ignore[attr-defined] return unsharded_flat_param def _all_gather_flat_param( self, padded_unsharded_flat_param: Tensor, ) -> Tensor: """ All-gathers the handle's flattened parameter to the destination ``padded_unsharded_flat_param``, and switches to using the all-gathered tensor. """ p_assert( hasattr(self, "process_group") and hasattr(self, "world_size"), "Expects a process group and world size to have been set via `shard()`", ) sharded_flat_param = self.flat_param.data expected_numel = sharded_flat_param.numel() * self.world_size p_assert( padded_unsharded_flat_param.numel() == expected_numel, f"Expects {expected_numel} numel but got {padded_unsharded_flat_param.numel()}", ) dist.all_gather_into_tensor( padded_unsharded_flat_param, sharded_flat_param, self.process_group, ) return padded_unsharded_flat_param def _use_unsharded_flat_param( self, padded_unsharded_flat_param: torch.Tensor, ) -> None: """ Switches to using the *unpadded* unsharded flattened parameter, which is a view into the *padded* unsharded flattened parameter. """ unsharded_size = self.flat_param._unpadded_unsharded_size self.flat_param.data = padded_unsharded_flat_param[ : unsharded_size.numel() ].view( unsharded_size ) # this `.view()` is not autograd visible in_forward = self._training_state == HandleTrainingState.FORWARD in_pre_backward = self._training_state == HandleTrainingState.BACKWARD_PRE if self._use_orig_params: # We use `Tensor` views in the forward so that they are tracked by # autograd. We use them in the pre-backward as well to support # reentrant activation checkpointing, which needs the views to be # tracked by autograd in the backward pass's recomputed forward. self._use_unsharded_views( as_params=(not in_forward and not in_pre_backward) ) elif in_forward: self._use_unsharded_views(as_params=False) def post_unshard(self): """ Runs the post-unshard logic. This includes freeing the low precision shard if needed. """ if self._uses_param_mixed_precision and self.uses_sharded_strategy: self._free_low_precision_sharded_param() self._check_on_compute_device(self.flat_param) def _free_low_precision_sharded_param(self): """Frees the low precision sharded flattened parameter.""" self._check_low_precision_shard() # `_mp_shard` is allocated in the pre-unshard stream, consumed in the # unshard stream for sharded strategies, and consumed in both the # unshard and default streams for `NO_SHARD`. For sharded strategies, # the current stream here is the unshard stream, and for `NO_SHARD`, # it is the default stream. For `NO_SHARD`, only recording for the # default stream suffices since the default stream waits for the # unshard stream. _no_dispatch_record_stream( self.flat_param._mp_shard, torch.cuda.current_stream() # type: ignore[attr-defined] ) _free_storage(self.flat_param._mp_shard) # type: ignore[attr-defined] @torch.no_grad() def unshard_grad(self): """ Unshards the handle's ``FlatParameter`` 's gradient. If all ranks have ``None`` gradient, then all original parameters will as well. This method performs an all-reduce and an all-gather. The additional all-reduce is tolerable since this method is not meant to be used on the computation critical path. Postcondition: ``_saved_grad_shard`` is defined and contains the value to set ``flat_param.grad`` after gradients are resharded. """ if not self.uses_sharded_strategy: self._use_unsharded_grad_views() return flat_param = self.flat_param self._check_unsharded(flat_param) # Check if all ranks have a `None` gradient num_grad_none = torch.zeros(1, dtype=torch.int32, device=self.device) num_grad_none[0] = flat_param.grad is None dist.all_reduce(num_grad_none, group=self.process_group) if num_grad_none[0] == self.world_size: flat_param._saved_grad_shard = None # type: ignore[attr-defined] self._use_unsharded_grad_views() return padded_unsharded_grad = torch.empty( flat_param._padded_unsharded_size, # type: ignore[attr-defined] device=self.device, ) if flat_param.grad is None: # In the case that only some ranks have `None` gradient, we use # zeros to approximate as a best effort attempt if self._debug_level == dist.DebugLevel.DETAIL: warnings.warn( f"[Rank {self.rank}] Only some but not all ranks have a " "`None` `FlatParameter` gradient, so FSDP is using zeros to " "approximate those ranks' sharded gradients being `None`" ) flat_param._saved_grad_shard = None # type: ignore[attr-defined] sharded_grad = torch.zeros(flat_param._sharded_size, device=self.device) # type: ignore[attr-defined] else: self._check_sharded(flat_param.grad) flat_param._saved_grad_shard = flat_param.grad # type: ignore[attr-defined] sharded_grad = flat_param._saved_grad_shard # type: ignore[attr-defined] dist.all_gather_into_tensor( padded_unsharded_grad, sharded_grad, self.process_group ) unsharded_size = self.flat_param._unpadded_unsharded_size flat_param.grad = padded_unsharded_grad[: unsharded_size.numel()].view( unsharded_size ) self._use_unsharded_grad_views() def reshard_grad(self): if self._use_orig_params: self._use_sharded_grad_views() if not self.uses_sharded_strategy: return self.flat_param.grad = self.flat_param._saved_grad_shard # type: ignore[attr-defined] delattr(self.flat_param, "_saved_grad_shard") def prepare_gradient_for_backward(self): """ Prepares the gradient for the backward computation by saving and clearing any existing sharded gradient in ``.grad`` to enable computing a new unsharded gradient. """ p_assert( self._training_state in (HandleTrainingState.BACKWARD_PRE, HandleTrainingState.IDLE), "Expects to be in `BACKWARD_PRE` or `IDLE` (if prefetching)", ) flat_param = self.flat_param if flat_param.grad is not None and ( flat_param.grad.size() != flat_param._unpadded_unsharded_size or flat_param.grad.device != flat_param.device # grad on CPU ): self._check_on_compute_device(self.flat_param) grad_offloaded = flat_param.grad.device != self.device p_assert( not grad_offloaded or self._offload_params, f"Expects the sharded gradient to be on {self.device} " f"but got {flat_param.grad.device}", ) prev_iter_synced_gradients = ( flat_param.grad.size() == flat_param._local_shard.size() # type: ignore[attr-defined] ) if prev_iter_synced_gradients: # TODO (awgu): Gradient accumulation outside `no_sync()` # does not work with CPU offloading. The issue should be # that, in the post-backward hook, we cannot do an addition # between a CPU tensor (the existing sharded gradient) and # a GPU tensor (the new sharded gradient). if not grad_offloaded: flat_param._saved_grad_shard = flat_param.grad.data # type: ignore[attr-defined] sharded_grad = flat_param._saved_grad_shard # type: ignore[attr-defined] else: p_assert( hasattr(flat_param, "_cpu_grad"), "`_cpu_grad` should be defined if the gradient is on CPU", ) sharded_grad = flat_param._cpu_grad # type: ignore[attr-defined] # If user specified to keep the gradient in low precision, then # the gradient may still be of the low precision dtype if the # user did not set the gradient to `None` after the previous # backward, in which case FSDP should cast back to the full # precision dtype so that FSDP can accumulate in that dtype in # the post-backward hook and assign to `.grad` in that dtype in # the post-backward callback. local_shard_dtype = flat_param._local_shard.dtype # type: ignore[attr-defined] if ( self._keep_low_precision_grads and sharded_grad.dtype != local_shard_dtype ): sharded_grad.data = sharded_grad.to(local_shard_dtype) else: padded_unsharded_size = flat_param._padded_unsharded_size # type: ignore[attr-defined] p_assert( flat_param.grad.size() == padded_unsharded_size, "Expects `.grad` to be the unsharded gradient in " f"`no_sync()` with size {padded_unsharded_size} " f"but got size {flat_param.grad.size()}", ) flat_param.grad = None def prepare_gradient_for_optim(self): """ Prepares the gradient for optimizer computation by moving the sharded gradient to the ``.grad`` attribute. """ def cast_grad_to_param_dtype_if_needed(flat_param): if self._keep_low_precision_grads: assert flat_param.grad is not None # mypy if flat_param.grad.dtype != self._fwd_bwd_param_dtype: flat_param.grad.data = flat_param.grad.to(self._fwd_bwd_param_dtype) if self._use_orig_params: self._use_sharded_grad_views() flat_param = self.flat_param # TODO (awgu): We should replace these conditional checks to encode # the logical intention more directly. if hasattr(flat_param, "_cpu_grad"): # NOTE: This branch includes `NO_SHARD`. self._check_sharded(flat_param) self._check_on_cpu(flat_param) flat_param.grad = flat_param._cpu_grad # type: ignore[attr-defined] cast_grad_to_param_dtype_if_needed(flat_param) elif hasattr(flat_param, "_saved_grad_shard"): self._check_sharded(flat_param) self._check_on_compute_device(flat_param) self._check_on_compute_device(flat_param._saved_grad_shard) # type: ignore[attr-defined] # If no sharded gradient was computed this iteration, then there is # no need to forward `_saved_grad_shard` to `grad` if flat_param._post_backward_called: # type: ignore[attr-defined] flat_param.grad = flat_param._saved_grad_shard # type: ignore[attr-defined] cast_grad_to_param_dtype_if_needed(flat_param) else: p_assert( not self.uses_sharded_strategy or not flat_param._post_backward_called, # type: ignore[attr-defined] "All sharded parameters that received a gradient in the " "post-backward should use `_saved_grad_shard`", ) # Delete `_saved_grad_shard` since its existence indicates a previous # gradient to accumulate with in the post-backward hook if hasattr(flat_param, "_saved_grad_shard"): delattr(flat_param, "_saved_grad_shard") @contextlib.contextmanager def to_cpu(self): """ Moves the unpadded unsharded flattened parameter to CPU while in the context and moves it back