Source code for torch.distributed.fsdp.fully_sharded_data_parallel

import functools
import traceback
from dataclasses import dataclass
from enum import Enum, auto
from typing import (

import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.distributed import ProcessGroup
from torch.distributed.distributed_c10d import _get_default_group
from torch.nn.parameter import Parameter

from .flatten_params_wrapper import FlattenParamsWrapper
from .wrap import _recursive_wrap

from .utils import (

    from collections import OrderedDict  # noqa: F401

class CPUOffload:
    CPU offlaoding config. Currently, only parameter and gradient CPU
    offload are supported.
    offload_params: Offloading parameters to CPUs when these parameters are
                    not used for computation on GPUs. This implicitly enables
                    gradient offloading to CPUs in order for parameters and
                    gradients to be on the same device to work with optimizer.
    offload_params: bool = False
    # TODO: state dict offloading

class BackwardPrefetch(Enum):
    Specify where to prefetch next layer's full parameters
    during backward pass.
    BACKWARD_PRE: prefetch right before current layer's backward computation
                  starts, this approach will increase backward communication
                  and computation overalpping and potentialy improve training
                  performance, but it may increase the peak memory usage as
                  the prefetched full parameters will be kept in the GPU memory
                  until next layer's backward computation is done.
    BACKWARD_POST: prefetch right after current layer's backward computation finishes,
                   this approach will not increase peak memory as prefetching happens
                   after current layer's full parameters are freed.
                   It could potentially improve backward communication and computation
                   overlapping as it avoids all_gather and reduce_scatter are blocked
                   each other in the single NCCL stream. However, based on our experiments,
                   for some models, the backward post backward hook fire order is not always
                   the reversed forward computation order, so this
                   approach may prefetch full parameters for layers ahead of next layer,
                   this 'ahead' all_gather could delay next layer's all_gather in the
                   single NCCL stream and cause the next layer's computation delay. So it may
                   cause some performance regession for some models.
    BACKWARD_PRE = auto()
    BACKWARD_POST = auto()
    # TODO, BACKWARD_PRE_CPU, prefetch full parameters and keep them in the CPU memory

class TrainingState_(Enum):
    Simple enum to indicate what state FSDP is in. Used for asserting
    to make sure APIs are called in the correct state.
        ``BACKWARD_PRE`` and ``BACKWARD_POST`` states are used to ensure we
        receives backward hooks in the correct order. It is used to catch
        unexpected order of hooks being called (likely due to our
        hook registration logic or autograd engine logic changes).

    IDLE = auto()
    FORWARD = auto()
    BACKWARD_PRE = auto()
    BACKWARD_POST = auto()

