# mypy: allow-untyped-defs import inspect import itertools from collections.abc import Sequence from dataclasses import dataclass, field from enum import auto, Enum from typing import Any, Callable, cast, Optional import torch import torch.nn as nn from torch._prims_common import make_contiguous_strides_for from torch.distributed._functional_collectives import AsyncCollectiveTensor from torch.distributed.tensor import DTensor, Replicate, Shard from torch.distributed.tensor._dtensor_spec import DTensorSpec, TensorMeta from torch.distributed.tensor.device_mesh import _mesh_resources from torch.distributed.tensor.placement_types import _StridedShard, Placement from ._fsdp_api import CPUOffloadPolicy, MixedPrecisionPolicy, OffloadPolicy from ._fsdp_common import ( _chunk_with_empty, _from_local_no_grad, _get_dim_chunked_size, _raise_assert_with_print, _to_dtype_if_needed, compiled_autograd_enabled, FSDPMeshInfo, HSDPMeshInfo, ) """ [Note: FSDP tensors] FSDP considers the following tensors: - Original parameter: parameter passed to :class:`FSDPParam`, i.e. the one on the module when applying FSDP - Sharded parameter: sharding the original parameter on dim-0 (or a user-specified dim) as a DTensor over the main mesh - All-gather inputs: the ``torch.Tensor`` or ``Tensor`` s passed to all-gather, derived from the sharded parameter - All-gather output: the ``torch.Tensor`` or ``Tensor`` s resulting from all-gathering the all-gather inputs - Unsharded parameter: parameter used for forward/backward computation, derived from the all-gather output; autograd leaf We define these tensors to describe the general framework that can accomodate extensions, where: - all-gather-inputs = pre-all-gather-transform(sharded-parameter) - unsharded-parameter = post-all-gather-transform(all-gather-outputs) For the default ``torch.Tensor`` case, there is only one all-gather input, and it shares the same underlying tensor data as the sharded parameter, meaning that they can be thought of as the same tensors. The same applies for the all-gather output and unsharded parameter. For non-``torch.Tensor`` extensions, these equivalences may no longer hold due to the pre/post-all-gather transforms, and some may have multiple all-gather inputs/outputs (e.g. quantized data and scales). [Note: FSDP and autograd] FSDP dynamically frees and allocates the unsharded parameter. Since autograd can pack a reference to it or a view to save for backward, we use storage resizing to implement the freeing/allocation since that preserves the aliasing. This implies that we construct the unsharded parameter object once and write to it in-place thereafter. For the default ``torch.Tensor` original parameter case, the all-gather output and unsharded parameter share the same data, so we use storage resizing on the all-gather output. """ lib = torch.library.Library("fsdp", "FRAGMENT") # noqa: TOR901 lib.define("copy_(Tensor(a!) tensor, Tensor data) -> ()") @torch.library.impl(lib, "copy_", "Meta") @torch.library.impl(lib, "copy_", "CUDA") @torch.library.impl(lib, "copy_", "XPU") @torch.library.impl(lib, "copy_", "HPU") @torch.library.impl(lib, "copy_", "CPU") @torch.library.impl(lib, "copy_", "MTIA") def copy_(tensor, data): tensor.copy_(data) """ [Note: Avoiding functionalization for fsdp.copy_ and inductor.resize_storage_bytes_] Currently we don't functionalize `fsdp.copy_` op or `inductor.resize_storage_bytes_` op (i.e. they show up as a mutation op in the middle of the AOT joint graph). Reason: Traceable FSDP2 compiled autograd BWD graph have the following traits: (1) Two inputs of the graph were aliased to each other (one from hook closed-over tensors, one from FWD saved tensors). (2) One of them is mutated (copy_ and resize_ to handle the all-gathered param). (3) They are both subclasses. The combination of these traits is not supported by AOTAutograd (it's difficult to reason about subclass aliasing). So this doesn't work at all for Traceable FSDP2. The compromise we use is to avoid functionalization for the FSDP2 copy_ and resize_ ops. This avoids the problem above, because from AOTAutograd point-of-view there are no mutations that functionalization needs to handle. (Although we need to be careful not to DCE those mutable ops.) We can avoid this functionalization because: (1) The nn.Parameter is never used before its .copy_() is called in eager code (i.e. no alias of it is created), so it's safe to call .copy_() in the middle of the graph to update its content and start using the nn.Parameter downstream. (2) We always re-allocate the buffer for nn.Parameter to store the AllGather output and to be used in downstream user ops. So calling resize-to-0 in the middle of the graph to free nn.Parameter memory after use should always be okay (since we always allocate anew next time we need it, we strictly don't need to keep the old tensor storage around anymore). Q: Wouldn't the extra resize_ and copy_ ops hurt both memory usage and performance? A: Yes it would. As an optimization, we have an Inductor post-grad FX pass to remove those resize_ and copy_ ops for unsharded params that have this pattern: resize_(full) -> copy_ -> resize_(0). TODO: Now that we are maintaining the invariant of "no aliased + mutated graph inputs" in both the forward and backward, it is now more feasible to functionalize all of the mutable FSDP ops. Some of the pros and cons are: Cons (of functionalizing those ops): (1) By not functionalizing them as we are today, we are making it more likely that they will run at the "correct" time in the generated code. If we start to functionalize them, we will need to make sure that Inductor reinplaces them in a way where it properly moves the mutations back to exactly where they should have run, or we risk suffering worse peak memory than eager. (We probably already need to do something similar in Inductor's reinplacing for copy_: https://github.com/pytorch/pytorch/issues/135305#issuecomment-2334888089) Pros (of functionalizing): (1) Better safety, we don't need to worry about the graph passes in inductor/partitioning handling input mutations mid-graph quite as much (to be fair we've already done some amount of auditing, but we might have to do some more). (2) Better perf: each mutation midway through the graph prevents Inductor from pattern matching across it. But maybe there are few enough mutations induced by FSDP for this to matter. """ @torch.library.impl(lib, "copy_", "Functionalize") def copy__functionalize(tensor, data): torch._sync(tensor) torch._sync(data) tensor_inner = torch._from_functional_tensor(tensor) data_inner = torch._from_functional_tensor(data) with torch._C._ExcludeDispatchKeyGuard( torch._C.DispatchKeySet(torch._C.DispatchKey.Functionalize) ): torch.ops.fsdp.copy_.default(tensor_inner, data_inner) if not torch._running_with_deploy(): torch.fx.node.has_side_effect(torch.ops.fsdp.copy_.default) class ShardedState(Enum): """ - ``SHARDED``: The sharded parameter is registered to the module. It is the only contributor to parameter memory. - ``SHARDED_POST_FORWARD``: The unsharded parameter is resharded to a smaller world size. Since this data should not be used for computation, we do not register it to the module. Users should reshard the module before any in-place modifications. Both it and the sharded parameter contribute to parameter memory. - ``UNSHARDED``: The unsharded parameter is registered to the module. Both it and the sharded parameter contribute to parameter memory. """ SHARDED = auto() SHARDED_POST_FORWARD = auto() UNSHARDED = auto() @dataclass class ParamModuleInfo: """ For a parameter, this stores the module and the parameter name to be able to do a parameter swap via ``setattr(module, param_name, ...)`` or to get the parameter via ``getattr(module, param_name)``. We additionally save shared modules and shared parameter names to update them accordingly. """ # Parameter names are unprefixed, e.g. "weight", not "lin.weight" module: nn.Module param_name: str shared_modules: list[nn.Module] = field(default_factory=list) shared_param_names: list[str] = field(default_factory=list) @dataclass class ExtensionsData: # User-defined metadata passed from pre to post-all-gather all_gather_metadata: Optional[Any] = None # Save the all-gather input sizes to unflatten the all-gather outputs to ND all_gather_input_sizes: Sequence[torch.Size] = () # ND def clear(self): self.all_gather_metadata = None self.all_gather_input_sizes = () class FSDPParam: """ This class manages a parameter with FSDP or FSDP variants applied, implementing dim-0 per-parameter sharding. """ orig_dtype: torch.dtype param_dtype: Optional[torch.dtype] reduce_dtype: Optional[torch.dtype] _orig_size: torch.Size # ND sharded_size: torch.Size # ND contiguous_sharded_stride: tuple[int, ...] padded_sharded_param_size: torch.Size # ND sharded_post_forward_size: torch.Size # ND contiguous_sharded_post_forward_stride: tuple[int, ...] _sharded_param_data: torch.Tensor # 1D sharded_param: nn.Parameter # ND _sharded_post_forward_param_data: Optional[torch.Tensor] # 1D _sharded_post_forward_param: Optional[nn.Parameter] # ND _unsharded_param: nn.Parameter # ND unsharded_accumulated_grad: Optional[torch.Tensor] # ND _sharding_spec: DTensorSpec # DTensor attributes (only defined for DTensor `param`): _tp_spec: DTensorSpec all_gather_outputs: list[torch.Tensor] # 1D # All-gather extension attributes _extensions_data: ExtensionsData _unsharded_inner_tensors: list[torch.Tensor] def __init__( self, param: nn.Parameter, module_info: ParamModuleInfo, mesh_info: FSDPMeshInfo, post_forward_mesh_info: Optional[FSDPMeshInfo], device: torch.device, shard_placement_fn: Optional[Callable[[nn.Parameter], Optional[Shard]]], mp_policy: MixedPrecisionPolicy, offload_policy: OffloadPolicy, ): self._module_info: ParamModuleInfo = module_info self.mesh_info = mesh_info self.post_forward_mesh_info = post_forward_mesh_info self.device = device self.mp_policy = mp_policy self.offload_to_cpu: bool = isinstance(offload_policy, CPUOffloadPolicy) self.pin_memory = ( self.offload_to_cpu and cast(CPUOffloadPolicy, offload_policy).pin_memory ) self.grad_offload_event: Optional[torch.Event] = None self._init_sharded_param(param, device, shard_placement_fn) if self.post_forward_mesh_info: self._init_sharded_post_forward_param_metadata(param) self._init_extensions() self.all_gather_outputs: list[torch.Tensor] = [] self.unsharded_accumulated_grad = None self._param_fqn: Optional[str] = None # prefixed from root module # TODO: Remove this padding logic once DTensor pads the local tensor: # https://github.com/pytorch/pytorch/issues/113045 self._post_load_hook_handle = ( module_info.module.register_load_state_dict_post_hook( lambda *args, **kwargs: self.reset_sharded_param() ) ) @torch.no_grad() def _init_sharded_param( self, param: nn.Parameter, device: torch.device, shard_placement_fn: Optional[Callable], ): if param.device != device and param.device.type != "meta": raise AssertionError( f"Expects the parameter to already be moved to device {device} but got {param.device}" ) if not param.is_contiguous(): raise NotImplementedError( f"FSDP does not support non-contiguous parameters yet: {param.shape=} {param.stride()=}" ) fsdp_placement = shard_placement_fn(param) if shard_placement_fn else None if fsdp_placement is None: fsdp_placement = Shard(0) elif fsdp_placement.dim < 0: fsdp_placement = Shard(fsdp_placement.dim + param.ndim) assert isinstance(fsdp_placement, Shard), f"{fsdp_placement}" self.fsdp_placement = fsdp_placement shard_dim = fsdp_placement.dim # TODO: Replace the sharded DTensor parameter construction logic with # `distribute_tensor` after https://github.com/pytorch/pytorch/issues/116101 # TODO: Simplify the following sharded parameter padding logic after # https://github.com/pytorch/pytorch/issues/113045 self.is_dtensor = isinstance(param, DTensor) if self.is_dtensor: self._tp_spec = cast(DTensor, param)._spec dp_mesh, tp_mesh = (self.mesh_info.mesh, self._tp_spec.mesh) dp_global_mesh = _mesh_resources.get_root_mesh(dp_mesh) tp_global_mesh = _mesh_resources.get_root_mesh(tp_mesh) if dp_global_mesh != tp_global_mesh or ( dp_global_mesh is None or tp_global_mesh is None ): raise AssertionError( "FSDP requires the DP and TP mesh to have the same parent mesh but got: \n" f"DP's global mesh: {dp_global_mesh}\nTP's global mesh: {tp_global_mesh}" ) name_dims_error = "FSDP requires named DeviceMesh dims for ND parallelism" assert dp_mesh.mesh_dim_names is not None, name_dims_error assert tp_mesh.mesh_dim_names is not None, name_dims_error submesh_names = dp_mesh.mesh_dim_names + tp_mesh.mesh_dim_names self._spmd_mesh = dp_global_mesh[submesh_names] if len(self._tp_spec.placements) != 1: raise NotImplementedError( f"FSDP only supports 1D TP, not {self._tp_spec.placements}" ) split_factor = self._tp_spec.num_shards_map[shard_dim] assert 2 <= self._spmd_mesh.ndim <= 3, ( f"_spmd_mesh.ndim can only be 2 or 3 but got {self._spmd_mesh.ndim}." ) self._spmd_placements: tuple[Placement, ...] dp_shard_tp_placement = ( ( _StridedShard(shard_dim, split_factor=split_factor) if split_factor > 1 else fsdp_placement ), self._tp_spec.placements[0], ) if self._spmd_mesh.ndim == 2: self._spmd_placements = dp_shard_tp_placement else: assert self.mesh_info.replicate_mesh_dim == 0 self._spmd_placements = (Replicate(),) + dp_shard_tp_placement self._sharding_spec = DTensorSpec( self._spmd_mesh, self._spmd_placements, tensor_meta=self._tp_spec.tensor_meta, ) # TODO: Enable uneven sharding for FSDP+TP. if split_factor > 1: # FSDP has strided sharding on tensor dim 0 num_shards = self._sharding_spec.num_shards_map[0] tensor_size_dim_0 = self._sharding_spec.shape[0] if tensor_size_dim_0 % num_shards != 0: raise NotImplementedError( "FSDP+TP sharding does not support uneven sharding for now: " f"tensor dim 0 has size {tensor_size_dim_0} which cannot be " f"evenly sharded into {num_shards} shards." ) param_data = cast(DTensor, param)._local_tensor else: self._spmd_mesh = self.mesh_info.mesh if isinstance(self.mesh_info, HSDPMeshInfo): self._spmd_placements = (Replicate(), fsdp_placement) else: self._spmd_placements = (fsdp_placement,) self._sharding_spec = DTensorSpec( self._spmd_mesh, self._spmd_placements, tensor_meta=TensorMeta(param.size(), param.stride(), param.dtype), ) param_data = param assert param_data.is_contiguous(), f"{param_data.shape=} {param_data.stride()=}" shard_dim = fsdp_placement.dim if shard_dim >= param_data.ndim: raise AssertionError( f"Shard dim {shard_dim} is invalid for {param_data.ndim}D tensor: {param.shape}" ) self._orig_size = param_data.size() self._contiguous_orig_stride = make_contiguous_strides_for(self._orig_size) shard_rank = self.mesh_info.shard_mesh_rank shard_world_size = self.mesh_info.shard_mesh_size if shard_dim > 0 and param_data.size(shard_dim) % shard_world_size != 0: # If sharding on nonzero dim, require even sharding for now because # the uneven sharding (1) requires extra copies before/after FSDP # collectives and (2) introduces extra complexity to handle padding # and unpadding raise NotImplementedError( f"FSDP does not support uneven sharding on dim {shard_dim}: " f"{param_data.size()} (world size: {shard_world_size})" ) chunks = _chunk_with_empty(param_data, shard_world_size, dim=shard_dim) sharded_param = chunks[shard_rank] self.sharded_size = _get_dim_chunked_size( sharded_param, param_data.size(), dim=shard_dim ) self.contiguous_sharded_stride = make_contiguous_strides_for(self.sharded_size) padded_sharded_size = chunks[0].size() # 0th always padded self.padded_sharded_param_size = padded_sharded_size # Pre-pad the sharded parameter to avoid padding before all-gather padded_sharded_param = param_data.new_zeros(padded_sharded_size) if sharded_param.numel() > 0: padded_sharded_param.narrow( dim=shard_dim, start=0, length=sharded_param.size(shard_dim) ).copy_(sharded_param) if self.offload_to_cpu and not padded_sharded_param.is_meta: padded_sharded_param = padded_sharded_param.cpu() if self.pin_memory: padded_sharded_param = padded_sharded_param.pin_memory( device=self.device ) self._sharded_param_data = padded_sharded_param.view(-1) length = sharded_param.size(shard_dim) if sharded_param.numel() > 0 else 0 sharded_param = padded_sharded_param.narrow( dim=shard_dim, start=0, length=length ) assert sharded_param.is_contiguous(), f"{self.fsdp_placement=}" self.sharded_param = nn.Parameter(self.to_sharded_dtensor(sharded_param)) self.sharded_param.requires_grad_(param.requires_grad) # Let `param_data` be freed normally when its ref count reaches 0 when # the `fully_shard` call returns to allow provided parameters to alias self._setattr_on_modules(self.sharded_param) self.sharded_state = ShardedState.SHARDED def _init_sharded_post_forward_param_metadata(self, param: torch.Tensor) -> None: mesh_info = self.post_forward_mesh_info assert mesh_info is not None # mypy param_data = param._local_tensor if isinstance(param, DTensor) else param chunks = _chunk_with_empty(param_data, mesh_info.shard_mesh_size, dim=0) self.sharded_post_forward_size = _get_dim_chunked_size( chunks[mesh_info.shard_mesh_rank], param_data.size(), dim=self.fsdp_placement.dim, ) self.contiguous_sharded_post_forward_stride = make_contiguous_strides_for( self.sharded_post_forward_size ) def init_dtype_attrs(self, mp_policy: MixedPrecisionPolicy): param_dtype, reduce_dtype = (mp_policy.param_dtype, mp_policy.reduce_dtype) self.orig_dtype = self.sharded_param.dtype # Clamp `reduce_dtype` to `None` if no casting is required: since # gradients are computed in `param_dtype`, if `reduce_dtype` matches, # then we do not need extra casting if reduce_dtype == param_dtype: reduce_dtype = None # Clamp `param_dtype` to `None` if no casting is required if param_dtype == self.orig_dtype: param_dtype = None self.param_dtype = param_dtype self.reduce_dtype = reduce_dtype # None indicates that the mixed precision is not enabled def _init_extensions(self) -> None: inner_tensor = self._sharded_local_tensor has_fsdp_pre_all_gather = hasattr(inner_tensor, "fsdp_pre_all_gather") has_fsdp_post_all_gather = hasattr(inner_tensor, "fsdp_post_all_gather") if has_fsdp_pre_all_gather != has_fsdp_post_all_gather: raise AssertionError( "Both fsdp_pre_all_gather and fsdp_post_all_gather should be defined " f"if using all-gather extensions: {inner_tensor}" ) if has_fsdp_pre_all_gather: self._extensions_data = ExtensionsData() self._unsharded_inner_tensors: list[torch.Tensor] = [] def init_all_gather_outputs( self, all_gather_input_numels: list[int], all_gather_input_dtypes: list[torch.dtype], world_size: int, device: torch.device, force_recreate: bool = False, ): if not force_recreate and len(self.all_gather_outputs) > 0: return # already initialized self.all_gather_outputs = [ torch.empty(torch.Size([numel * world_size]), dtype=dtype, device=device) for numel, dtype in zip(all_gather_input_numels, all_gather_input_dtypes) ] def init_unsharded_param(self): """ [Note: Invariants for torch.compile Traceable FSDP2] 1. Under compile, we always re-populate the content of `self._unsharded_param` per AllGather using the slow path. 2. Under compile, we always recreate `self.all_gather_outputs` per AllGather. This is to ensure the buffer creation is internal to the graph and avoid `self.all_gather_outputs` being captured as a graph input. 3. Under compile, at the end of `free_unsharded_param()`, we always clean up `self.all_gather_outputs` and `self._unsharded_inner_tensors`, to avoid them being captured as graph output. With these invariants, only these tensors will be inputs to the graph: - Sharded parameters - Placeholders for the `self._unsharded_param` nn.Parameter """ if not compiled_autograd_enabled() and hasattr( self, "_unsharded_param" ): # after the 1st all-gather inner_tensor = self._sharded_local_tensor if not hasattr(inner_tensor, "fsdp_post_all_gather"): return # already initialized for tensor in self._unsharded_inner_tensors: alloc_storage(tensor) all_gather_outputs = self._unflatten_all_gather_outputs() inner_tensor.fsdp_post_all_gather( all_gather_outputs, self._extensions_data.all_gather_metadata, self.param_dtype or self.orig_dtype, out=self._unsharded_param, ) self._extensions_data.clear() return inner_tensor = self._sharded_local_tensor if not compiled_autograd_enabled() and hasattr( inner_tensor, "fsdp_post_all_gather" ): all_gather_outputs = self._unflatten_all_gather_outputs() ( unsharded_tensor, self._unsharded_inner_tensors, ) = inner_tensor.fsdp_post_all_gather( all_gather_outputs, self._extensions_data.all_gather_metadata, self.param_dtype or self.orig_dtype, ) self._extensions_data.clear() else: # For the default path (no post-all-gather), the all-gather output # gives the unsharded parameter data directly assert len(self.all_gather_outputs) == 1, f"{len(self.all_gather_outputs)}" unsharded_tensor = self.all_gather_outputs[0] unsharded_param = torch.as_strided( unsharded_tensor, self._orig_size, self._contiguous_orig_stride, storage_offset=0, ) if self.is_dtensor: unsharded_param = _from_local_no_grad(unsharded_param, self._tp_spec) if hasattr(self, "_unsharded_param"): assert compiled_autograd_enabled() with ( torch.no_grad(), torch.autograd._unsafe_preserve_version_counter(self._unsharded_param), ): # NOTE: Under compile, if an unsharded param goes through # resize_(full) -> copy_ -> resize_(0) pattern, we will remove those # resize_ and copy_ ops in a compiler graph pass # `remove_fsdp2_unsharded_param_graph_input_usage` to