# Copyright (c) Meta Platforms, Inc. and affiliates. # All rights reserved. # # This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. import logging import os from itertools import chain from typing import Any, Callable, cast, Dict, List, Optional, Tuple import torch import torch.distributed as dist from torch import nn from torch.distributed._composable.fsdp import CPUOffloadPolicy, fully_shard from torch.distributed._tensor import distribute_tensor, DTensor from torch.distributed._tensor.placement_types import DTensorSpec, TensorMeta from torch.distributed.checkpoint.state_dict import ( _init_optim_state, get_optimizer_state_dict, set_model_state_dict, set_optimizer_state_dict, StateDictOptions, ) from torch.distributed.device_mesh import DeviceMesh from torch.distributed.fsdp import ShardingStrategy from torch.nn.modules.module import _IncompatibleKeys from torch.optim import Optimizer from torchao.dtypes.nf4tensor import NF4Tensor, to_nf4 from torchtune.modules import TransformerDecoder from torchtune.modules.attention import MultiHeadAttention from torchtune.modules.model_fusion import DeepFusionModel, EarlyFusionModel from torchtune.modules.peft import get_adapter_state_dict from torchtune.utils import get_device, get_logger from torchtune.utils._logging import deprecated _log: logging.Logger = get_logger() torch_version = torch.__version__ # TODO: Fix issues with DSD before uncommenting. See #2313 and #2277. # _DISTRIBUTED_STATE_DICT_API_IS_AVAILABLE = ( # "dev" not in torch_version and torch_version_ge("2.6.0") # ) or ("dev" in torch_version and torch_version.split("dev")[1] >= "20241220") _DISTRIBUTED_STATE_DICT_API_IS_AVAILABLE = False def _get_sharding_strategy(strategy: str) -> ShardingStrategy: """Helper function to convert sharding strategy strings to ShardingStrategy enum.""" return getattr(ShardingStrategy, strategy) def is_distributed() -> bool: """Check if all environment variables required to initialize torch.distributed are set and distributed is properly installed. This indicates a distributed run. https://pytorch.org/docs/stable/distributed.html#environment-variable-initialization Checks the following conditions: * torch.distributed is available * master port and master address environment variables are set * world size is >1 * rank environment variable is set Returns: bool: True if all of the above conditions hold, False otherwise. """ port = os.environ.get("MASTER_PORT", "") addr = os.environ.get("MASTER_ADDR", "") size = int(os.environ.get("WORLD_SIZE", 1)) rank = int(os.environ.get("RANK", -1)) avlb = dist.is_available() return bool(port and addr and size >= 1 and rank >= 0 and avlb) def _broadcast_tensor(tensor: torch.Tensor, src: int = 0) -> torch.Tensor: """Broadcasts a tensor from a source to all other processes. Args: tensor (torch.Tensor): torch.Tensor to broadcast. src (int, optional): Source rank. Defaults to 0. Returns: torch.Tensor: Broadcasted tensor. """ if dist.is_available() and dist.is_initialized(): device = tensor.device if dist.get_backend() == "nccl": tensor = tensor.to(get_device("cuda")) dist.broadcast(tensor, src=src, group=None) return tensor.to(device) else: return tensor def get_distributed_backend(device_type: str, offload_ops_to_cpu: bool = False) -> str: """Gets the PyTorch Distributed backend based on device type. Args: device_type (str): Device type to get backend for. offload_ops_to_cpu (bool, optional): Flag to check if any operations should be offloaded to CPU. Examples of these kinds of operations are CPU offload for FSDP and asynchronous save for distributed checkpointing. Defaults to False. Example: >>> get_distributed_backend("cuda") 'nccl' >>> get_distributed_backend("cpu") 'gloo' >>> get_distributed_backend("cuda", offload_ops_to_cpu=True) 'cuda:nccl,cpu:gloo' Returns: str: Distributed backend for use in ``torch.distributed.init_process_group``. """ default_device_backend_map = dist.Backend.default_device_backend_map backend = "nccl" if device_type in default_device_backend_map: backend = default_device_backend_map[device_type] if offload_ops_to_cpu: backend = f"{device_type}:{backend},cpu:gloo" return backend @deprecated( msg="The