Model-Optimizer/modelopt/torch/utils/distributed.py

# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# 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.

"""Utility functions for using torch.distributed."""

import functools
import io
import os
import time
from collections.abc import Callable
from typing import Any

import torch
import torch.distributed

# TODO: update this import to from torch.distributed.tensor import DTensor after we deprecate Torch<2.5
from torch.distributed._tensor import DTensor

__all__ = [
    "DistributedProcessGroup",
    "ParallelState",
    "backend",
    "barrier",
    "is_available",
    "is_initialized",
    "is_master",
    "rank",
    "size",
]


def is_available() -> bool:
    """Returns whether the distributed package is available."""
    return torch.distributed.is_available()


def is_initialized() -> bool:
    """Returns whether the distributed package is initialized."""
    return is_available() and torch.distributed.is_initialized()


def backend() -> str | None:
    """Returns the distributed backend."""
    if is_initialized():
        return "torch"
    return None


def size(group=None) -> int:
    """Returns the number of processes."""
    if backend() == "torch":
        return torch.distributed.get_world_size(group=group)
    return 1


def rank(group=None) -> int:
    """Returns the rank of the current process."""
    if backend() == "torch":
        return torch.distributed.get_rank(group=group)
    return 0


def is_master(group=None) -> bool:
    """Returns whether the current process is the master process."""
    return rank(group=group) == 0


def _serialize(obj: Any) -> torch.Tensor:
    buffer = io.BytesIO()
    torch.save(obj, buffer)
    storage = torch.UntypedStorage.from_buffer(buffer.getvalue(), dtype=torch.uint8)
    tensor = torch.ByteTensor(storage)
    return tensor


def _deserialize(tensor: torch.Tensor, size: int | None = None) -> Any:
    buffer = tensor.numpy().tobytes()
    if size is not None:
        buffer = buffer[:size]
    obj = torch.load(io.BytesIO(buffer), weights_only=False)
    return obj


def _broadcast(tensor: torch.Tensor, src: int = 0, group=None) -> None:
    if backend() == "torch":
        torch.distributed.broadcast(tensor, src, group)


def broadcast(obj: Any, src: int = 0, group=None) -> Any:
    """Broadcasts an object from the source to all other processes."""
    if size() == 1:
        return obj

    # serialize
    if rank() == src:
        tensor = _serialize(obj).cuda()

    # broadcast the tensor size
    tensor_size = (
        torch.LongTensor([tensor.numel()]).cuda() if rank() == src else torch.LongTensor([0]).cuda()
    )
    _broadcast(tensor_size, src=src, group=group)

    # broadcast the tensor
    if rank() != src:
        tensor = torch.ByteTensor(size=(tensor_size.item(),)).cuda()
    _broadcast(tensor, src=src, group=group)

    # deserialize
    if rank() != src:
        obj = _deserialize(tensor.cpu())
    return obj


def _allgather(tensors: list[torch.Tensor], tensor: torch.Tensor) -> None:
    if backend() == "torch":
        torch.distributed.all_gather(tensors, tensor)


def allgather(obj: Any) -> list[Any]:
    """Gathers an object from all processes into a list."""
    if size() == 1:
        return [obj]

    # serialize
    tensor = _serialize(obj).cuda()

    # gather the tensor size
    tensor_size = torch.LongTensor([tensor.numel()]).cuda()
    tensor_sizes = [torch.LongTensor([0]).cuda() for _ in range(size())]
    _allgather(tensor_sizes, tensor_size)
    tensor_sizes = [int(tensor_size.item()) for tensor_size in tensor_sizes]
    max_size = max(tensor_sizes)

    # gather the tensor
    tensors = [torch.ByteTensor(size=(max_size,)).cuda() for _ in tensor_sizes]
    if tensor_size != max_size:
        padding = torch.ByteTensor(size=(max_size - tensor_size,)).cuda()
        tensor = torch.cat((tensor, padding), dim=0)
    _allgather(tensors, tensor)

    # deserialize
    objs = []
    for tensor_size, tensor in zip(tensor_sizes, tensors):
        obj = _deserialize(tensor.cpu(), size=tensor_size)
        objs.append(obj)
    return objs


def allreduce(obj: Any, reduction: str = "sum") -> Any:
    """Reduces an object from all processes."""
    objs = allgather(obj)
    if reduction == "sum":
        return sum(objs)
    else:
        raise NotImplementedError(reduction)


def barrier(group=None) -> None:
    """Synchronizes all processes."""
    if size() == 1:
        return
    if backend() == "torch":
        torch.distributed.barrier(group=group)


def master_only(func):
    """Decorator to run a function only on the master process and broadcast the result."""