[docs]class FullyShardedDataParallel(nn.Module): """ A wrapper for sharding Module parameters across data parallel workers. This is inspired by `Xu et al.`_ as well as the ZeRO Stage 3 from DeepSpeed_. FullyShardedDataParallel is commonly shorten to FSDP. .. _`Xu et al.`: .. _DeepSpeed: Example:: >>> import torch >>> from torch.distributed.fsdp import FullyShardedDataParallel as FSDP >>> torch.cuda.set_device(device_id) >>> sharded_module = FSDP(my_module) >>> optim = torch.optim.Adam(sharded_module.parameters(), lr=0.0001) >>> x = sharded_module(x, y=3, z=torch.Tensor([1])) >>> loss = x.sum() >>> loss.backward() >>> optim.step() .. warning:: The optimizer must be initialized *after* the module has been wrapped, since FSDP will shard parameters in-place and this will break any previously initialized optimizers. .. warning: Module should be already placed on the destination device or device is set properly using torch.cuda.set_device(device_id). FSDP will get compute device from module first, if module device is CPU, FSDP will then get compute device from current device. Args: module (nn.Module): module to be wrapped with FSDP. process_group (Optional[ProcessGroup]): process group for sharding cpu_offload (Optional [CPUOffload]): CPU offloading config. Currently, only parameter and gradient CPU offload is supported. It can be enabled via passing in ``cpu_offload=CPUOffload(offload_params=True)``. Note that this currently implicitly enables gradient offloading to CPU in order for params and grads to be on same device to work with optimizer. This API is subject to change. Default is ``None`` in which case there will be no offloading. fsdp_auto_wrap_policy: (Optional [callable]): A callable specifying a policy to recursively wrap layers with FSDP. Note that this policy currently will only apply to child modules of the passed in module. The remainder modules are always wrapped in the returned FSDP root instance. ``default_auto_wrap_policy`` written in ``torch.distributed.fsdp.wrap`` is an example of ``fsdp_auto_wrap_policy`` callable, this policy wraps layers with parameter sizes larger than 100M. Users can supply the customized ``fsdp_auto_wrap_policy`` callable that should accept following arguments: ``module: nn.Module``, ``recurse: bool``, ``unwrapped_params: int``, extra customized arguments could be added to the customized ``fsdp_auto_wrap_policy`` callable as well. Example:: >>> def custom_auto_wrap_policy( >>> module: nn.Module, >>> recurse: bool, >>> unwrapped_params: int, >>> # These are customizable for this policy function. >>> min_num_params: int = int(1e8), >>> ) -> bool: >>> return unwrapped_params >= min_num_params backward_prefetch: (Optional[BackwardPrefetch]): This is an experimental feature that is subject to change in the the near future. It allows users to enable two different backward_prefetch algorithms to help backward communication and computation overlapping. Pros and cons of each algorithm is explained in the class ``BackwardPrefetch``. """ def __init__( self, module: nn.Module, process_group: Optional[ProcessGroup] = None, cpu_offload: Optional[CPUOffload] = None, fsdp_auto_wrap_policy: Optional[Callable] = None, backward_prefetch: Optional[BackwardPrefetch] = None, ): torch._C._log_api_usage_once("torch.distributed.fsdp") super().__init__() # if fsdp_auto_wrap_policy is specified, submodules should not be # already wrapped, otherwise we'd attempt to double wrap them resulting # in errors. if fsdp_auto_wrap_policy is not None: self._check_wrapped( module, check_fn=lambda mod: not isinstance(mod, FullyShardedDataParallel), err_fn=lambda mod: f"Expected {mod} to NOT be FullyShardedDataParallel if auto_wrap is enabled.", ) _recursive_wrap( module, auto_wrap_policy=fsdp_auto_wrap_policy, wrapper_cls=FullyShardedDataParallel, # Note that we have the recursive_wrap skip wrapping for # the outermost (this) module otherwise it will result in a # double-wrap causing issues. only_wrap_children=True, # FSDP arguments follow. process_group=process_group, cpu_offload=cpu_offload, backward_prefetch=backward_prefetch, # Note that recursive_wap should not call FSDP with wrapping # enabled, as this recursive call handles all wrapping, # including for nested children. fsdp_auto_wrap_policy=None, ) self.process_group = process_group or _get_default_group() self.rank = self.process_group.rank() self.world_size = self.process_group.size() # device for computation, if module is on GPU, use module.device; # if module is on CPU, use current device; self.compute_device = _get_default_cuda_device(module) # Free full params and keep shard only after forward self.reshard_after_forward = True # setting two factors to avoid underflow and overflow self.gradient_predivide_factor: float = self._get_gradient_predivide_factor( self.world_size ) self.gradient_postdivide_factor: float = ( self.world_size / self.gradient_predivide_factor ) self.numel_padded_per_param: List[int] = [] self.cpu_offload = cpu_offload or CPUOffload() self.backward_prefetch = backward_prefetch # Only handle params which are not already sharded. This enables # sharding individual layers of a Module, with an outer wrapper to # shard any leftover parameters. params = [] for param_name, param in module.named_parameters(): if not hasattr(param, "_is_sharded"): params.append(param) self._fsdp_wrapped_module: nn.Module = FlattenParamsWrapper( module, param_list=params ) del module # free original module in case it helps garbage collection if self._fsdp_wrapped_module.flat_param is not None: self.params = [self._fsdp_wrapped_module.flat_param] else: self.params = [] # Shard module parameters in place self._shard_parameters() # Make sure all parameters are sharded. for n, p in self.named_parameters(): if not hasattr(p, "_is_sharded"): raise RuntimeError(f"found unsharded parameter: {n} ; {p.size()}") self._reset_lazy_init() # Enum to indicate if we're in the forward/backward pass, idle, etc. self.training_state = TrainingState_.IDLE # Flag to guard against preparing gradients multiple times per backward pass. self._pre_backward_hook_has_run = False # Used for prefetching all gather full params in post backward hook self._need_rebuild_full_params = False # If specified, offload parameter shard to CPU. if self.cpu_offload.offload_params: for p in self.params: self._offload_to_cpu(p) @classmethod def _check_wrapped(cls, begin_module, check_fn, err_fn): for _, mod in begin_module.named_modules(): if not check_fn(mod): raise ValueError(err_fn(mod)) @property def module(self) -> FlattenParamsWrapper: """make model.module accessible, just like DDP.""" assert isinstance(self._fsdp_wrapped_module, FlattenParamsWrapper) return self._fsdp_wrapped_module # setting two factors 'self.gradient_predivide_factor' # and 'self.gradient_postdivide_factor' to avoid underflow and overflow def _get_gradient_predivide_factor(self, world_size: int) -> float: factor: int = 1 while world_size % factor == 0 and world_size / factor > factor: factor *= 2 return float(factor) def _offload_to_cpu(self, p): """ Offloads parameter to CPU from self.compute_device. If the parameter is already on CPU then this is a noop. """ cpu_device = torch.device("cpu") if p.device == cpu_device: return with torch.no_grad(): = def _cast_buffers( self, device: Optional[torch.device] = None, memo: Optional[Set] = None ) -> None: """Move all buffers to the given *device*. If *device* is not given, then it will default to ``self.compute_device``. In the case of nested FSDP instances, we will respect the child instance's ``compute_device`` configuration. Args: device (torch.device, Optional): device to cast buffers to (defaults to compute_device) memo (Set, Optional): set of modules that have already been processed """ if memo is None: memo = set() for module in self.modules(): if module is not self and isinstance(module, FullyShardedDataParallel): # Allow any child FSDP instances to handle their own buffers. module._cast_buffers(device=device, memo=memo) elif module not in memo: memo.add(module) for name, buf in module.named_buffers(recurse=False): if buf is None: continue buf = or self.compute_device) setattr(module, name, buf) @torch.no_grad() def _shard_parameters(self) -> None: """ At initialization we wrap a module with full parameters and shard the parameters in-place. Sharding is implemented by viewing each parameter as a 1D Tensor and retaining only a single slice, where the slice size is determined by the number of data parallel workers. After this initial sharding is complete, the user can initialize a ``torch.optim.Optimizer`` in the usual way, i.e.:: .. code-block:: python optim = torch.optim.Adam(sharded_module.parameters(), lr=0.0001) The optimizer will see only a single slice of parameters and will thus allocate less memory for optimizer state, avoiding redundancy across data parallel workers. """ self.numel_padded_per_param = [] for p in self.params: assert not hasattr( p, "_is_sharded" ), "Param should have not been sharded yet." assert ( p.is_floating_point() ), "Autograd does not support operations for integer type." # Sharding is done only when world_size is larger than 1. p._is_sharded = self.world_size > 1 # type: ignore[attr-defined] p._orig_size = p.size() # type: ignore[attr-defined] if not p._is_sharded: # type: ignore[attr-defined] self.numel_padded_per_param.append(0) continue # Save the original storage and free it later on. # Since we're modifying the tensor's storage directly, # make sure the tensor is the sole occupant of the storage. assert ( p.storage_offset() == 0 ), "The tensor is not the sole occupant of the storage." orig_storage = # Replace p with the relevant shard. local_shard, num_padded = self._get_shard(p) p.set_(local_shard) # type: ignore[call-overload] self.numel_padded_per_param.append(num_padded) # Free storage that contains the original full data. if orig_storage.size() > 0: orig_storage.resize_(0) # type: ignore[attr-defined] assert len(self.numel_padded_per_param) == len( self.params ), "numel_padded_per_param is not populated correctly." def _get_shard(self, tensor: torch.Tensor) -> Tuple[torch.Tensor, int]: """Return the local shard of a full tensor.""" # Shard using torch.chunk to match all-gather/reduce-scatter. chunks = torch.flatten(tensor).chunk(self.world_size) if len(chunks) < (self.rank + 1): # If there are not enough chunks to shard across ranks, create an # empty chunk that will just be padded with zeros to be the # appropriate size. chunk = chunks[0].new_empty(0) else: chunk = chunks[self.rank] # Determine number of padding elements. num_to_pad = chunks[0].numel() - chunk.numel() assert ( num_to_pad >= 0 ), "Chunk's size should be equal or smaller than \ the first chunk's size." # We always need to clone here, because regardless of padding the # original parameter, of which this chunk is a view of, is deallocated # after _get_shard. shard = chunk.clone() if num_to_pad > 0: shard = F.pad(shard, [0, num_to_pad]) return shard, num_to_pad def __getattr__(self, name: str) -> Any: """Forward missing attributes to wrapped module.""" try: return super().__getattr__(name) # defer to nn.Module's logic except AttributeError: return getattr(self.module, name) def __getitem__(self, key: int) -> Any: """Forward indexing calls in case the module is a nn.Sequential.""" return self.module.