verl/utils/fsdp_utils.py

# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import functools
import itertools
import json
import math
import os
from contextlib import contextmanager, nullcontext
from typing import Dict

import torch
import torch.distributed as dist
import torch.nn as nn
from packaging import version
from torch.distributed import DeviceMesh
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp._runtime_utils import _lazy_init
from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy, transformer_auto_wrap_policy
from transformers.trainer_pt_utils import get_module_class_from_name

if version.parse(torch.__version__) >= version.parse("2.6"):
    from torch.distributed.fsdp import CPUOffloadPolicy, FSDPModule, MixedPrecisionPolicy, fully_shard
elif version.parse(torch.__version__) >= version.parse("2.4"):
    from torch.distributed._composable.fsdp import CPUOffloadPolicy, FSDPModule, MixedPrecisionPolicy, fully_shard
else:
    fully_shard, MixedPrecisionPolicy, FSDPModule, CPUOffloadPolicy = None, None, None, None


def init_fn(x: torch.nn.Module):
    if torch.distributed.get_rank() != 0:
        x = x.to_empty(device=torch.cuda.current_device(), recurse=False)
        torch.cuda.empty_cache()
    return x


def get_init_weight_context_manager(use_meta_tensor=True, mesh: DeviceMesh = None):
    from accelerate import init_empty_weights

    cpu_init_weights = lambda: torch.device("cpu")
    if use_meta_tensor:
        if mesh is None:
            init_context = init_empty_weights if torch.distributed.get_rank() != 0 else cpu_init_weights
        else:
            init_context = init_empty_weights if mesh.get_coordinate()[-1] != 0 else cpu_init_weights
    else:
        init_context = cpu_init_weights
    return init_context


# Copyright 2020-present the HuggingFace Inc. team.
# Adapted from https://github.com/huggingface/transformers/src/transformers/trainer.py
def get_fsdp_wrap_policy(module, config=None, is_lora=False):
    """Get FSDP wrap policy for the module.

    Args:
        module: The module to get wrap policy for
        config: Configuration for wrap policy
        is_lora: Whether to enable lambda policy for LoRA modules
    """
    if config is None:
        config = {}

    # NOTE: This is a temporary workaround to be compatible with the OmegaConf & dataclass. We will remove this once we have make all config in verl from OmegaConf to data class.
    def _get_attr(attr_name, default_value=None):
        if hasattr(config, "get"):
            return config.get(attr_name, default_value)
        else:
            return config.__getattribute__(attr_name)

    if _get_attr("disable", False):
        return None

    default_transformer_cls_names_to_wrap = getattr(module, "_no_split_modules", None)
    fsdp_transformer_layer_cls_to_wrap = _get_attr("transformer_layer_cls_to_wrap", default_transformer_cls_names_to_wrap)
    min_num_params = _get_attr("min_num_params", 0)
    auto_wrap_policy = None

    policies = []

    from torch.distributed.fsdp.wrap import _or_policy, lambda_auto_wrap_policy

    # Add lambda policy for LoRA modules if is_lora is True
    if is_lora:

        def lambda_policy_fn(module):
            return bool(len(list(module.named_children())) == 0 and getattr(module, "weight", None) is not None and module.weight.requires_grad)

        lambda_policy = functools.partial(lambda_auto_wrap_policy, lambda_fn=lambda_policy_fn)
        policies.append(lambda_policy)

    if min_num_params > 0:
        size_policy = functools.partial(size_based_auto_wrap_policy, min_num_params=min_num_params)
        policies.append(size_policy)
    elif fsdp_transformer_layer_cls_to_wrap is not None:
        transformer_cls_to_wrap = set()
        for layer_class in fsdp_transformer_layer_cls_to_wrap:
            transformer_cls = get_module_class_from_name(module, layer_class)
            if transformer_cls is None:
                raise Exception("Could not find the transformer layer class to wrap in the model.")
            else:
                transformer_cls_to_wrap.add(transformer_cls)

        transformer_policy = functools.partial(
            transformer_auto_wrap_policy,
            transformer_layer_cls=transformer_cls_to_wrap,
        )
        policies.append(transformer_policy)

    if len(policies) > 0:
        auto_wrap_policy = functools.partial(_or_policy, policies=policies)