    @functools.wraps(func)
    def wrapper(*args, **kwargs):
        return broadcast(func(*args, **kwargs) if is_master() else None)

    return wrapper


class DistributedProcessGroup:
    """A convenient wrapper around torch.distributed.ProcessGroup objects."""

    def __init__(self, group: torch.distributed.ProcessGroup | int | None = None):
        """Initialize the distributed process group."""
        self.group = group

    def is_initialized(self) -> bool:
        """Check if the distributed process group is initialized."""
        return backend() == "torch" and self.group != -1

    def rank(self) -> int:
        """Get the rank of the current process group."""
        return rank(group=self.group) if self.is_initialized() else -1

    def world_size(self) -> int:
        """Get the world size of the current process group."""
        return size(group=self.group) if self.is_initialized() else -1

    def __repr__(self) -> str:
        return f"group: {self.group}, initialized: {self.is_initialized()}, world size: {self.world_size()}"

    @staticmethod
    def get_dist_syncd_obj(
        obj: Any,
        groups: "DistributedProcessGroup | list[DistributedProcessGroup]",
        op: Callable,
    ):
        """Get the distributed synchronized object across the specified distributed groups."""

        def _get_dist_syncd_obj_across_group(obj, group: DistributedProcessGroup):
            if not group.is_initialized():
                return obj
            obj_list = [None] * group.world_size()
            torch.distributed.all_gather_object(obj_list, obj, group=group.group)
            return op(obj_list)

        for group in groups if isinstance(groups, list) else [groups]:
            obj = _get_dist_syncd_obj_across_group(obj, group)

        return obj


class ParallelState:
    """A class to manage various parallel groups such as data parallel, tensor parallel etc.

    Specify the parallel groups of type :class:`torch.distributed.ProcessGroup` for the current module.
    If the parallel group is not used, it should be set to `-1`.
    if a parallel group is `None`, it will use the default PyTorch distributed process group which is the whole world.
    """

    def __init__(
        self,
        data_parallel_group: torch.distributed.ProcessGroup | int | None = None,
        tensor_parallel_group: torch.distributed.ProcessGroup | int | None = -1,
    ):
        """Initialize the parallel state."""
        self.data_parallel_group = DistributedProcessGroup(data_parallel_group)
        self.tensor_parallel_group = DistributedProcessGroup(tensor_parallel_group)

    def __repr__(self) -> str:
        return f"data_parallel_group: {self.data_parallel_group}, tensor_parallel_group: {self.tensor_parallel_group}"


def get_group(ranks: list[int]):
    """Returns the process group if torch.distributed.is_initialized()."""
    # NCCL has an issue with calling barrier. So we just use the gloo backebnd for group barriers.
    return torch.distributed.new_group(ranks, backend="gloo") if is_initialized() else None


def is_dtensor_sharded(model):
    """Returns True if the model is using DTensor."""
    return any(isinstance(param, DTensor) for param in model.parameters()) or any(
        isinstance(param, DTensor) for param in model.buffers()
    )


class FileLock:
    """Mutex object for writing to a file atomically using the O_EXCL directive on Unix filesystems."""

    def __init__(
        self,
        lockfile_path: str,
        all_acquire: bool = False,
        poll_time: float = 0.25,
    ):
        """Constructor.

        Args:
            lockfile_path: Path to a nonexistent file to be used as the locking mechanism.
            all_acquire: Will keep retrying to acquire a lock if True.
            poll_time: Sleep interval between retries.
        """
        self.lockfile_path = lockfile_path
        self.all_acquire = all_acquire
        self.poll_time = poll_time
        self.handle = None

    def try_acquire(self):
        try:
            self.handle = os.open(self.lockfile_path, os.O_CREAT | os.O_EXCL)
            return True
        except FileExistsError:
            return False

    def wait(self):
        while os.path.exists(self.lockfile_path):
            time.sleep(self.poll_time)

    def release(self):
        if self.handle is not None:
            os.close(self.handle)
        os.remove(self.lockfile_path)

    def __enter__(self):
        while True:
            if self.try_acquire() or not self.all_acquire:
                break
            self.wait()
        return self

    def __exit__(self, exc_type, exc_value, traceback):
        self.release()