__getitem__(key) # type: ignore[operator] def _reset_lazy_init(self) -> None: """ Reset instance so :func:`_lazy_init` will run on the next forward. Currently this is only called in __init__ """ self._is_root: Optional[bool] = None self._streams: Dict[str, torch.cuda.Stream] = {} self._fsdp_graph_order: List[nn.Module] = [] self._my_fsdp_idx_in_graph: Optional[int] = None for p in self.params: if hasattr(p, "_local_shard"): # reset attributes that are added in _init_param_attributes, as # part of _lazy_init del p._local_shard # type: ignore[attr-defined] def _lazy_init(self) -> None: """Initialization steps that should happen lazily, typically right before the first forward pass. """ # Initialize param attributes lazily, in case the param's dtype or # device changes after __init__. for p in self.params: self._init_param_attributes(p) # Initialize _is_root and setup streams. These steps would ideally # happen in __init__, but _is_root can only be determined after the # entire model hierarchy is setup, thus we run it lazily. if self._is_root is None: # _is_root means that we are in the outermost module's forward. self._set_is_root() self._setup_streams() if self._is_root: # Buffers stay on GPU, and don't get sharded. Since _cast_buffers # applies recursively, we only call this from the root instance. self._cast_buffers() # Don't free the full params for the outer-most (root) instance, # In most cases, root instance contains params in the last layers # or has no params. In these cases, those params will be needed # immediately after for the backward pass. Note that this only # applies currently when freeing parameters at end of layer's # forward pass. self.reshard_after_forward = False # Due to the use of streams, we need to make sure the previous # ``optim.step()`` is done before we all-gather parameters. self._wait_for_previous_optim_step() @torch.no_grad() def _init_param_attributes(self, p: Parameter) -> None: """ We manage several attributes on each Parameter instance. The first two are set by :func:`_shard_parameters`: ``_is_sharded``: ``True`` if the Parameter is sharded or ``False`` if the Parameter is intentionally not sharded (in which case we will all-reduce grads for this param). Currently the only way `_is_sharded = False` is if world_size = 1. ``_orig_size``: the size of the original Parameter (before sharding) A few attributes are set here: ``_local_shard``: a single shard of the parameter. This is needed to recover the shard after rebuilding full parameter in forward and backward. ``_full_param_padded``: the full weight (padded to be evenly divisible by ``world_size``), used for computation in the forward and backward pass. It is initialized with the appropriate size and then has its storage freed. This will be resized in place and only materialized (via all-gather) as needed. Another attribute is set by :func:`_register_post_backward_hooks`: ``_shard_bwd_hook``: it holds the parameter's AccumulateGrad object and the registered post hook handle. """ assert hasattr(p, "_is_sharded") and hasattr( p, "_orig_size" ), "Parameters should have been sharded during construction." # If _local_shard has been set in the first lazy init and # current parameter is pointed to _local_shard, no need to # set the _local_shard again. if hasattr(p, "_local_shard"): # If CPU offloading, p._local_shard should have been placed on CPU # during its first lazy construction. if self.cpu_offload.offload_params: assert p._local_shard.device == torch.device( # type: ignore[attr-defined] "cpu" ), ( "Expected p._local_shard to be on CPU, " # type: ignore[attr-defined] f"but it's on {p._local_shard.device}" # type: ignore[attr-defined] ) return # A single shard of the parameters. Also makes p._local_shard to be on # CPU if we are CPU offloading, since would be on CPU during # init. if self.cpu_offload.offload_params: assert p.device == torch.device( "cpu" ), ("Expected param to be on CPU when cpu_offloading is enabled. " "If CPU offloading is enabled correctly, you may be " "accidentally moving the model to CUDA after FSDP initialization." ) p._local_shard = # type: ignore[attr-defined] # If CPU offloading, pin the memory to enable faster CPU -> GPU device # transfer. if self.cpu_offload.offload_params: assert p._local_shard.device == torch.device("cpu") # type: ignore[attr-defined] p._local_shard.pin_memory() # type: ignore[attr-defined] # When offloading parameters, also move the grad shard to CPU during # backward pass. In this case, it's important to pre-allocate the # CPU grad shard in pinned memory so that we can do a non-blocking # transfer. p._cpu_grad = torch.zeros_like( # type: ignore[attr-defined] p, device=torch.device("cpu") ).pin_memory() # We also maintain a full-sized parameter of type self.compute_dtype. # We resize the storage to size 0 at init (here) and only materialize # as needed. The storage may contain padding elements so that it is # evenly divisible by world_size, although these padding elements will # be removed before the relevant computation. if p._is_sharded: # type: ignore[attr-defined] p._full_param_padded = torch.zeros( # type: ignore[attr-defined] p.numel() * self.world_size, device=self.compute_device, dtype=p.dtype, ) _free_storage(p._full_param_padded) # type: ignore[attr-defined] def _set_is_root(self) -> None: """If ``True``, implies that no other :class:`FullyShardedDataParallel` instance wraps this one. Called once by :func:`_lazy_init`. """ if self._is_root is not None: return # No FSDP