to the previous device upon exit. For now, this assumes the ``FlatParameter`` is the unpadded unsharded flattened parameter since (1) there is no reason to include the padding in the copy and (2) there is no use case for the sharded flattened parameter. Precondition: ``self.flat_param`` 's data is the unpadded unsharded flattened parameter on the compute device, and the handle uses a sharded strategy. Postcondition: Same as the precondition. """ self._check_sharded_strategy() p_assert( self.flat_param.size() == self.flat_param._unpadded_unsharded_size, f"Expects size {self.flat_param._unpadded_unsharded_size} but got {self.flat_param.size()}", ) self._check_on_compute_device(self.flat_param) # Check that the unpadded unsharded flattened parameter is a view into # the padded unsharded flattened parameter as expected # NOTE: This check is not strictly needed for correctness but is a # useful sanity check since the tensor should only be used internally. unpadded_storage_ptr = self.flat_param._typed_storage()._data_ptr() padded_storage_ptr = ( self._get_padded_unsharded_flat_param()._typed_storage()._data_ptr() ) p_assert( unpadded_storage_ptr == padded_storage_ptr, "Expects the unpadded parameter to be a view into the padded parameter", ) self.flat_param_to(torch.device("cpu")) self._free_unsharded_flat_param() try: yield finally: p_assert( self.flat_param.size() == self.flat_param._unpadded_unsharded_size, f"Expects size {self.flat_param._unpadded_unsharded_size} but got {self.flat_param.size()}", ) padded_unsharded_flat_param = self._alloc_padded_unsharded_flat_param() # Copy from CPU to the compute device padded_unsharded_flat_param[: self.flat_param.numel()].copy_( self.flat_param ) self._use_unsharded_flat_param(padded_unsharded_flat_param) def reshard(self, free_unsharded_flat_param: bool): """ Runs the reshard logic. This includes freeing the unsharded flattened parameter if ``free_unsharded_flat_param`` and switching to using the sharded flattened parameter. """ # Switch to the sharded `FlatParameter` before freeing to prevent # "use-after-free"-type bugs with external profiling tools, where for # `use_orig_params=True`, the `param` does not point to valid memory # when setting `param.data = ...` in `_use_sharded_views()`. self._use_sharded_flat_param() if free_unsharded_flat_param: self._free_unsharded_flat_param() def post_reshard(self): """ Runs the post-reshard logic. This includes freeing any memory that can now be freed given that the ``FlatParameter`` points to the full precision sharded flattened parameter. Precondition: ``self.flat_param`` 's data points to the full precision sharded flattened parameter. """ # For `NO_SHARD`, `_mp_shard` is not freed in the post-unshard since # it is also the low precision *unsharded* flattened parameter. Hence, # we delay the free until the reshard. if ( self._uses_param_mixed_precision and not self.uses_sharded_strategy and not self._force_full_precision # did not use the low precision shard ): self._free_low_precision_sharded_param() def _free_unsharded_flat_param(self): """ Frees the padded unsharded flattened parameter. The tensor to free depends on the calling context since the unshard may have forced full precision, in which case a different tensor is used. """ self._check_sharded_strategy() unsharded_flat_param = self._get_padded_unsharded_flat_param() self._check_storage_allocated(unsharded_flat_param) self._check_on_compute_device(unsharded_flat_param) # Do not free the memory until all ops in the current stream finish _no_dispatch_record_stream(unsharded_flat_param, torch.cuda.current_stream()) _free_storage(unsharded_flat_param) def _use_sharded_flat_param(self) -> None: """Switches to using the sharded flattened parameter.""" flat_param = self.flat_param if self._offload_params: device = flat_param._local_shard.device # type: ignore[attr-defined] p_assert( device == torch.device("cpu"), f"Expects the local shard to be on CPU but got {device}", ) flat_param.data = flat_param._local_shard # type: ignore[attr-defined] if self._use_orig_params: self._use_sharded_views() # For the post-forward reshard, we may try to use sharded gradient # views (or unsharded gradient views if a gradient was accumulated # in `no_sync()`), but for the post-backward reshard, we delay the # call to after the reduce-scatter. if self._training_state == HandleTrainingState.FORWARD: # TODO: Change `_unpadded_unsharded_size` if we change the # gradient to be computed directly with padding. accumulated_grad_in_no_sync = ( flat_param.grad is not None and self.uses_sharded_strategy and flat_param.grad.shape == flat_param._unpadded_unsharded_size ) if accumulated_grad_in_no_sync: self._use_unsharded_grad_views() else: self._use_sharded_grad_views() ######### # VIEWS # ######### @staticmethod def _get_unflat_views( flat_param: FlatParameter, tensor: Optional[torch.Tensor] = None, ) -> Iterator[Tensor]: """ Returns unflattened ``Tensor`` views into ``tensor`` if it is not ``None`` or ``flat_param`` otherwise, where the unflattening is based on ``flat_param`` 's metadata. In other words, to get views into the unsharded flattened parameter, pass ``tensor`` as ``None``, but to get views into tensor optimizer state, pass ``tensor`` as the optimizer state tensor. """ if tensor is None: tensor = flat_param p_assert( tensor.numel() == flat_param._unpadded_unsharded_size.numel(), f"Expects {flat_param._unpadded_unsharded_size.numel()} numel but got " f"{tensor.numel()} numel", ) views = ( _ext_post_unflatten_transform(subtensor.view(shape), param_extension) for (subtensor, shape, param_extension) in zip( torch.split(tensor, flat_param._numels, dim=0), # type: ignore[arg-type] flat_param._shapes, flat_param._param_extensions, ) ) return views def _use_unsharded_views(self, as_params: bool) -> None: """ Unflattens the unsharded flattened parameter by setting the original module parameter variables to be views into it. Args: as_params (bool): If ``True``, then registers the original parameters as ``nn.Parameter`` s; if ``False``, then registers the original parameters only as ``Tensor`` s. ``False`` should be used during forward/backward computation and when hiding the original parameters from :meth:`nn.Module.named_parameters`. """ self._check_unsharded(self.flat_param) views = self._get_unflat_views(self.flat_param) for i, (view, (param_name, module, _)) in enumerate( zip(views, self.flat_param._param_infos) ): if hasattr(module, param_name): delattr(module, param_name) if self._use_orig_params and as_params: if type(view) is DTensor: # A `DTensor` `view` is not compatible with assigning # `param.data = view`, so we cannot preserve the parameter # variable. setattr(module, param_name, nn.Parameter(view)) continue param = self.flat_param._params[i] # type: ignore[index] setattr(module, param_name, param) param.data = view elif as_params: module.register_parameter(param_name, nn.Parameter(view)) else: # `as_params=False` param_var: Tensor = view if self._use_orig_params: if self._training_state == HandleTrainingState.FORWARD: assert self.flat_param._tensors is not None # Save the `Tensor` for the pre-backward self.flat_param._tensors[i] = view # save for pre-backward elif self._training_state == HandleTrainingState.BACKWARD_PRE: # Use the saved `Tensor` variable from the forward to # preserve the autograd graph so that the post-backward # hook fires (e.g. for reentrant AC) assert self.flat_param._tensors is not None # mypy tensor = self.flat_param._tensors[i] p_assert( tensor is not None, "Expects `Tensor` to have been saved in forward", ) tensor.data = view # type: ignore[union-attr] assert tensor is not None # mypy param_var = tensor setattr(module, param_name, param_var) if ( self._use_orig_params and self._training_state == HandleTrainingState.FORWARD ): module._parameters[param_name] = param_var # type: ignore[assignment] for i, ( param_name, module, _, prim_param_name, prim_module, prim_module_name, ) in enumerate(self.flat_param._shared_param_infos): if hasattr(module, param_name): delattr(module, param_name) p_assert( hasattr(prim_module, prim_param_name), f"Module {prim_module_name} is missing parameter {prim_param_name}", ) prim_param: Union[Tensor, nn.Parameter] = getattr( prim_module, prim_param_name ) p_assert( not as_params or isinstance(prim_param, nn.Parameter), f"as_params={as_params} type(prim_param)={type(prim_param)}", ) if self._use_orig_params and as_params: shared_param = self.flat_param._shared_params[i] # type: ignore[index] setattr(module, param_name, shared_param) shared_param.data = prim_param elif as_params: assert isinstance(prim_param, nn.Parameter) module.register_parameter(param_name, prim_param) else: setattr(module, param_name, prim_param) if ( self._use_orig_params and self._training_state == HandleTrainingState.FORWARD ): module._parameters[param_name] = prim_param # type: ignore[assignment] def _use_unsharded_grad_views(self) -> None: """ Unflattens the unsharded flattened parameter's gradient by setting the original module parameter variables' gradients to be views into it. """ # Expects the gradient to be in `flat_param.grad` if self.flat_param.grad is None: assert self.flat_param._params is not None # mypy assert self.flat_param._shared_params is not None # mypy for param in chain( self.flat_param._params, # type: ignore[attr-defined] self.flat_param._shared_params, # type: ignore[attr-defined] ): param.grad = None return self._check_unsharded(self.flat_param.grad) views = self._get_unflat_views(self.flat_param, self.flat_param.grad) for i, (view, (param_name, module, _)) in enumerate( zip(views, self.flat_param._param_infos) ): p_assert( hasattr(module, param_name), f"{self.flat_param._fqns[i]} is