functionality of `init_distributed` is covered by `torch.distributed.init_process_group`. " ) def init_distributed(**kwargs: Dict[str, Any]) -> bool: """Initialize process group required for ``torch.distributed``. Args: **kwargs (Dict[str, Any]): Additional arguments to pass to torch.distributed.init_process_group. Returns: bool: True if torch.distributed is initialized. Raises: RuntimeError: If torch.distributed is already initialized. """ if is_distributed(): if dist.is_initialized(): raise RuntimeError("torch.distributed already initialized.") dist.init_process_group(**kwargs) return True else: return False def set_torch_num_threads() -> None: """ Sets the number of threads used by torch to utilize all physical CPU cores for intra-op parallelism. Currently, this function sets num_threads to be the number of physical CPU cores divided by the number of GPUs as we use one process per GPU, and this avoids CPU oversubscription. Note that this is currently a rough approximation, and doesn't take into account environments where things like CPU affinity is set. """ num_threads = os.cpu_count() // ( torch.cuda.device_count() if torch.cuda.is_available() else 1 ) torch.set_num_threads(num_threads) _log.info(f"Set intra op parallelism no. of threads to {num_threads}") @deprecated( msg="`get_world_size_and_rank` will move to `torchtune.utils._device` in future releases. " "Please use `torchtune.utils.get_world_size_and_rank` instead." ) def get_world_size_and_rank() -> Tuple[int, int]: """Function that gets the current world size (aka total number of ranks) and rank number of the current process in the default process group. Returns: Tuple[int, int]: world size, rank """ if dist.is_available() and dist.is_initialized(): return torch.distributed.get_world_size(), torch.distributed.get_rank() else: return 1, 0 def validate_no_params_on_meta_device(model: nn.Module) -> None: """ Utility to validate that model has no params or buffers on meta device. If a meta param or buffer is found, an error indicating the param name will be raised. Args: model (nn.Module): model to check for meta params Raises: RuntimeError: If meta params or buffers exist in model """ for n, p in chain(model.named_parameters(), model.named_buffers()): if p.is_meta: raise RuntimeError(f"Unexpected param or buffer {n} on meta device.") def load_from_full_model_state_dict( model: "FSDPModule", # noqa full_sd: Dict[str, Any], device: torch.device, strict: bool = False, cpu_offload: bool = False, ) -> _IncompatibleKeys: """ Converting full state dict into a sharded state dict and loading it into FSDP model Args: model (FSDPModule): Model to generate fully qualified names for cpu_state_dict full_sd (Dict[str, Any]): a full state dict to load into the model device (torch.device): device used to move full state dict tensors strict (bool): flag to check if to load the model in strict mode cpu_offload (bool): flag to check if offload to CPU is enabled Returns: ``NamedTuple`` with ``missing_keys`` and ``unexpected_keys`` fields: * **missing_keys** is a list of str containing the missing keys * **unexpected_keys** is a list of str containing the unexpected keys Raises: NotImplementedError: If got FSDP with more than 1D. """ # PyTorch nightly versions from December 20, 2024, support the following features: # - `set_model_state_dict` with the `cpu_offload` option # - Multiple devices in local state dict # - Relative optimizations for improved memory performance # Please keep the version check `_DISTRIBUTED_STATE_DICT_API_IS_AVAILABLE` until these changes are # released in the PyTorch stable version. has_nf4 = any( hasattr(param, "_local_tensor") and isinstance(param._local_tensor, NF4Tensor) for param in model.parameters() ) meta_sharded_sd = model.state_dict() # NF4Tensor is not supported in `set_model_state_dict` right now, running with the previous logic right # now, would support in the future and remove the following code if _DISTRIBUTED_STATE_DICT_API_IS_AVAILABLE and not has_nf4: for param_name in full_sd.keys(): sharded_meta_param = meta_sharded_sd.get(param_name) full_sd[param_name] = full_sd[param_name].to(sharded_meta_param.dtype) options = StateDictOptions( full_state_dict=True, broadcast_from_rank0=True, strict=strict, cpu_offload=cpu_offload, ) return set_model_state_dict( model=model, model_state_dict=full_sd, options=options ) else: sharded_sd = {} for param_name, full_tensor in full_sd.items(): sharded_meta_param = meta_sharded_sd.get(param_name) full_tensor = full_tensor.to(sharded_meta_param.dtype).to(device) if hasattr(sharded_meta_param, "_local_tensor") and isinstance( sharded_meta_param._local_tensor, NF4Tensor ): block_size = sharded_meta_param._local_tensor.block_size scaler_block_size = sharded_meta_param._local_tensor.scaler_block_size full_tensor = to_nf4( full_tensor, block_size=block_size, scaler_block_size=scaler_block_size, ) # replicating logic from `_fsdp_param.py`` `_init_sharded_param` # otherwise `distribute_tensor(DTensor(local=NF4))` # requires dispatching `c10d.scatter_`` # long-term solution is `swap_tensor` mesh = sharded_meta_param.device_mesh if mesh.ndim > 1: raise NotImplementedError( f"only support 1D FSDP but got {mesh.ndim=}" ) shard_mesh_dim = 0 shard_world_size = mesh.size(shard_mesh_dim) shard_rank = cast( torch.distributed.ProcessGroup, mesh.get_group(shard_mesh_dim) ).rank() chunk = list(torch.chunk(full_tensor, shard_world_size, dim=0))[ shard_rank ] sharded_param = full_tensor.new_zeros(chunk.size()) sharded_param[: chunk.size(0)].copy_(chunk) # TODO: change to from_local API (need to add view support for NF4) sharded_tensor = DTensor( local_tensor=sharded_param, spec=DTensorSpec( mesh=sharded_meta_param.device_mesh, placements=sharded_meta_param.placements, tensor_meta=TensorMeta( shape=sharded_meta_param.size(), dtype=sharded_meta_param.dtype, stride=sharded_meta_param.stride(), ), ), requires_grad=sharded_meta_param.requires_grad, ) elif not hasattr(sharded_meta_param, "device_mesh"): # In cases where parts of the model aren't sharded, some parameters will be plain tensors sharded_tensor = full_tensor else: sharded_tensor = distribute_tensor( full_tensor, sharded_meta_param.device_mesh, sharded_meta_param.placements, ) if cpu_offload: sharded_tensor = sharded_tensor.cpu() sharded_sd[param_name] = nn.Parameter(sharded_tensor) # choose `assign=True` since we cannot call `copy_` on meta tensor return model.load_state_dict(sharded_sd, strict=strict, assign=True) def _gather_nf4_tensor(sharded_param: nn.Parameter) -> nn.Parameter: """ Manually gather NF4Tensor parameter since it does not support all_gather """ mesh = sharded_param.device_mesh nf4_tensor = sharded_param._local_tensor quant_params, metadata = nf4_tensor.fsdp_pre_all_gather(mesh) full_quant_params = [] for quant_param in quant_params: d0, *dn = quant_param.shape shape = (d0 * mesh.get_group().size(), *dn) full_quant_param = torch.empty( shape, device=quant_param.device, dtype=quant_param.dtype ) dist.all_gather_into_tensor( full_quant_param, quant_param, mesh.get_group(), async_op=False ) full_quant_params.append(full_quant_param) full_param, _ = nf4_tensor.fsdp_post_all_gather( full_quant_params, metadata, nf4_tensor.dtype ) return full_param def gather_cpu_state_dict( model: "FSDPModule", # noqa is_rank_zero: bool, device: Optional[torch.device] = None, adapter_weights_only: bool = False, ) -> Dict[str, Any]: """ Converting sharded state dict into a full state dict on CPU Returning non-empty result only on rank0 to avoid peaking CPU memory Currenltly we can used distributed state dict API to process model without NF4Tensor. Otherwise, we need to manually gather any NF4 tensors until all-gather is supported in the NF4Tensor subclass TODO: add support for NF4Tensor at distributed state dict API Args: model (FSDPModule): Model to generate fully qualified names for cpu_state_dict is_rank_zero (bool): flag to check if the process is on rank 0 device (Optional[torch.device]): device to use for sharded tensors. Default: None adapter_weights_only (bool): flag to check if only trainable parameters should be returned. Default: False Returns: Dict[str, Any]: State dict on CPU """ # TODO: Disabling DSD as it has issues. Add back changes in #2138 once DSD issue is fixed. cpu_state_dict = {} sharded_sd = model.state_dict() for param_name, param in