instance wraps this, else _is_root would be set to False. self._is_root = True # If final backward callback is never been queued, state should be IDLE. # If final backward callback is queued, the callback should be finished # and the state was reset to be IDLE. # This should be asserted at the beginning of forward pass in the root instance only. # For children instances, if they are checkpointed, state will not be reset to # IDLE after each inner forward/backward. self._assert_state(TrainingState_.IDLE) for n, m in self.named_modules(): # `n != ""` excludes self. if n != "" and isinstance(m, FullyShardedDataParallel): # We relax the assert for non-root instance, when the nested initialized module is wrapped # again in FSDP later, for example after training to run inference. assert ( m._is_root is None or not m._is_root ), "Non-root instance's _is_root flag should have not been set yet \ or has already been set as False." if m._is_root is None: m._is_root = False def _setup_streams(self) -> None: """Create streams to overlap data transfer and computation.""" if len(self._streams) > 0 or not self._is_root: return if torch.cuda.is_available(): # Stream for all-gathering parameters. self._streams["all_gather"] = torch.cuda.Stream() # Stream for overlapping grad reduction with the backward pass. self._streams["post_backward"] = torch.cuda.Stream() # We share streams with all children instances, which allows them to # overlap transfers across the forward pass without synchronizing with # the default stream. for n, m in self.named_modules(): if n != "" and isinstance(m, FullyShardedDataParallel): m._streams = self._streams m._fsdp_graph_order = self._fsdp_graph_order def _wait_for_previous_optim_step(self) -> None: """ The outer-most :class:`FullyShardedDataParallel` instance (i.e., the root instance) needs to synchronize with the default stream to ensure the previous optimizer step is done. """ if not torch.cuda.is_available(): return self._streams["all_gather"].wait_stream(torch.cuda.current_stream()) def _need_prefetch_pre_backward_hook(self) -> bool: if ( self.backward_prefetch == BackwardPrefetch.BACKWARD_PRE and self._fsdp_graph_order is not None and self._my_fsdp_idx_in_graph is not None and self._my_fsdp_idx_in_graph > 0 and self._fsdp_graph_order[self._my_fsdp_idx_in_graph - 1].training_state != TrainingState_.BACKWARD_POST ): return True else: return False def _need_prefetch_post_backward_hook(self) -> bool: if ( self.backward_prefetch == BackwardPrefetch.BACKWARD_POST and self._fsdp_graph_order is not None and self._my_fsdp_idx_in_graph is not None and self._my_fsdp_idx_in_graph > 0 and self._fsdp_graph_order[self._my_fsdp_idx_in_graph - 1].training_state != TrainingState_.BACKWARD_POST and self._fsdp_graph_order[self._my_fsdp_idx_in_graph - 1]._need_rebuild_full_params ): return True else: return False def forward(self, *args: Any, **kwargs: Any) -> Any: self._lazy_init() # Start of a forward pass. self.training_state = TrainingState_.FORWARD # All-gather full parameters, moving them to compute_device if # necessary. self._rebuild_full_params() # Wait for all_gather full parameters to finish before computation torch.cuda.current_stream().wait_stream(self._streams["all_gather"]) # Register backward hooks to reshard params and reduce-scatter grads. # These need to be re-registered every forward pass in some cases where grad_fn # is mutated. self._register_post_backward_hooks() outputs = self.module(*args, **kwargs) if self not in self._fsdp_graph_order: self._my_fsdp_idx_in_graph = len(self._fsdp_graph_order) self._fsdp_graph_order.append(self) if self.reshard_after_forward: self._free_full_params() # Switch to original local shards of params. We maintain this invariant throughout # the code, i.e., `` == p._local_shard`` after each function. This # also ensures that after the first forward, the optimizer state will be # initialized with the correct dtype and (sharded) size, since optimizer # state is typically initialized lazily in ``optim.step()``. self._use_param_local_shard() # Register pre-backward hooks to all-gather the params for the backward # pass (if output's grad was needed). This won't register anything if # we are in eval mode. outputs = self._register_pre_backward_hooks(outputs) # Done with a forward pass. self.training_state = TrainingState_.IDLE return outputs def _register_pre_backward_hooks(self, outputs: Any) -> Any: """Register pre-backward hook to run before the wrapped module's backward. Hooks should be attached to all outputs from the forward. Returns: outputs: new outputs with hooks registered if they requires gradient. """ # Reset before each backward pass self._need_rebuild_full_params = False if not torch.is_grad_enabled(): return outputs # don't register hooks if grad isn't enabled if self._is_root: # This actually means that only root instance has # _post_backward_callback_queued defined. Accidentally accessing this field # will assert on all other instances, giving us a nice bug checker. self._post_backward_callback_queued = False # Reset before each backward pass self._pre_backward_hook_has_run = False def _pre_backward_hook(*unused: Any) -> None: # Run ``_pre_backward_hook`` only once per backward pass if self._pre_backward_hook_has_run: return # try to queue final backward callback only once for root, so # that final backward callback is attached to the outer most # backward graph task and called after all the backward # calls are completed. if self._is_root: self._queue_wait_for_post_backward() if self._need_prefetch_pre_backward_hook(): # Always wait for all_gather before rebuilding