recover performance. self._unsharded_param.untyped_storage().resize_( self._unsharded_param.numel() * self._unsharded_param.itemsize ) torch.ops.fsdp.copy_(self._unsharded_param, unsharded_param) else: self._unsharded_param = nn.Parameter( unsharded_param, requires_grad=self.sharded_param.requires_grad ) def _unflatten_all_gather_outputs(self) -> tuple[torch.Tensor, ...]: return tuple( t.view(-1, *s[1:]) for t, s in zip( self.all_gather_outputs, self._extensions_data.all_gather_input_sizes ) ) def to_sharded(self) -> None: self._setattr_on_modules(self.sharded_param) self.free_unsharded_param() self.sharded_state = ShardedState.SHARDED def to_sharded_post_forward(self) -> None: if self.is_dtensor: raise NotImplementedError( "Resharding to smaller mesh with TP is not supported yet" ) self._assert_in_states(ShardedState.UNSHARDED) assert self.post_forward_mesh_info is not None # mypy assert len(self.all_gather_outputs) == 1 shard_world_size = self.post_forward_mesh_info.shard_mesh_size if (numel := self.all_gather_outputs[0].numel()) % shard_world_size != 0: _raise_assert_with_print( f"All-gather output size ({numel}) must be divisible by the shard " f"world size ({shard_world_size})" ) shard_rank = self.post_forward_mesh_info.shard_mesh_rank sharded_numel = numel // shard_world_size self._sharded_post_forward_param_data = ( self.all_gather_outputs[0].narrow( 0, sharded_numel * shard_rank, sharded_numel ) ).clone() # clone to be able to free all-gather output sharded_post_forward_tensor = torch.as_strided( self._sharded_post_forward_param_data, size=self.sharded_post_forward_size, stride=self.contiguous_sharded_post_forward_stride, storage_offset=0, ) self._sharded_post_forward_param = nn.Parameter( self.to_sharded_post_forward_dtensor(sharded_post_forward_tensor) ) self._setattr_on_modules(self._sharded_post_forward_param) self.free_unsharded_param() self.sharded_state = ShardedState.SHARDED_POST_FORWARD def to_unsharded(self) -> None: # Assume that the data has been allocated and all-gathered set_requires_grad_if_needed(self.sharded_param, self._unsharded_param) self._setattr_on_modules(self._unsharded_param) if self.sharded_state == ShardedState.SHARDED_POST_FORWARD: # The data is allocated in the default stream via the post-forward # reshard and must be kept alive for the next all-gather copy-in. # Since we call this method after the copy-out, the data's lifetime # is ensured without further synchronization. self._sharded_post_forward_param = None self._sharded_post_forward_param_data = None # free self.sharded_state = ShardedState.UNSHARDED def _setattr_on_modules(self, param: nn.Parameter) -> None: unsafe_setattr_param( self._module_info.module, self._module_info.param_name, param ) for shared_module, shared_param_name in zip( self._module_info.shared_modules, self._module_info.shared_param_names ): unsafe_setattr_param(shared_module, shared_param_name, param) def to_sharded_dtensor(self, tensor: torch.Tensor) -> DTensor: """ Converts a local tensor representing either the sharded parameter or sharded gradient to DTensor. """ if tensor.shape != self.sharded_size: _raise_assert_with_print( f"Expects size {self.sharded_size} but got {tensor.shape}" ) return _from_local_no_grad( tensor, self._sharding_spec, ) def to_sharded_post_forward_dtensor(self, tensor: torch.Tensor) -> DTensor: if tensor.shape != self.sharded_post_forward_size: _raise_assert_with_print( f"Expects size {self.sharded_post_forward_size} but got {tensor.shape}" ) assert isinstance(self.post_forward_mesh_info, HSDPMeshInfo) # TODO: Prefer this DTensor to be read-only and generalize the # placement once we support TP. post_forward_sharding_spec = DTensorSpec( self.post_forward_mesh_info.mesh, (Replicate(), Shard(0)), tensor_meta=self._sharding_spec.tensor_meta, ) return _from_local_no_grad(tensor, post_forward_sharding_spec) def to_accumulated_grad_if_needed(self) -> None: # Access `_unsharded_param` to bypass the sharded state check since we # prefer to reshard before upcasting the gradient to save memory if ( self.reduce_dtype is None or self._unsharded_param.grad is None or self._unsharded_param.grad.dtype == self.reduce_dtype ): return unsharded_grad = self._unsharded_param.grad self._unsharded_param.grad = None self.unsharded_accumulated_grad = unsharded_grad.to(self.reduce_dtype) def accumulate_unsharded_grad_if_needed(self) -> None: if ( self.unsharded_accumulated_grad is not None and self.unsharded_param.grad is not None ): self.unsharded_accumulated_grad += self.unsharded_param.grad self.unsharded_param.grad = None def alloc_all_gather_outputs(self) -> None: for tensor in self.all_gather_outputs: alloc_storage(tensor) def free_unsharded_param(self) -> None: if compiled_autograd_enabled(): """ Assumptions under compile: - `self._unsharded_param` is NOT an alias of `self.all_gather_outputs`. Instead, we resize `self._unsharded_param` storage size to full and then explicitly *copy* the data from `self.all_gather_outputs` to `self._unsharded_param` in `init_unsharded_param()`. (For full-graph FSDP2 case, we will then remove the resize_ and copy_ ops in a compiler graph pass to recover performance.) - `self.all_gather_outputs` and `self._unsharded_inner_tensors` are NOT graph inputs. They are created within the graph and is guaranteed to be freed by the end of the graph. They don't leak outside of the graph. """ self._unsharded_param.untyped_storage().resize_(0) self.all_gather_outputs = [] self._unsharded_inner_tensors = [] else: for tensor in itertools.chain( self.all_gather_outputs, self._unsharded_inner_tensors ): free_storage(tensor) @property def all_gather_inputs(self) -> list[torch.Tensor]: # 1D self._assert_in_states(ShardedState.SHARDED, ShardedState.SHARDED_POST_FORWARD) if self.sharded_state == ShardedState.SHARDED: if not compiled_autograd_enabled() and hasattr( self._sharded_local_tensor, "fsdp_pre_all_gather" ): sharded_local_tensor = self._sharded_local_tensor if self.offload_to_cpu: sharded_local_tensor = sharded_local_tensor.to( self.device, non_blocking=True ) pre_all_gather_signature = inspect.signature( sharded_local_tensor.fsdp_pre_all_gather ) num_fn_params = len(pre_all_gather_signature.parameters) # Old signature only passes mesh; keep for BC for now assert num_fn_params in ( 1, 5, ), ( f"Invalid fsdp_pre_all_gather: {pre_all_gather_signature}\n" "Expects fsdp_pre_all_gather(self, mesh: DeviceMesh, " "module: nn.Module, mp_policy: MixedPrecisionPolicy)" ) if num_fn_params == 1: ( all_gather_inputs, self._extensions_data.all_gather_metadata, ) = sharded_local_tensor.fsdp_pre_all_gather( self.shard_mesh_from_root ) else: ( all_gather_inputs, self._extensions_data.all_gather_metadata, ) = sharded_local_tensor.fsdp_pre_all_gather( self.shard_mesh_from_root, self._orig_size, self._contiguous_orig_stride, self._module_info.module, self.mp_policy, ) if ( sharded_local_tensor.size() != self.padded_sharded_param_size and any( all_gather_input.size() != self.padded_sharded_param_size for all_gather_input in all_gather_inputs ) ): # NOTE: Since this error can only be raised on the # ranks that have padding, this can manifest as a NCCL # watchdog timeout, as the other ranks will not error. raise AssertionError( "When a parameter is unevenly sharded by FSDP " f"(orig size={self._orig_size}, FSDP world size={self.mesh_info.mesh.size()}), " "fsdp_pre_all_gather must return all-gather inputs with the padded sharded size " f"{self.padded_sharded_param_size} but got {[t.size() for t in all_gather_inputs]}" ) self._extensions_data.all_gather_input_sizes = [ t.size() for t in all_gather_inputs ] return [t.view(-1) for t in all_gather_inputs] sharded_param_data = self._sharded_param_data if self.offload_to_cpu: sharded_param_data = sharded_param_data.to( self.device, non_blocking=True ) return [_to_dtype_if_needed(sharded_param_data, self.param_dtype)] elif self.sharded_state == ShardedState.SHARDED_POST_FORWARD: if not compiled_autograd_enabled() and hasattr( self._sharded_local_tensor, "fsdp_pre_all_gather" ): raise NotImplementedError all_gather_input = _to_dtype_if_needed( cast(torch.Tensor, self._sharded_post_forward_param_data), self.param_dtype, ) return [all_gather_input] return [torch.empty(0)] # mypy @property def unsharded_param(self) -> nn.Parameter: # ND return self._unsharded_param @property def unsharded_grad_data(self) -> torch.Tensor: grad = self.unsharded_param.grad assert grad is not None, "Expects unsharded_param.grad to not be None" return self._get_grad_inner_tensor(grad) @property def unsharded_accumulated_grad_data(self) -> torch.Tensor: grad = self.unsharded_accumulated_grad assert grad is not None, "Expects unsharded_accumulated_grad to not be None" return self._get_grad_inner_tensor(grad) def _get_grad_inner_tensor(self, grad: torch.Tensor) -> torch.Tensor: if self.is_dtensor: if isinstance(grad, AsyncCollectiveTensor): grad = grad.wait() assert isinstance(grad, DTensor), f"{type(grad)}" placements = self._tp_spec.placements if placements != grad.placements: assert len(self._tp_spec.placements) == len(grad.placements), ( f"{self._tp_spec=} {grad.placements=}" ) grad = grad.redistribute(placements=placements) grad = grad._local_tensor return grad @property def _sharded_local_tensor(self) -> torch.Tensor: return cast(DTensor, self.sharded_param)._local_tensor @property def shard_mesh(self): mesh = self.mesh_info.mesh if mesh.ndim == 1: return mesh elif mesh.ndim == 2: assert mesh.mesh_dim_names is not None return mesh[mesh.mesh_dim_names[-1]] raise ValueError(f"Invalid mesh: {mesh}") @property def shard_mesh_from_root(self): mesh = self.mesh_info.mesh if mesh.ndim == 1: return mesh else: assert mesh.mesh_dim_names is not None shard_dim_name = mesh.mesh_dim_names[-1] root_mesh = _mesh_resources.get_root_mesh(mesh) return root_mesh[shard_dim_name] def _assert_in_states(self, *states: ShardedState) -> None: if self.sharded_state not in states: _raise_assert_with_print( f"Expects to be in one of {states}, not {self.sharded_state}" ) def reset_sharded_param(self): # For ops like `nn.Module._apply` or `load_state_dict(assign=True)` # that change the sharded parameter tensor, we may need to re-pad the # sharded local tensor and re-save the reference. module_info = self._module_info new_param = getattr(module_info.module, module_info.param_name) if new_param is not self.sharded_param: if torch.__future__.get_swap_module_params_on_conversion(): raise AssertionError( f"Expects swap_tensors to preserve object but got {new_param} " f"instead of {self.sharded_param}" ) self.sharded_param = new_param local_tensor = new_param._local_tensor if local_tensor.is_meta: return updated_local_tensor = False padded_sharded_size = self.padded_sharded_param_size shard_dim = self.fsdp_placement.dim length = local_tensor.size(shard_dim) if local_tensor.numel() > 0 else 0 if local_tensor.size() != padded_sharded_size: assert shard_dim == 0, ( f"Shard({shard_dim}) requires even sharding: {local_tensor.size()=}" ) padded_local_tensor = local_tensor.new_zeros(padded_sharded_size) padded_local_tensor.narrow(dim=shard_dim, start=0, length=length).copy_( local_tensor ) local_tensor = padded_local_tensor updated_local_tensor = True if self.pin_memory and not local_tensor.is_pinned(): local_tensor = local_tensor.cpu().pin_memory(device=self.device) updated_local_tensor = True self._sharded_param_data = local_tensor.view(-1) assert isinstance(self.sharded_param, DTensor) # mypy if updated_local_tensor: # Only change the local tensor object if needed self.sharded_param._local_tensor = local_tensor.narrow( dim=shard_dim, start=0, length=length ) assert self.sharded_param._local_tensor.is_contiguous() self._sharding_spec = self.sharded_param._spec def __repr__(self): return f"FSDPParam(fqn={self._param_fqn}, orig_size={self._orig_size})" def alloc_storage(tensor: torch.Tensor) -> None: size = tensor.numel() * tensor.itemsize if (storage := tensor.untyped_storage()).size() != size: storage.resize_(size) def free_storage(tensor: torch.Tensor) -> None: if (storage := tensor.untyped_storage()).size() != 0: storage.resize_(0) # NOTE: These bypass `nn.Module.__setattr__` checks, which incur non-trivial # CPU overhead, if the module did not override it. For FSDP, we know we do not # need those checks when transitioning between sharded/unsharded parameters. def unsafe_setattr_param( module: nn.Module, param_name: str, param: nn.Parameter ) -> None: if getattr(module.__setattr__, "__func__", None) is nn.Module.__setattr__: module._parameters[param_name] = param else: # slow path setattr(module, param_name, param) def set_requires_grad_if_needed( src_tensor: torch.Tensor, dst_tensor: torch.Tensor ) -> None: # Only call `requires_grad_` if needed to avoid the Python <> C++ context # switch overhead if src_tensor.requires_grad != dst_tensor.requires_grad: dst_tensor.requires_grad_(src_tensor.requires_grad)