sharded_sd.items(): if param.is_cpu: # Move back to device if offloaded to CPU param = param.to(device) if hasattr(param, "_local_tensor"): if isinstance(param._local_tensor, NF4Tensor): param = _gather_nf4_tensor(param) else: # Gather DTensor param = param.full_tensor() if isinstance(param, NF4Tensor): # upcasting NF4 to original dtype param = param.to(param.dtype) if is_rank_zero: cpu_state_dict[param_name] = param.cpu() torch.distributed.barrier() if adapter_weights_only: cpu_state_dict = get_adapter_state_dict(cpu_state_dict, device=None) return cpu_state_dict def get_full_optimizer_state_dict( model: "FSDPModule", # noqa opt: Optimizer, is_rank_zero: bool, device: Optional[torch.device] = None, ) -> Dict[str, Any]: """ Converting optimizer state from sharded to full For example, "exp_avg" in AdamW is `DTensor`, "exp_avg.full_tensor()" converts it to plain tensor on rank 0 Returning non-empty cpu state dict on rank 0 """ options = StateDictOptions( full_state_dict=True, broadcast_from_rank0=True, cpu_offload=True ) full_state_dict = get_optimizer_state_dict( model=model, optimizers=opt, options=options ) if is_rank_zero: return full_state_dict else: return {} def load_from_full_optimizer_state_dict( model: "FSDPModule", # noqa opt: Optimizer, full_sd: Dict[str, Any], device: torch.device, ) -> None: """ Converting full optimizer state to sharded state dict and loading it into optimizer """ if _DISTRIBUTED_STATE_DICT_API_IS_AVAILABLE: options = StateDictOptions( full_state_dict=True, broadcast_from_rank0=True, cpu_offload=device is torch.device("cpu"), ) set_optimizer_state_dict( model=model, optimizers=opt, optim_state_dict=full_sd, options=options ) else: PARAMS = "params" # noqa: N806 _init_optim_state(opt) param_groups = opt.state_dict()["param_groups"] state = opt.state_dict()["state"] full_param_groups = full_sd["param_groups"] full_state = full_sd["state"] for param_group, full_param_group in zip(param_groups, full_param_groups): for key, value in full_param_group.items(): if key == PARAMS: continue param_group[key] = value for pid, full_pid in zip(param_group[PARAMS], full_param_group[PARAMS]): if pid not in state: continue param_state = state[pid] full_param_state = full_state[full_pid] for attr, full_tensor in full_param_state.items(): sharded_tensor = param_state[attr] if isinstance(sharded_tensor, DTensor): # exp_avg is DTensor param_state[attr] = distribute_tensor( full_tensor, sharded_tensor.device_mesh, sharded_tensor.placements, ) else: # step is plain tensor param_state[attr] = full_tensor opt.load_state_dict( { "param_groups": param_groups, "state": state, } ) def get_shard_conditions( name: str, module: nn.Module, names_to_match: Optional[List[str]] = None, *args, **kwargs, ) -> bool: """ Returs True for layers named {}.layers.i or layers that exactly match names_to_match, otherwise, returns False. This is a helper function for sharding a model with FSDP. In :func:`~torchtune.training.shard_model`, we iterate over the model's named modules and apply fully_shard using this condition. As part of our sharding strategy, we want each layer to be sharded separately, as this is generally efficient. We may also want to shard certain modules that are not layers, such as the embedding module. #TODO: a more robust way would be to shard on the module type, not the name. Args: name (str): Name of the module. module (nn.Module): Module to be sharded. names_to_match (Optional[List[str]]): List of names to match, if any. *args: Variable length argument list to be passed to the Embedding module. **kwargs: Arbitrary keyword arguments to be passed to the Embedding module. Returns: bool: True if the module name matches the condition, False otherwise. Examples: >>> names_to_match = ["embedding"] >>> layer_names = ["layers.0", "decoder.layers.1", "encoder.layers.2.attention", "my_wrapper.layer.1.something", "embedding"] >>> matches = [] >>> for name in layer_names: >>> if shard_condition_is_layer_or_match(name, None): matches.append(name) >>> print(matches) >>> ["layers.0", "decoder.layers.1", "embedding"] """ if names_to_match and name in names_to_match: return True name_list = name.split(".") if len(name_list) >= 