missing", ) param = getattr(module, param_name) if param.shape != view.shape or param.dtype != view.dtype: # NOTE: This is a hack using `.data` to side step the # check that parameter/gradient sizes and dtypes match. Here, # `param` can have the sharded size, and `grad` can have the # unsharded size. Orthgonally, `param` can have the full # precision dtype from `reshard()`, and `grad` can have the # parameter low precision dtype. Both of these mismatches # happen when running in `no_sync()`. if param.grad is None: param.grad = torch.empty_like(param) param.grad.data = view else: param.grad = view for i, ( param_name, module, module_name, prim_param_name, prim_module, _, ) in enumerate(self.flat_param._shared_param_infos): p_assert( hasattr(module, param_name), f"{module_name + '.' + param_name if module_name else param_name} is missing", ) # did not save FQN info in `_shared_param_infos` param = getattr(module, param_name) prim_param = getattr(prim_module, prim_param_name) if ( param.shape != prim_param.grad.shape or param.dtype != prim_param.grad.dtype ): # NOTE: This is the same hack to use `.data` to side step the # size check. if param.grad is None: param.grad = torch.empty_like(param) param.grad.data = prim_param.grad else: param.grad = prim_param.grad @contextlib.contextmanager def unflatten_as_params(self) -> Generator: """ Assumes the flattened parameter is unsharded. When in the context, unflattens the original parameters as ``nn.Parameter`` views into the flattened parameter, and after the context, restores the original parameters as ``Tensor`` views into the flattened parameter. """ self._use_unsharded_views(as_params=True) try: yield finally: self._use_unsharded_views(as_params=False) @torch.no_grad() def _use_sharded_views(self) -> None: """ Sets the original module parameter variables' data to be flattened views into the sharded flattened parameter. The views are kept as flattened to simplify the case where a parameter is sharded across ranks. Parameters whose data is not present in the sharded flattened parameter have their data set to a size-0 empty tensor. We do not delete them to ensure to preserve expected behaviors like model printability. Parameters whose data is present must preserve their variables to be passable to an optimizer. """ if not self.uses_sharded_strategy: # For `NO_SHARD`, use the *unflattened* unsharded views since we # have the unsharded parameter self._use_unsharded_views(as_params=True) return self._check_sharded(self.flat_param) start, end = self.flat_param._shard_indices # type: ignore[attr-defined] offset = 0 assert self.flat_param._params is not None for i, (param, (param_name, module, _)) in enumerate( zip(self.flat_param._params, self.flat_param._param_infos) ): setattr(module, param_name, param) in_sharded_flat_param = ( i >= start and i <= end and self.flat_param._shard_param_offsets # type: ignore[attr-defined] ) if in_sharded_flat_param: param_start, param_end = self.flat_param._shard_param_offsets[i - start] # type: ignore[attr-defined] numel_in_shard = param_end - param_start + 1 param.data = self.flat_param[offset : offset + numel_in_shard] offset += numel_in_shard else: # Allow the original data to be freed via garbage collection param.data = torch.empty( 0, dtype=self.flat_param.dtype, # in case `flat_param` changed dtype device=self.flat_param.device, requires_grad=False, ) assert self.flat_param._shared_params is not None for i, ( param, (param_name, module, _, prim_param_name, prim_module, _), ) in enumerate( zip(self.flat_param._shared_params, self.flat_param._shared_param_infos) ): setattr(module, param_name, param) prim_param = getattr(prim_module, prim_param_name) param.data = prim_param # could be both empty and non-empty if self._training_state == HandleTrainingState.BACKWARD_POST: assert self.flat_param._tensors is not None # mypy # Clear the saved `Tensor`s since they are unneeded now for i in range(len(self.flat_param._tensors)): self.flat_param._tensors[i] = None # type: ignore[index] @torch.no_grad() def _use_sharded_grad_views(self) -> None: """ Sets the original module parameter variables' gradients to be flattened views into the sharded flattened parameter's gradient. This is a no-op if there is no gradient. Parameters whose data is not present in the sharded flattened parameter and parameters with ``requires_grad=False`` have their gradients set to ``None``. Since the gradient variables do not need to be preserved, this method does not manipulate existing ``Tensor`` data directly and creates new ``Tensor`` variables instead. """ flat_param = self.flat_param self._check_sharded(flat_param) grad = self.sharded_grad if grad is None: assert flat_param._params is not None # mypy assert flat_param._shared_params is not None # mypy for param in