full params, just # in case full params have already been prefetched in previous layer's # pre-backward hook. torch.cuda.current_stream().wait_stream(self._streams["all_gather"]) # All-gather full parameters, moving them to compute device if # necessary. self._rebuild_full_params() # Wait for all_gather to finish before computation torch.cuda.current_stream().wait_stream(self._streams["all_gather"]) # Prefetch next layer's full params in backward pass, # since it is prefetching, no need to wait for all_gather stream. if self._need_prefetch_pre_backward_hook(): self._fsdp_graph_order[self._my_fsdp_idx_in_graph - 1]._rebuild_full_params() # type: ignore[operator] self._pre_backward_hook_has_run = True # Prepare p.grad so that it is in the right shape, device, accumulated values, etc. self._prep_grads_for_backward() # Start of a backward pass for the first time in an backward pass. self._assert_state([TrainingState_.IDLE]) self.training_state = TrainingState_.BACKWARD_PRE def _register_hook(t: torch.Tensor) -> torch.Tensor: if t.requires_grad: t.register_hook(_pre_backward_hook) self._need_rebuild_full_params = True return t # Attach hooks to Tensor outputs. outputs = _apply_to_tensors(_register_hook, outputs) return outputs def _register_post_backward_hooks(self) -> None: """ Register backward hooks to reshard params and reduce-scatter grads. This is called during forward pass. The goal is to attach a hook on each of the parameter's gradient generating function (``grad_acc`` below) so that the hook is called *after* all gradients for that param are computed. Goals: 1. We want the hook to fire once and only once *after* all gradients are accumulated for a param. 2. If it fires more than once, we end up incorrectly shard the grad multiple times. (could lead to dimension too small) 3. If it fires once but too early or doesn't fire, we leave gradients unsharded. (could lead to dimension too large) Due to multiple-pass forward, this function can be called on the same parameter multiple times in a single forward pass. If we register the hook multiple time, we end up getting called multiple times. We could try to get a new hook every time and delete the previous one registered. However, due to *unknown reason* (I have debugged it for a long time!), in mixed precision mode, we get two different ``grad_acc`` objects below during different calls of this function (in the same forward pass). If we keep the last one, the hook end up firing too early. In full precision mode, we luckily get the *same* ``grad_acc`` object, so deleting and re-registering still ensured the hook fire once after all gradients are generated. Empirically, keep the first hook register per forward pass seems to work the best. We do need to remove the hook at the end of the backward pass. Otherwise, the next forward pass will not register a new hook, which is needed for a new forward pass. """ if not torch.is_grad_enabled(): return # don't register grad hooks if grad isn't enabled for p in self.params: if p.requires_grad: if hasattr(p, "_shard_bwd_hook"): continue # Register a hook on the first call, empirically, autograd # fires it at the end for this param, which makes sense. p_tmp = p.expand_as(p) # Get a grad_fn on p_tmp. assert ( p_tmp.grad_fn is not None ), "p_tmp grad_fn should not be None, it is used to access \ p's AccumulateGrad object and register post hook on it." grad_acc = p_tmp.grad_fn.next_functions[0][ 0 ] # Gets its AccumulateGrad object. handle = grad_acc.register_hook( functools.partial(self._post_backward_hook, p) ) p._shard_bwd_hook = (grad_acc, handle) # type: ignore[attr-defined] @torch.no_grad() def _post_backward_hook(self, param: Parameter, *unused: Any) -> None: """ At the start of :func:`_post_backward_hook`, ``param.grad`` contains the full gradient for the local batch. The reduce-scatter op will replace ``param.grad`` with a single shard of the summed gradient across all GPUs. This shard will align with the current GPU rank. For example:: before reduce_scatter: param.grad (GPU #0): [1, 2, 3, 4] param.grad (GPU #1): [5, 6, 7, 8] after reduce_scatter: param.grad (GPU #0): [6, 8] # 1+5, 2+6 param.grad (GPU #1): [10, 12] # 3+7, 4+8 The local GPU's ``optim.step`` is responsible for updating a single shard of params, also corresponding to the current GPU's rank. This alignment is created by :func:`_shard_parameters`, which ensures that the local optimizer only sees the relevant parameter shard. """ # First hook callback will see PRE state. If we have multiple params, # then subsequent hook callbacks will see POST state. self._assert_state([TrainingState_.BACKWARD_PRE, TrainingState_.BACKWARD_POST]) self.training_state = TrainingState_.BACKWARD_POST if param.grad is None: return if param.grad.requires_grad: raise RuntimeError( "FSDP only works with gradients that don't require gradients" ) self._free_full_params([param]) # Switch to local shard after backward. self._use_param_local_shard([param]) # Prefetch previous layer's full params in backward pass post backward hook, # If next layer's backward computation is done and full params are freed, # no need to prefetch the full params again. # Only prefetch full params if any of the next layer's outputs requires grad if self._need_prefetch_post_backward_hook(): self._fsdp_graph_order[self._my_fsdp_idx_in_graph - 1]._rebuild_full_params() # type: ignore[operator] # Next layer's computation will start right after this all_gather, # Wait for all_gather to finish before computation. torch.cuda.current_stream().wait_stream(self._streams["all_gather"]) # Wait for all work in the current