    return auto_wrap_policy


@torch.no_grad()
def offload_fsdp_model_to_cpu(model: FSDP, empty_cache: bool = True):
    if fsdp_version(model) == 2:
        offload_fsdp2_model_to_cpu(model, empty_cache)
        return

    assert isinstance(model, FSDP)
    # lazy init FSDP model
    _lazy_init(model, model)
    assert model._is_root, "Only support root model offloading to CPU"
    for handle in model._all_handles:
        if handle._offload_params:
            continue
        flat_param = handle.flat_param
        assert flat_param.data.data_ptr() == flat_param._local_shard.data_ptr() and id(flat_param.data) != id(flat_param._local_shard) and flat_param.data.size() == flat_param._local_shard.size()
        handle.flat_param_to(torch.device("cpu"), non_blocking=True)
        # the following still keeps id(._local_shard) != id(.data)
        flat_param._local_shard = flat_param.data
        assert id(flat_param._local_shard) != id(flat_param.data)
    if empty_cache:
        torch.cuda.empty_cache()


@torch.no_grad()
def offload_fsdp2_model_to_cpu(model, empty_cache: bool = True):
    for param in model.parameters():
        param.data = param.data.to(torch.device("cpu"), non_blocking=True)
    if empty_cache:
        torch.cuda.empty_cache()


@torch.no_grad()
def load_fsdp_model_to_gpu(model: FSDP):
    if fsdp_version(model) == 2:
        load_fsdp2_model_to_gpu(model)
        return

    assert isinstance(model, FSDP)
    # lazy init FSDP model
    _lazy_init(model, model)
    assert model._is_root, "Only support root model loading to GPU"
    device_id = torch.cuda.current_device()
    for handle in model._all_handles:
        if handle._offload_params:
            continue
        flat_param = handle.flat_param
        handle.flat_param_to(torch.device(f"cuda:{device_id}"), non_blocking=True)
        # the following still keeps id(._local_shard) != id(.data)
        flat_param._local_shard = flat_param.data


@torch.no_grad()
def load_fsdp2_model_to_gpu(model):
    device = torch.cuda.current_device()
    for param in model.parameters():
        param.data = param.data.to(device, non_blocking=True)


@torch.no_grad()
def offload_fsdp_optimizer(optimizer):
    if not optimizer.state:
        return
    for param_group in optimizer.param_groups:
        for param in param_group["params"]:
            state = optimizer.state[param]
            for key, value in state.items():
                if isinstance(value, torch.Tensor):
                    state[key] = value.to("cpu", non_blocking=True)


@torch.no_grad()
def load_fsdp_optimizer(optimizer, device_id):
    if not optimizer.state:
        return
    for param_group in optimizer.param_groups:
        for param in param_group["params"]:
            state = optimizer.state[param]
            for key, value in state.items():
                if isinstance(value, torch.Tensor):
                    state[key] = value.to(device_id, non_blocking=True)


@contextmanager
def meta_device_init():
    """
    Create model parameters with meta device.

    Note buffers in model will still be initialized in default device (e.g., CPU),
    since the buffers can be non-persistent and filled with expected values that can
    NOT be captured in meta device.
    """
    device = torch.device("meta")
    old_register_parameter = nn.Module.register_parameter
    registered = set()

    def register_empty_parameter(module, name, param):
        old_register_parameter(module, name, param)
        # we will skip register shared parameters as it
        # is already registered previously
        if param is not None and param not in registered:
            param_cls = type(module._parameters[name])
            kwargs = module._parameters[name].__dict__
            kwargs["requires_grad"] = param.requires_grad
            module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
            registered.add(module._parameters[name])

    try:
        nn.Module.register_parameter = register_empty_parameter
        yield
    finally:
        registered.clear()
        nn.Module.register_parameter = old_register_parameter


def parallel_load_safetensors(filepath):
    """
    Parallel load safetensors from huggingface checkpoint

    Huggingface checkpoint contains:

    - config.json: a json file for model configuration
    - model.safetensor.index.json: a json file for safetensors (parameters & buffers) index
    - model-000x-of-ooxx.safetensors: a binary file for safetensors (parameters & buffers) chunks