chain(flat_param._params, flat_param._shared_params): # type: ignore[attr-defined] param.grad = None return self._check_sharded(grad) start, end = flat_param._shard_indices # type: ignore[attr-defined] offset = 0 assert flat_param._params is not None for i, param in enumerate(flat_param._params): in_sharded_flat_param = ( i >= start and i <= end and flat_param._shard_param_offsets # type: ignore[attr-defined] ) if in_sharded_flat_param: param_start, param_end = flat_param._shard_param_offsets[i - start] # type: ignore[attr-defined] numel_in_shard = param_end - param_start + 1 assert flat_param._is_grad_none is not None # mypy if param.requires_grad and not flat_param._is_grad_none[i]: if self._keep_low_precision_grads or param.dtype != grad.dtype: # NOTE: This is a hack using `.data` to side step the # check that parameter/gradient dtypes match. Here, # `param` has full precision; `grad` has low precision. if param.grad is None: # `.grad` must have the same shape as `param` param.grad = torch.empty_like(param) param.grad.data = grad[ offset : offset + numel_in_shard ].reshape(param.shape) else: param.grad = grad[offset : offset + numel_in_shard].reshape( param.shape ) else: param.grad = None offset += numel_in_shard else: param.grad = None assert flat_param._shared_params is not None for i, (param, (_, _, _, prim_param_name, prim_module, _)) in enumerate( zip(flat_param._shared_params, flat_param._shared_param_infos) ): in_sharded_flat_param = hasattr(prim_module, prim_param_name) if in_sharded_flat_param and param.requires_grad: prim_param = getattr(prim_module, prim_param_name) param.grad = prim_param.grad # share the same reference else: param.grad = None @torch.no_grad() def _writeback_orig_params(self) -> bool: """ Iterates over the original parameters and writes back any parameters that changed storages (due to a non-inplace operator) to the handle's ``FlatParameter``. This method preserves the ``FlatParameter` 's device even if an original parameter's device changes. Raises: RuntimeError: If an original parameter or gradient changes storages but no longer has the expected flattened shape. Returns: ``True`` if some writeback happened, and ``False`` otherwise. """ if self.uses_sharded_strategy and not self.is_sharded(self.flat_param): # For `NO_SHARD`, we may still need to writeback return False flat_param = self.flat_param start, end = flat_param._shard_indices # type: ignore[attr-defined] offset = 0 assert flat_param._params is not None wroteback = False for i, (param, (param_name, module, _)) in enumerate( zip(flat_param._params, flat_param._param_infos) ): if not hasattr(module, param_name): # Do not writeback if original parameters are deregistered # (e.g. during model checkpointing) continue in_sharded_flat_param = ( i >= start and i <= end and self.flat_param._shard_param_offsets # type: ignore[attr-defined] ) if not in_sharded_flat_param: continue param_start, param_end = flat_param._shard_param_offsets[i - start] # type: ignore[attr-defined] numel_in_shard = param_end - param_start + 1 # Check for parameter writeback param_changed = getattr(module, param_name) is not param needs_param_writeback = ( param_changed # changed parameter variable itself or not _same_storage(param, flat_param) # changed `.data` ) if param_changed: # NOTE: The gradient is not preserved after a parameter change. param = getattr(module, param_name) flat_param._params[i] = param if needs_param_writeback: expected_shape = torch.Size([numel_in_shard]) self._writeback_tensor( param, flat_param, i, expected_shape, offset, True ) wroteback = True # Check for gradient writeback # NOTE: Since this method is called in the pre-unshard, which is # only called during computation in the pre-forward or # pre-backward, the sharded gradient should be guaranteed to be in # `.grad`, not in `._saved_grad_shard`. if param.grad is None and flat_param.grad is not None: expected_shape = torch.Size([numel_in_shard]) self._writeback_tensor( None, flat_param.grad, i, expected_shape, offset, False ) elif param.grad is not None: # For `NO_SHARD` + CPU offloading, `_cpu_grad` is always in # memory and owns the gradient storage, so it will never # require gradient writeback. flat_param_grad = ( flat_param.grad if self.uses_sharded_strategy or not self._offload_params else flat_param._cpu_grad # type: ignore[attr-defined] ) needs_grad_writeback = flat_param_grad is None or not _same_storage( param.grad, flat_param_grad ) if needs_grad_writeback: if flat_param_grad is None: flat_param_grad = torch.zeros_like(flat_param) expected_shape = torch.Size([numel_in_shard]) self._writeback_tensor( param.grad, flat_param_grad, i, expected_shape, offset, False ) flat_param.grad = flat_param_grad offset += numel_in_shard # TODO (awgu): Handle shared parameters. We need to re-generate the # shared parameter data