2: return name_list[-2] == "layers" and str.isdigit(name_list[-1]) return False def shard_model( model: TransformerDecoder, shard_conditions: List[Callable[[str, nn.Module], bool]], *, cpu_offload: bool, reshard_after_forward: bool = True, dp_mesh: Optional[DeviceMesh] = None, ) -> None: """ Utility to shard a model with FSDP using the PyTorch Distributed fully_shard API. This method will over the model's named modules from the bottom-up and apply shard modules based on whether they meet any of the criteria from shard_conditions. Args: model (TransformerDecoder): Model to shard with FSDP. shard_conditions (List[Callable[[str, nn.Module], bool]]): A list of functions to determine which modules to shard with FSDP. Each function should take module name (relative to root) and the module itself, returning True if FSDP should shard the module and False otherwise. If any of shard_conditions return True for a given module, it will be sharded by FSDP. cpu_offload (bool): If set to True, FSDP will offload parameters, gradients, and optimizer states to CPU. reshard_after_forward (bool): Whether to reshard parameters and buffers after the forward pass. Setting this to True corresponds to the FULL_SHARD sharding strategy from FSDP1, while setting it to False corresponds to the SHARD_GRAD_OP sharding strategy. dp_mesh (Optional[DeviceMesh]): Device mesh to use for FSDP sharding under mutliple parallelism. Default to None. Raises: ValueError: If no layer modules were sharded, indicating that no shard_condition was triggered. """ fsdp_kwargs = {"reshard_after_forward": reshard_after_forward, "mesh": dp_mesh} if cpu_offload: fsdp_kwargs["offload_policy"] = CPUOffloadPolicy() # Shard the model with FSDP, iterating in reverse to start with # lowest-level modules first num_layers_sharded = 0 for n, m in reversed(list(model.named_modules())): if any([shard_condition(n, m) for shard_condition in shard_conditions]): fully_shard(m, **fsdp_kwargs) num_layers_sharded += 1 if num_layers_sharded == 0: raise ValueError( "No layer modules were sharded. Please check if shard conditions are working as expected." ) # Finally shard the entire model to account for any stragglers fully_shard(model, **fsdp_kwargs) def prepare_mha_for_tp( model: nn.Module, tp_mesh: DeviceMesh, ) -> nn.Module: """ Utility to scale MultiHeadAttention parameters(num_heads, num_kv_heads, embed_dim) across tensor parallel devices. Each device will handle a portion of the attention computations. Args: model (nn.Module): Model whose attention parameters will be scaled by TP size. tp_mesh (DeviceMesh): Tensor parallel device mesh. Returns: nn.Module: The model with scaled MultiHeadAttention parameters. Raises: ValueError: If attention heads, kv heads, or embed dimension is not divisible by TP size. Examples: >>> from torchtune.modules import TransformerDecoder >>> from torch.distributed.device_mesh import DeviceMesh >>> model = TransformerDecoder( num_heads=32, num_kv_heads=32, embed_dim=4096, ) >>> tp_mesh = DeviceMesh("cuda", torch.arange(2)) # 2 GPUs >>> model = prepare_mha_for_tp(model, tp_mesh) >>> # Now each GPU has: >>> # num_heads = 16 (32/2) >>> # num_kv_heads = 16 (32/2) >>> # embed_dim = 2048 (4096/2) """ # Handle fusion models by extracting decoder is_fusion_model = isinstance(model, (DeepFusionModel, EarlyFusionModel)) decoder = model.decoder if is_fusion_model else model tp_size = tp_mesh.size() for m in list(decoder.modules()): if isinstance(m, MultiHeadAttention): # Adjust attention module to use the local number of heads if m.num_heads % tp_size != 0: raise ValueError( f"Number of attention heads ({m.num_heads}) must be divisible by " f"tensor parallel size ({tp_size})." ) if m.num_kv_heads % tp_size != 0: raise ValueError( f"Number of KV heads ({m.num_kv_heads}) must be divisible by " f"tensor parallel size ({tp_size})." ) if m.embed_dim % tp_size != 0: raise ValueError( f"Embedding dimension ({m.embed_dim}) must be divisible by " f"tensor parallel size ({tp_size})." ) m.num_heads = m.num_heads // tp_size m.num_kv_heads = m.num_kv_heads // tp_size m.embed_dim = m.embed_dim // tp_size if is_fusion_model: model.decoder = decoder return model