    Or (when model is small),

    - model.safetensors: a binary file for all parameters and buffers

    Each rank will own a part of model chunks and load them directly into GPU memory.
    """
    from safetensors.torch import load_file

    safetensors2param = {}

    index_file = os.path.join(filepath, "model.safetensors.index.json")
    if os.path.exists(index_file):
        index = json.load(open(index_file, "rb"))
        for param_name, filename in index["weight_map"].items():
            safetensors2param.setdefault(filename, []).append(param_name)
    else:
        # in this case, the model is small and we can load it all at once
        param_file = os.path.join(filepath, "model.safetensors")
        assert os.path.exists(param_file), f"Cannot find {param_file}"
        states = load_file(param_file)
        for param_name in states:
            safetensors2param.setdefault("model.safetensors", []).append(param_name)
        del states

    total_files = len(safetensors2param)
    ckpt_chunks = sorted(safetensors2param.keys())
    world_size = dist.get_world_size()
    size = int(math.ceil(total_files / world_size))
    ckpt_chunks = [ckpt_chunks[rank * size : rank * size + size] for rank in range(world_size)]

    shard_states = {}
    device = torch.cuda.current_device()
    for rank, files in enumerate(ckpt_chunks):
        if rank == dist.get_rank():
            for file in files:
                file = os.path.join(filepath, file)
                states = load_file(file, device=device)
                # print(f"rank {rank} loading {file}...")
                shard_states.update(states)
        else:
            for file in files:
                for param_name in safetensors2param[file]:
                    shard_states[param_name] = rank
    return shard_states


def parallel_init_module_fn(module: torch.nn.Module, shard_states: Dict[str, torch.nn.Parameter]):
    """
    Generate a function to initialize sub-modules in the `module` with `shard_states`
    from huggingface checkpoint.

    Args:
        module (torch.nn.Module): the global module to be initialized
        shard_states (Dict[str, torch.nn.Parameter]): the shard states from huggingface checkpoint

    Returns:
        init_fn (Callable): a function to initialize sub-modules in the `module` with `shard_states`
    """

    state2fqn = {}
    for name, state in itertools.chain(module.named_parameters(remove_duplicate=False), module.named_buffers(remove_duplicate=False)):
        state2fqn.setdefault(state, []).append(name)
    # remove standalone parameters and buffers
    shared = {s for s, names in state2fqn.items() if len(names) > 1}
    materialized_states = {}

    @torch.no_grad()
    def create_and_sync_state(param_name, state, is_param):
        assert param_name in shard_states, f"{param_name} not loaded"
        device = torch.cuda.current_device()
        if is_param:
            param = torch.nn.Parameter(torch.empty_like(state.data, device=device), requires_grad=state.requires_grad)
        else:  # buffer
            param = torch.empty_like(state.data, device=device)
        loaded = shard_states[param_name]
        if isinstance(loaded, (torch.nn.Parameter, torch.Tensor)):
            # NOTE: loaded.dtype can be different with param.dtype
            param.data.copy_(loaded.data)
            dist.broadcast(param.data, src=dist.get_rank())
        else:
            assert isinstance(loaded, int)  # the rank that holds the state
            dist.broadcast(param.data, src=loaded)
        shard_states.pop(param_name)
        del loaded
        return param

    def init_fn(sub_mod: torch.nn.Module, recurse: bool = True):
        param_and_buffers = tuple(sub_mod.named_parameters(recurse=False)) + tuple(sub_mod.named_buffers(recurse=False))
        # param_and_buffers = sorted(sub_mod.named_parameters(recurse=False), key=lambda x: x[0])
        for name, state in param_and_buffers:
            if not state.is_meta:
                continue
            is_param = name in sub_mod._parameters
            fqn = state2fqn[state].pop(0)
            # non-persistent buffers will not be saved in state dict, we can safely skip it
            if (not is_param) and fqn not in shard_states:
                if state.is_meta:
                    raise RuntimeError(f"find a non-persistent buffer ({fqn}) initiated with device meta. Such buffer is not saved in checkpoint and user should guarantee to init in CPU / GPU device.")
                continue
            # for shared parameter, we get it from the first time it is created
            if state in shared:
                if state not in materialized_states:
                    materialized_states[state] = create_and_sync_state(fqn, state, is_param)
                else:
                    if fqn in shard_states:
                        shard_states.pop(fqn)
                materialize_state = materialized_states[state]
            # for not shared parameter, we create it directly
            else:
                materialize_state = create_and_sync_state(fqn, state, is_param)
            if is_param:
                sub_mod._parameters[name] = materialize_state
            else:
                sub_mod._buffers[name] = materialize_state
        if recurse:
            for module in sub_mod.children():
                init_fn(module, recurse=True)