structures in case sharedness changed. for i, ( param_name, module, _, prim_param_name, prim_module, _, ) in enumerate(flat_param._shared_param_infos): if getattr(module, param_name) is not getattr(prim_module, prim_param_name): raise NotImplementedError( "Changing shared parameters is not supported yet" ) return wroteback def _writeback_tensor( self, src_tensor: Optional[Tensor], dst_tensor: Tensor, tensor_index: int, expected_shape: torch.Size, offset: int, is_param: bool, # else gradient ) -> None: """ Writes back ``src_tensor`` to ``dst_tensor`` at offset ``offset``, where ``src_tensor`` should have shape ``expected_shape``. ``is_param`` indicates if the tensor is the parameter (if ``True``) or gradient (if ``False``). If ``src_tensor`` is ``None``, then the effect is zeroing instead of copying. ``tensor_index`` gives the index of ``src_tensor`` in the metadata structures. Raises: RuntimeError: If the ``src_tensor`` does not have the expected shape. """ p_assert( len(expected_shape) == 1, f"Expects a 1D expected shape but got {expected_shape}", ) if self._debug_level == dist.DebugLevel.DETAIL: rank = self.rank if hasattr(self, "rank") else dist.get_rank() src_shape = src_tensor.shape if src_tensor is not None else None src_device = src_tensor.device if src_tensor is not None else None warnings.warn( f"[Rank {rank}] {'Parameter' if is_param else 'Gradient'} needs " f"writeback in {self._training_state}\n" f"expected shape={expected_shape} shape={src_shape} " f"expected device={dst_tensor.device} device={src_device}" ) if src_tensor is not None and src_tensor.shape != expected_shape: # NOTE: Gradient shape mismatch is not possible in practice since # the gradient shape is enforced to match that of the parameter and # we already check for parameter shape mismatch. raise RuntimeError( f"Cannot writeback when the {'parameter' if is_param else 'gradient'} " f"shape changes\nExpects {expected_shape} but got {src_tensor.shape}" ) if src_tensor is not None: dst_tensor[offset : offset + expected_shape.numel()].copy_(src_tensor) else: dst_tensor[offset : offset + expected_shape.numel()].zero_() assert self.flat_param._is_grad_none is not None self.flat_param._is_grad_none[tensor_index] = True def _clear_grads_if_needed(self): """ When ``use_orig_params=True``, sets the underlying ``flat_param.grad`` to ``None`` if *all* of the original parameters' ``.grad`` are ``None``. This is targeting ``optim.zero_grad(set_to_none=True)``, in which case we want to free the gradients as soon after the ``zero_grad()`` call as possible. """ if not self._use_orig_params: return flat_param = self.flat_param assert flat_param._params is not None if all(param.grad is None for param in flat_param._params): flat_param.grad = None def _deregister_orig_params(self): for (param_name, module, _) in self.flat_param._param_infos: if hasattr(module, param_name): delattr(module, param_name) for (param_name, module, _, _, _, _) in self.flat_param._shared_param_infos: if hasattr(module, param_name): delattr(module, param_name) ########### # HELPERS # ########### def flat_param_to(self, *args, **kwargs): """Wraps an in-place call to ``.to()`` for ``self.flat_param``.""" self.flat_param.data = self.flat_param.to(*args, **kwargs) if self._use_orig_params: # Refresh the views because their storage may have changed if self.is_sharded(self.flat_param): self._use_sharded_views() else: self._use_unsharded_views(as_params=True) def _get_modules(self) -> Set[nn.Module]: """Returns a :class:`set` of the modules whose parameters are included in this handle's flattened parameter.""" return {pi.module for pi in self.flat_param._param_infos}.union( {spi.module for spi in self.flat_param._shared_param_infos} ) def is_sharded(self, tensor: Tensor) -> bool: """ Returns if ``tensor`` is *currently* sharded. For ``NO_SHARD``, we choose to have this always return ``False`` for clarity. """ if ( not hasattr(self.flat_param, "_sharded_size") or not self.uses_sharded_strategy ): # `_sharded_size` is defined iff `handle.shard()` has been called return False sharded_size = self.flat_param._sharded_size # type: ignore[attr-defined] return tensor.size() == sharded_size def parameter_module_names(self) -> Iterator[Tuple[str, str]]: shared_param_infos = [ ParamInfo(param_name, module, module_name) for ( param_name, module, module_name, _, _, _, ) in self.flat_param._shared_param_infos ] for param_name, _, module_name in chain( self.flat_param._param_infos, shared_param_infos ): yield (param_name, module_name) def shared_parameter_module_names(self) -> Iterator[Tuple[str, str]]: for param_name, _, module_name in [ ParamInfo(param_name, module, module_name) for ( param_name, module, module_name, _, _, _, ) in self.flat_param._shared_param_infos ]: yield (param_name, module_name) @property def _fqns_in_shard(self) -> List[str]: """Returns the FQNs of the parameters present in this rank's shard.""" fqns_in_shard: List[str] = [] start, end = self.flat_param._shard_indices # type: ignore[attr-defined] for i in range(len(self.flat_param._fqns)): if i >= start and i <= end and self.flat_param._shard_param_offsets: # type: ignore[attr-defined] fqns_in_shard.append(self.flat_param._fqns[i]) return fqns_in_shard @property def sharded_grad(self) -> Optional[Tensor]: """Returns the handle's sharded gradient.""" flat_param = self.flat_param # Priority for non-`None`: `_cpu_grad` > `_saved_grad_shard` > `grad` # - CPU offloading: `_cpu_grad` # - No CPU offloading + sharded strategies: `_saved_grad_shard` # - No CPU offloading + `NO_SHARD`: `grad` if hasattr(flat_param, "_cpu_grad"): grad = flat_param._cpu_grad # type: ignore[attr-defined] elif hasattr(flat_param, "_saved_grad_shard"): grad = flat_param._saved_grad_shard # type: ignore[attr-defined] else: # If in the forward, then there may be an accumulated gradient, # which will be in `.grad` p_assert( flat_param.grad is None or not self.uses_sharded_strategy or self._training_state == HandleTrainingState.FORWARD, "Sharded strategies should use `_cpu_grad` or `_saved_grad_shard` " "unless in FORWARD (for the post-forward reshard)", ) grad = flat_param.grad return grad def _reset_is_grad_none(self) -> None: """ Resets the ``_is_grad_none`` mask as needed. This method should only be called in the post-backward after gradient computation, in which case if a parameter requires gradient, then it will surely receive a gradient and we may reset its mask entry to ``False``. """ if not self._use_orig_params: return p_assert( self._training_state == HandleTrainingState.BACKWARD_POST, "Expects to only be called in the post-backward after gradient computation", ) flat_param = self.flat_param assert flat_param._params is not None # mypy for i, param in enumerate(flat_param._params): # As long as the parameter requires gradient, it should receive a # meaningful gradient (even if the gradient happens to be zeros) if param.requires_grad: assert flat_param._is_grad_none is not None # mypy flat_param._is_grad_none[i] = False ####################### # CHECKS & INVARIANTS # ####################### def _check_sharded_strategy(self): p_assert(self.uses_sharded_strategy, "Expects sharded strategy") def _check_on_compute_device(self, tensor: Tensor): p_assert( tensor.device == self.device, f"Expects tensor to be on the compute device {self.device}", ) def _check_on_cpu(self, tensor: Tensor): p_assert( tensor.device == torch.device("cpu"), f"Expects tensor to be on CPU but got {tensor.device}", ) @staticmethod def _check_storage_freed(tensor: Tensor): storage_size: int = tensor._typed_storage()._size() p_assert( storage_size == 0, f"Expects storage to be freed but got storage with size {storage_size}", ) @staticmethod def _check_storage_allocated(tensor: Tensor): storage_size: int = tensor._typed_storage()._size() p_assert(storage_size > 0, "Expects storage to be allocated") def _check_low_precision_shard(self): p_assert( self._uses_param_mixed_precision, "Not using low precision for parameters", ) p_assert( getattr(self.flat_param, "_mp_shard", None) is not None, "Expects `_mp_shard` to exist", ) device = self.flat_param._mp_shard.device # type: ignore[attr-defined] p_assert( device == self.device, f"Expects the low precision shard to be on {self.device} but got {device}", ) def _check_unsharded(self, tensor: Tensor): msg_prefix = "Expects tensor to be unsharded " p_assert(tensor is not None, msg_prefix + "but got `None`") unsharded_size = self.flat_param._unpadded_unsharded_size p_assert( tensor.size() == unsharded_size, msg_prefix + f"with size {unsharded_size} but got {tensor.size()}", ) def _check_sharded(self, tensor: Tensor): msg_prefix = "Expects tensor to be sharded " p_assert(tensor is not None, msg_prefix + "but got `None`") sharded_size = self.flat_param._sharded_size # type: ignore[attr-defined] p_assert( tensor.size() == sharded_size, msg_prefix + f"with size {sharded_size} but got {tensor.size()}", ) ############## # PROPERTIES # ############## @property def uses_sharded_strategy(self) -> bool: return self._sharding_strategy != HandleShardingStrategy.NO_SHARD @property def _uses_param_mixed_precision(self) -> bool: return self._fwd_bwd_param_dtype != self._orig_param_dtype @property def _uses_reduce_mixed_precision(self) -> bool: return self._reduce_dtype != self._orig_param_dtype @property def _force_full_precision(self) -> bool: return ( self._training_state == HandleTrainingState.SUMMON_FULL_PARAMS and self._uses_param_mixed_precision ) # A handles key represents the group of `FlatParamHandle`s involved in a given # module's forward. These will be all-gathered together in the pre-forward and # pre-backward. _HandlesKey = Tuple[FlatParamHandle, ...]