stream to finish, then start the # reductions in post_backward stream. self._streams["post_backward"].wait_stream(torch.cuda.current_stream()) with["post_backward"]): orig_grad_data = if self.gradient_predivide_factor > 1: # Average grad by world_size for consistency with PyTorch DDP. param.grad.div_(self.gradient_predivide_factor) if param._is_sharded: # type: ignore[attr-defined] grad_flatten = torch.flatten(param.grad) chunks = list(grad_flatten.chunk(self.world_size)) num_pad = self.world_size * chunks[0].numel() - param.grad.numel() input_flattened = F.pad(grad_flatten, [0, num_pad]) output = torch.zeros_like(chunks[0]) dist._reduce_scatter_base( output, input_flattened, group=self.process_group ) if self.gradient_postdivide_factor > 1: # Average grad by world_size for consistency with PyTorch DDP. output.div_(self.gradient_postdivide_factor) = output else: # Currently the only way for _is_sharded to be False is if # world_size == 1. This could be relaxed in the future, e.g, # no sharding like PyTorch DDP, in which case grads should be # all-reduced here. assert ( self.world_size == 1 ), "Currently the only way for _is_sharded to be False is \ world_size == 1" # Regardless of sharding or not, offload the grad to CPU if we are # offloading params. This is so param and grad reside on same device # which is needed for the optimizer step. if self.cpu_offload.offload_params: # We specify non_blocking=True # and ensure the appropriate synchronization is done by waiting # streams in _wait_for_post_backward. param._cpu_grad.copy_( # type: ignore[attr-defined] param.grad.detach(), non_blocking=True ) # Don't let this memory get reused until after the transfer. # Point to CPU grad to offload it. Note that # the transfer is async so it is not necessarily done until we # explicitly synchronize in backward. = param._cpu_grad # type: ignore[attr-defined] # After _post_backward_hook returns, orig_grad_data will eventually # go out of scope, at which point it could otherwise be freed for # further reuse by the main stream while the div/reduce_scatter/copy # are underway in the post_backward stream. See: # orig_grad_data.record_stream(self._streams["post_backward"]) def _queue_wait_for_post_backward(self) -> None: """Try to queue a `wait_for_post_backward` callback. Only called on root and only queue one callback at the beginning of outer most backward. """ assert ( self._is_root ), "_queue_wait_for_post_backward can only be called on root." if not self._post_backward_callback_queued: self._assert_state([TrainingState_.IDLE]) self._post_backward_callback_queued = True Variable._execution_engine.queue_callback(self._wait_for_post_backward) @torch.no_grad() def _wait_for_post_backward(self) -> None: """Wait for post-backward to finish. Only called on root instance.""" assert self._is_root, "_wait_for_post_backward can only be called on root." # Check if the root module has params and if any of them has # the `requires_grad` field set. If `requires_grad=False` for # all the params, the post_backward hook will not fire and the # state will remain in `TrainingState_.BACKWARD_PRE`. if any([p.requires_grad for p in self.params]): self._assert_state(TrainingState_.BACKWARD_POST) else: self._assert_state(TrainingState_.BACKWARD_PRE) torch.cuda.current_stream().wait_stream(self._streams["post_backward"]) if self.cpu_offload.offload_params: # We need to wait for the non-blocking GPU -> # CPU grad transfers to finish. We need to do this for GPU -> CPU # copies because when grad is on CPU, it won't wait for any CUDA # stream to finish GPU -> CPU copies unless we explicitly block the # host-side with synchronize(). torch.cuda.current_stream().synchronize() # A backward pass is done, clean up below. def _remove_shard_bwd_hook(fsdp_module: FullyShardedDataParallel) -> None: """Helper used below on all fsdp modules.""" for p in fsdp_module.params: if p.requires_grad: if hasattr(p, "_shard_bwd_hook"): assert len(p._shard_bwd_hook) == 2 and len( # type: ignore[attr-defined] p._shard_bwd_hook # type: ignore[attr-defined] ), ( # type: ignore[attr-defined] "p._shard_bwd_hook fields are not valid." ) p._shard_bwd_hook[1].remove() # type: ignore[attr-defined] delattr(p, "_shard_bwd_hook") # Update root and nested FSDP's hooks and flags. for m in self.modules(): # includes self if isinstance(m, FullyShardedDataParallel): _remove_shard_bwd_hook(m) m._pre_backward_hook_has_run = False if any(p.requires_grad for p in m.parameters()): # Check if the module has params and if any of them has # the `requires_grad` field set. If `requires_grad=False` for # all the params, the post_backward hook will not fire and the # state will remain in `TrainingState_.BACKWARD_PRE`. if any([p.requires_grad for p in m.params]): m._assert_state(TrainingState_.BACKWARD_POST) else: m._assert_state(TrainingState_.BACKWARD_PRE) else: # When `m` and its children has no params or has params but # none with `requires_grad==True`, there are two cases: # 1. output tensors are `requires_grad==True`. In this case, # pre-backward hook is still registered, so it is in BACKWARD_PRE state. # 2. output tensors are `requires_grad==False`. In this case, # pre-backward hook is not registered, so it is in IDLE state. m._assert_state([TrainingState_.BACKWARD_PRE, TrainingState_.IDLE]) m.training_state = TrainingState_.IDLE if m._is_root: # reset this flag for cases like "one forward pass + multiple backward passes" self._post_backward_callback_queued = False @torch.no_grad() def _rebuild_full_params(self) -> None: """ Gather all shards of params. """ def update_p_data(output_tensor: torch.Tensor) -> None: """ Helper function to update pointer. Args: output_tensor (torch.Tensor): this tensor contains the data we just gathered. """ = output_tensor # Trim any padding and reshape to match original size. =[: p._orig_size.numel()].view(p._orig_size) # type: ignore[attr-defined] with["all_gather"]): for p in self.params: if self.cpu_offload.offload_params: # Move params to GPU if needed. Note that we don't use # self._full_param_padded.device here because the attr is # not set always, i.e. when world_size=1 and # p._is_sharded = False. However when it is set, the # device is always self.compute_device. =, non_blocking=True) # e.g., when world_size == 1 if not p._is_sharded: # type: ignore[attr-defined] continue # If full param has been rebuilt or has not been freed, no need to call all gather elif ( # type: ignore[attr-defined] == p._full_param_padded.size().numel() # type: ignore[attr-defined] ): update_p_data(p._full_param_padded) # type: ignore[attr-defined] continue else: # If full param has not been rebuilt or has been freed, call all gather p_data = # type: ignore[attr-defined] p_full_size = p._full_param_padded.size() # type: ignore[attr-defined] assert ( p_full_size.numel() == p_data.numel() * self.world_size ), "Param full size should be equal to its shard size multiply world_size." assert ( == 0 # type: ignore[attr-defined] ), "Full param's storage should have been freed before if all gather is needed." # type: ignore[attr-defined] # Allocate based on full size from all shards. _alloc_storage(p._full_param_padded, size=p_full_size) # type: ignore[attr-defined] output_tensor = p._full_param_padded # type: ignore[attr-defined] # Fill output_tensor with ( for each shard in self.world_size) dist._all_gather_base( output_tensor, p_data, group=self.process_group ) # Set = output_tensor (with padding trimmed) update_p_data(output_tensor) @torch.no_grad() def _prep_grads_for_backward(self) -> None: """Make sure p.grad has the correct size/device, otherwise set it to None.""" for p in self.params: if p.grad is not None and ( p.grad.size() != p._orig_size # type: ignore[attr-defined] or p.grad.device != p.device ): p.grad = None @torch.no_grad() def _free_full_params(self, params: Optional[List[Parameter]] = None) -> None: """ Free up storage for full parameters. """ if params is None: params = self.params current_stream = torch.cuda.current_stream() for p in params: # e.g., world_size == 1 if not p._is_sharded: # type: ignore[attr-defined] continue # Don't let PyTorch reuse this memory until all work in the current # stream is complete. p._full_param_padded.record_stream(current_stream) # type: ignore[attr-defined] # There may be external references to the Tensor Storage that we # can't modify, such as references that are created by # ctx.save_for_backward in the forward pass. Thus when we # unshard parameters, we should reuse the original Tensor # Storage object and unshard it in-place. For now, just resize # the Storage to 0 to save memory. _free_storage(p._full_param_padded) # type: ignore[attr-defined] @torch.no_grad() def _use_param_local_shard(self, params: Optional[List[Parameter]] = None) -> None: """Use local shard for a list of params. Also implicitly offloads parameters back to CPU if we are CPU offloading.""" if params is None: params = self.params for p in params: if self.cpu_offload.offload_params: # Ensure local_shard resides in CPU if we are offloading params. assert p._local_shard.device == torch.device( # type: ignore[attr-defined] "cpu" ), ( "Expected p._local_shard to be on CPU" ) = p._local_shard # type: ignore[attr-defined] def _assert_state(self, state: Union[TrainingState_, List[TrainingState_]]) -> None: """Assert we are in the given state.""" # Since assert can be turned off and this error checking # is really important, we use explicit error checking # and raise a ValueError if needed. if isinstance(state, TrainingState_): state = [state] if self.training_state not in state: msg = ( f"expected to be in states {state} but current state " f"is {self.training_state}" ) # In case we are failing in the context of autograd hook, asserting # may not generate useful msg. So, let's print it to be sure. if self.rank == 0: print(f"Asserting FSDP instance is: {self}") print(f"ERROR: {msg}") traceback.print_stack() raise ValueError(msg)
def _get_default_cuda_device(module: nn.Module) -> torch.device: """Try to infer CUDA device from module parameters.""" try: compute_device = next(module.parameters()).device if compute_device.type == "cuda": return compute_device # e.g., if module does not have parameters, it will throw StopIteration, # in this case, instead of raising exception, return cuda device. except StopIteration: pass # Fall back to current CUDA device return torch.device("cuda") def _free_storage(data: torch.Tensor) -> None: """Free underlying storage of a Tensor.""" if > 0: # Since we're modifying the Tensor's Storage directly, make sure the Tensor # is the sole occupant of the Storage. assert ( data.storage_offset() == 0 ), "The tensor is not the sole occupant of the storage." # type: ignore[attr-defined] @torch.no_grad() def _alloc_storage(data: torch.Tensor, size: torch.Size) -> None: """Allocate storage for a tensor.""" if == size.numel(): # no need to reallocate return assert ( == 0 ), "Then tensor storage should have been resized to be 0." # type: ignore[attr-defined]


Access comprehensive developer documentation for PyTorch

View Docs


Get in-depth tutorials for beginners and advanced developers

View Tutorials


Find development resources and get your questions answered

View Resources