        # for debug
        # if len(shard_states) == 0: print("clear")
        return sub_mod

    return init_fn


def fsdp_version(model):
    if isinstance(model, FSDP):
        return 1
    elif isinstance(model, FSDPModule):
        return 2
    else:
        return 0


def get_fsdp_state_ctx(model, state_type, state_cfg, optim_cfg):
    if fsdp_version(model) == 1:
        return FSDP.state_dict_type(model, state_type, state_cfg, optim_cfg)
    else:
        return nullcontext()


def fsdp2_load_full_state_dict(model: torch.nn.Module, full_state: dict, device_mesh=None, cpu_offload=None):
    """
    Loads the full state dict (could be only on rank 0) into the sharded model. This is done by broadcasting the
    parameters from rank 0 to all other ranks. This function modifies the model in-place.

    Args:
        model (`torch.nn.Module`): The model to load the state dict into
        full_state (`dict`): The full state dict to load, can only be on rank 0
    """
    from torch.distributed.checkpoint.state_dict import StateDictOptions, set_model_state_dict

    # To broadcast, it needs to be instantiated in the GPU.
    if dist.get_rank() == 0:
        model = model.to(device=torch.cuda.current_device(), non_blocking=True)
    else:
        model = model.to_empty(device=torch.cuda.current_device())

    cpu_offload = cpu_offload is not None
    options = StateDictOptions(full_state_dict=True, cpu_offload=cpu_offload, broadcast_from_rank0=True)
    set_model_state_dict(model, full_state, options=options)

    # rotary_emb is not in state_dict, so we need to broadcast it manually
    for name, buf in model.named_buffers():
        dist.broadcast(buf, src=0)

    if cpu_offload:
        model.to("cpu", non_blocking=True)
        for buf in model.buffers():
            buf.data = buf.data.to(torch.cuda.current_device())


def apply_fsdp2(model, fsdp_kwargs, config):
    """model: AutoModelForCausalLM"""
    assert CPUOffloadPolicy is not None, "PyTorch version >= 2.4 is required for using fully_shard API (FSDP2)"

    default_transformer_cls_names_to_wrap = getattr(model, "_no_split_modules", None)
    fsdp_transformer_layer_cls_to_wrap = config.get("wrap_policy", {}).get("transformer_layer_cls_to_wrap", default_transformer_cls_names_to_wrap)

    if isinstance(fsdp_transformer_layer_cls_to_wrap, str):
        fsdp_transformer_layer_cls_to_wrap = [fsdp_transformer_layer_cls_to_wrap]

    assert len(fsdp_transformer_layer_cls_to_wrap) > 0 and fsdp_transformer_layer_cls_to_wrap[0] is not None

    modules = []
    for name, module in model.named_modules():
        if module.__class__.__name__ in fsdp_transformer_layer_cls_to_wrap or (isinstance(module, nn.Embedding) and not model.config.tie_word_embeddings):
            modules.append(module)

    for idx, module in enumerate(modules):
        fully_shard(module, **fsdp_kwargs)
    fully_shard(model, **fsdp_kwargs)  # fsdp2 will not reshard_after_forward for root module


def fsdp2_clip_grad_norm_(parameters, max_norm, norm_type=2.0, error_if_nonfinite=False, foreach=None):
    """torch.nn.utils.clip_grad_norm_ cann't run on cpu parameter DTensor"""
    from torch.nn.utils.clip_grad import _clip_grads_with_norm_, _get_total_norm

    if isinstance(parameters, torch.Tensor):
        parameters = [parameters]
    else:
        # prevent generators from being exhausted
        parameters = list(parameters)
    grads = [p.grad for p in parameters if p.grad is not None]
    total_norm = _get_total_norm(grads, norm_type, error_if_nonfinite, foreach)
    total_norm = total_norm.to(torch.cuda.current_device(), non_blocking=True)
    _clip_grads_with_norm_(parameters, max_norm, total_norm, foreach)
    return total_norm